{ "map": { "schemaVersion": 1, "version": "0.19.0", "title": "AI risk × development pace", "asOf": "2026-09-15", "scope": "Selected public positions on frontier capability development. This is a curated sample, not a census or ranking.", "coordinateSystem": "editorial-public-position-v1", "entities": [ { "id": "yudkowsky", "name": "Eliezer Yudkowsky", "kind": "person", "status": "placed", "position": { "x": -87, "y": 87, "xRange": [ -100, -72 ], "yRange": [ 72, 100 ] }, "confidence": "high", "basis": "primary", "scope": "Superintelligence and the frontier leading to it.", "caveat": "This is his advocated position, not this project's prediction.", "evidence": [ { "axis": "pace", "source": "yudkowsky-book", "note": "Argues against proceeding to superintelligence with present methods." }, { "axis": "concern", "source": "yudkowsky-book", "note": "Frames superhuman AI as an extinction threat." }, { "axis": "context", "source": "miri-position", "note": "MIRI separately advocates an enforced international ASI moratorium." } ], "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "right" }, "profile": { "summary": "Argues against building superintelligence with present methods.", "sources": [ "yudkowsky-book" ], "sections": [ { "title": "The claim being mapped", "text": "The book announcement presents an argument that superhuman AI would threaten humanity. The map records that argument; it does not present the outcome as established.", "sources": [ "yudkowsky-book" ] } ] } }, { "id": "pauseai", "name": "PauseAI", "kind": "movement", "status": "placed", "position": { "x": -87, "y": 66, "xRange": [ -100, -65 ], "yRange": [ 42, 94 ] }, "confidence": "high", "basis": "primary", "scope": "Its April 2026 proposal for a global pause on the most powerful general-AI training.", "caveat": "A movement has internal variation; this point represents its published proposal.", "evidence": [ { "axis": "pace", "source": "pauseai-proposal-2026", "note": "Advocates a coordinated global pause rather than a company pausing on its own.", "locator": "April 5, 2026 proposal; opening and Treaty Measures" }, { "axis": "concern", "source": "pauseai-proposal-2026", "note": "Its proposal treats dangerous AI progress as a reason for international restraint.", "locator": "April 5, 2026 proposal; opening and Treaty Measures" } ], "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "right" }, "profile": { "summary": "Calls for a coordinated global pause on the most powerful AI training.", "sources": [ "pauseai-proposal-2026" ], "sections": [ { "title": "The proposal's limits", "text": "Its April 2026 proposal calls for safety and democratic-control conditions before more powerful general AI is developed. It distinguishes that target from narrow applications such as medical image recognition.", "sources": [ "pauseai-proposal-2026" ] } ] } }, { "id": "bengio", "name": "Yoshua Bengio", "kind": "person", "status": "placed", "position": { "x": -62, "y": 50, "xRange": [ -85, -35 ], "yRange": [ 25, 85 ] }, "confidence": "medium", "basis": "mixed", "scope": "Superintelligence restriction, not a ban on every AI application or safety research.", "caveat": "Conditional superintelligence restraint is reported separately from his directly read account of control risks. His research proposal is not a proven safety solution.", "evidence": [ { "axis": "pace", "source": "bengio-signature", "note": "Reported signatory of the conditional superintelligence prohibition." }, { "axis": "pace", "source": "superintelligence-statement", "note": "The statement makes safety consensus and public support conditions for lifting the prohibition." }, { "axis": "concern", "source": "bengio-lawzero-2025", "note": "Warns about deception, self-preservation and loss of human control while proposing safety research.", "locator": "LawZero introduction", "role": "support" } ], "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "right" }, "profile": { "summary": "Supports restraint on superintelligence and research aimed at retaining human control.", "sources": [ "bengio-signature", "bengio-lawzero-2025" ], "sections": [ { "title": "Two different kinds of evidence", "text": "The prohibition endorsement is reported; his concern about loss of control is stated directly in his LawZero announcement. Neither establishes a numerical risk estimate.", "sources": [ "bengio-signature", "bengio-lawzero-2025" ] } ] } }, { "id": "amodei", "name": "Dario Amodei", "kind": "person", "status": "placed", "position": { "x": -28, "y": 81, "xRange": [ -48, -8 ], "yRange": [ 55, 96 ] }, "confidence": "high", "basis": "primary", "scope": "Frontier capability growth, especially unchecked self-improvement.", "caveat": "Pacing is not stopping. Coordinated steps remain proposals; no independent operational audit is claimed.", "evidence": [ { "axis": "pace", "source": "amodei-pacing", "note": "Explicitly proposes slower frontier capability growth." }, { "axis": "concern", "source": "amodei-pacing", "note": "Expresses serious loss-of-control and catastrophic-misuse concern." } ], "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "right" }, "profile": { "summary": "Calls for slower frontier capability growth while continuing AI development.", "sources": [ "amodei-pacing" ], "sections": [ { "title": "What pacing means here", "text": "His September essay proposes outside evaluators and coordination among companies and governments. It argues that extra time could improve safeguards; it does not call for all AI work to stop.", "sources": [ "amodei-pacing" ] } ] } }, { "id": "anthropic", "name": "Anthropic", "kind": "organization", "status": "placed", "position": { "x": 0, "y": 51, "xRange": [ -35, 45 ], "yRange": [ 32, 86 ] }, "confidence": "medium", "basis": "primary", "scope": "Published company safeguards and its announced embedded-evaluator commitment.", "caveat": "Giving outside evaluators access does not commit the company to slow all frontier development. The Responsible Scaling Policy separates company promises from industry recommendations; implementation is not independently audited.", "evidence": [ { "axis": "pace", "source": "anthropic-rsp-3-4", "note": "Appendix A promises development delays in specified competitor-related circumstances, not a general training halt.", "locator": "Appendix A, Anthropic in the lead and Competitors have strong safety measures", "role": "support" }, { "axis": "pace", "source": "amodei-pacing", "note": "The company commits to work with outside evaluators. Wider limits on developing more capable models require coordination.", "locator": "Three-step plan and Embedded Evaluators", "role": "counterpoint" }, { "axis": "concern", "source": "anthropic-rsp-3-4", "note": "The policy assesses catastrophic misuse and sabotage that could increase later global-catastrophe risk.", "locator": "Section 1 and risk-report requirements", "role": "support" } ], "implementation": "published-policy-and-announced-commitment", "lastReviewed": "2026-09-15", "label": { "preferred": "right" }, "profile": { "summary": "Promises outside scrutiny and conditional safeguards while continuing development.", "sources": [ "amodei-pacing", "anthropic-rsp-3-4" ], "sections": [ { "title": "Why the position was reassessed", "text": "The previous dot treated evaluator access as pace restraint. Reading the company policy separately supports conditional continuation with a wide range. This is an editorial correction, not evidence of a change of mind.", "sources": [ "anthropic-rsp-3-4", "amodei-pacing" ] }, { "title": "What can be checked", "text": "The policy publishes risk-report and external-review requirements. Announcing a requirement does not establish that it was met in a particular case.", "sources": [ "anthropic-rsp-3-4" ] } ] } }, { "id": "hassabis", "name": "Demis Hassabis", "kind": "person", "status": "placed", "position": { "x": -3, "y": 65, "xRange": [ -30, 25 ], "yRange": [ 38, 86 ] }, "confidence": "medium", "basis": "mixed", "scope": "Public governance proposal and subsequent endorsement.", "caveat": "Do not convert a personal endorsement into a verified Google DeepMind-wide slowdown.", "evidence": [ { "axis": "pace", "source": "hassabis-framework", "note": "Proposes standards that could coordinate a slowdown if necessary." }, { "axis": "concern", "source": "hassabis-framework", "note": "Warns that safeguards and understanding are not keeping pace." }, { "axis": "context", "source": "september-responses", "note": "Reportedly endorsed the direction of Amodei's proposal." } ], "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "right" }, "profile": { "summary": "Proposes shared frontier safeguards and a possible coordinated slowdown.", "sources": [ "hassabis-framework" ], "sections": [ { "title": "A conditional proposal", "text": "His July framework combines support for innovation with a standards body that could coordinate a slowdown if necessary. That proposal does not establish a Google DeepMind-wide operational change.", "sources": [ "hassabis-framework" ] } ] } }, { "id": "altman", "name": "Sam Altman", "kind": "person", "status": "placed", "position": { "x": -8, "y": 29, "xRange": [ -32, 18 ], "yRange": [ 12, 76 ] }, "confidence": "medium", "basis": "mixed", "scope": "Reported September 2026 pacing support, with a directly read earlier governance proposal.", "caveat": "The original social posts were unavailable to this review; the reporting is linked.", "evidence": [ { "axis": "pace", "source": "september-responses", "note": "Backed pacing and embedded independent evaluation." }, { "axis": "concern", "source": "september-responses", "note": "Identified loss of control as a failure mode." }, { "axis": "context", "source": "altman-governance-2023", "note": "His coauthored 2023 proposal considered limits on frontier capability growth while exempting lower-capability systems.", "locator": "A starting point; What's not in scope", "role": "context" } ], "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "right" }, "profile": { "summary": "Supports pacing in September reporting; earlier coauthored proposals also considered frontier limits.", "sources": [ "september-responses", "altman-governance-2023" ], "sections": [ { "title": "What is direct and what is reported", "text": "The 2023 article is directly available. The newer social response remains indirectly verified through reporting; the older proposal does not make that post independently verified.", "sources": [ "altman-governance-2023", "september-responses" ] } ] } }, { "id": "openai", "name": "OpenAI", "kind": "organization", "status": "placed", "position": { "x": -11, "y": 9, "xRange": [ -35, 20 ], "yRange": [ 0, 66 ] }, "confidence": "medium", "basis": "primary", "scope": "Selected frontier research workloads and controls.", "caveat": "The company reports selective restrictions, not a full training halt. Concern level remains an editorial inference.", "evidence": [ { "axis": "pace", "source": "openai-pacing", "note": "Reports restrictions on selected frontier research workloads pending stronger safeguards." }, { "axis": "concern", "source": "openai-pacing", "note": "Treats stronger cyber capabilities and alignment failures as requiring safeguards." } ], "implementation": "company-reported-restraint", "lastReviewed": "2026-09-15", "label": { "preferred": "right" }, "profile": { "summary": "Reports selective research pauses while upgrading safeguards.", "sources": [ "openai-pacing" ], "sections": [ { "title": "How to read the company claim", "text": "The company describes pausing certain tool-using research workloads, then resuming some under stronger controls. This supports targeted restraint; it does not establish a company-wide halt.", "sources": [ "openai-pacing" ] } ] } }, { "id": "musk", "name": "Elon Musk", "kind": "person", "status": "placed", "position": { "x": 15, "y": 50, "xRange": [ -75, 85 ], "yRange": [ 10, 85 ] }, "confidence": "medium", "basis": "mixed", "scope": "Personal statements combining frontier scaling, concern about human control and reported support for pacing.", "caveat": "The September endorsement is indirect and brief. Building plans indicate a personal preference, not a general regulatory policy or a company safety rating. The conflicting signals warrant a broad pace range.", "evidence": [ { "axis": "pace", "source": "musk-dwarkesh-2026", "note": "Advocates expanding intelligence through Grok's mission and describes removing bottlenecks to rapid AI-compute growth.", "locator": "Scaling, 00:00:00–00:36:46; Grok's mission, from 00:36:46", "role": "support" }, { "axis": "concern", "source": "musk-dwarkesh-2026", "note": "Questions whether humans could remain in charge of much smarter AI; wants human survival but acknowledges his proposed values are no guarantee.", "locator": "Grok and alignment, 00:36:46–00:59:56", "role": "support" }, { "axis": "pace", "source": "september-responses", "note": "The September roundup reports his endorsement of Amodei's pacing essay. This qualifies a simple acceleration-only reading.", "locator": "Elon Musk response; linked post not independently retrieved", "role": "counterpoint" }, { "axis": "context", "source": "musk-lex-2023", "note": "In 2023, described rapid compute expansion alongside a preference for delayed open sourcing and a pro-human safety rationale.", "locator": "01:23:13–01:29:39", "role": "context" } ], "profile": { "summary": "Scaling ambition and concern about control coexist in the reviewed statements.", "sources": [ "musk-dwarkesh-2026", "september-responses" ], "sections": [ { "title": "Why the anchor changed", "text": "The earlier dot represented only a reported endorsement. This review adds direct building and control statements, not a spending-based score. Its position near the middle of the pace axis summarizes conflicting signals; it does not measure a change in Musk's beliefs.", "sources": [ "musk-dwarkesh-2026", "september-responses" ] }, { "title": "The proposed safeguard is an argument", "text": "Musk argues that curiosity and truth-seeking could preserve humanity. The interviewer challenges that connection. The transcript does not establish that the proposed values reliably produce safe behavior.", "sources": [ "musk-dwarkesh-2026" ] }, { "title": "Keep the historical record separate", "text": "The 2023 discussion already combines building with a safety rationale. It also discusses delayed openness. Neither that interview nor the newer endorsement establishes a present company-wide halt.", "sources": [ "musk-lex-2023", "september-responses" ] } ] }, "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "right", "dx": 230, "dy": -80 } }, { "id": "deepmind", "name": "Google DeepMind", "kind": "organization", "status": "placed", "position": { "x": 20, "y": 46, "xRange": [ -10, 47 ], "yRange": [ 20, 80 ] }, "confidence": "medium", "basis": "primary", "scope": "Institutional frontier-safety framework.", "caveat": "Its framework supports conditional progress. No whole-lab slowdown is established by this evidence.", "evidence": [ { "axis": "pace", "source": "deepmind-framework", "note": "Continues frontier development under capability-linked safeguards." }, { "axis": "concern", "source": "deepmind-framework", "note": "Tracks severe risks and mitigation requirements." } ], "implementation": "published-framework", "lastReviewed": "2026-09-15", "label": { "preferred": "right" }, "profile": { "summary": "Links continued model development to capability-based safeguards.", "sources": [ "deepmind-framework" ], "sections": [ { "title": "How to read the company claim", "text": "Its updated framework includes safety-case reviews for some external launches and large-scale internal uses. A requirement in a framework is distinct from proof that a specific deployment met it.", "sources": [ "deepmind-framework" ] } ] } }, { "id": "microsoft-ai", "name": "Microsoft AI", "kind": "organization", "status": "placed", "position": { "x": 38, "y": 12, "xRange": [ 0, 68 ], "yRange": [ 0, 59 ] }, "confidence": "medium", "basis": "primary", "scope": "Microsoft AI model development, not the entire Microsoft group.", "caveat": "A human-control code is not a catastrophe probability. The consultation is not a verified slowdown.", "evidence": [ { "axis": "pace", "source": "microsoft-code", "note": "Plans continued model training and deployment guided by its code." }, { "axis": "concern", "source": "microsoft-code", "note": "Makes retaining human control a central requirement." } ], "implementation": "announced-commitment", "lastReviewed": "2026-09-15", "label": { "preferred": "right" }, "profile": { "summary": "Plans continued model development under human-control commitments.", "sources": [ "microsoft-code" ], "sections": [ { "title": "How to read the company claim", "text": "The code describes limits and evaluation goals for MAI models. A public consultation and a published rule do not show how reliably a model follows the rule.", "sources": [ "microsoft-code" ] } ] } }, { "id": "meta", "name": "Meta", "kind": "organization", "status": "placed", "position": { "x": 65, "y": 25, "xRange": [ 15, 84 ], "yRange": [ 0, 67 ] }, "confidence": "medium", "basis": "primary", "scope": "Meta's frontier development and published safeguards.", "caveat": "Publishing a risk framework does not reveal a probability belief; nor does openness imply low concern.", "evidence": [ { "axis": "pace", "source": "meta-framework", "note": "Describes continued scaling with release safeguards." }, { "axis": "concern", "source": "meta-framework", "note": "Explicitly addresses catastrophic risks and loss of control." } ], "implementation": "published-framework", "lastReviewed": "2026-09-15", "label": { "preferred": "right" }, "profile": { "summary": "Continues advanced-model development with stated release safeguards.", "sources": [ "meta-framework" ], "sections": [ { "title": "How to read the company claim", "text": "The framework addresses severe misuse and loss of control across open and closed releases. This is the company's account of its safeguards, not an independent safety finding.", "sources": [ "meta-framework" ] } ] } }, { "id": "lecun", "name": "Yann LeCun", "kind": "person", "status": "placed", "position": { "x": 43, "y": -48, "xRange": [ 20, 75 ], "yRange": [ -90, -12 ] }, "confidence": "medium", "basis": "mixed", "scope": "His public views, separate from Meta's institutional policy.", "caveat": "Skepticism about catastrophe is not a claim that all AI harms are negligible.", "evidence": [ { "axis": "pace", "source": "lecun-interview", "note": "Favors broad AI development and access over catastrophe-driven restrictions." }, { "axis": "concern", "source": "lecun-interview", "note": "Disputes inevitable loss-of-control arguments." }, { "axis": "context", "source": "september-responses", "note": "Rejected Amodei's latest slowdown case." } ], "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "right" }, "profile": { "summary": "Disputes takeover arguments and favors broad AI development and access.", "sources": [ "lecun-interview" ], "sections": [ { "title": "What his skepticism covers", "text": "In the 2024 interview, he argues that progress and safeguards can develop gradually. He also describes concentrated control of AI as a serious concern. Skepticism about extinction is not dismissal of every harm.", "sources": [ "lecun-interview" ] } ] } }, { "id": "andreessen", "name": "Marc Andreessen", "kind": "person", "status": "placed", "position": { "x": 77, "y": -74, "xRange": [ 55, 100 ], "yRange": [ -98, -33 ] }, "confidence": "medium", "basis": "mixed", "scope": "His dated 2023 manifesto, with a 2026 interview-publisher summary for context.", "caveat": "The detailed catastrophe argument remains the 2023 manifesto. A newer episode summary supports continued pro-growth advocacy but cannot substitute for a full updated two-axis interview review.", "profile": { "summary": "Argues for faster technological development and against catastrophe-driven restraint.", "sources": [ "andreessen-manifesto" ], "sections": [ { "title": "A dated argument with newer context", "text": "The 2023 manifesto supplies the detailed argument. A June 2026 episode description reports continued enthusiasm for AI growth; it is context, not a newly reviewed statement about every catastrophic-risk scenario.", "sources": [ "andreessen-manifesto", "andreessen-interview-summary-2026" ] } ] }, "evidence": [ { "axis": "pace", "source": "andreessen-manifesto", "note": "Explicitly opposes deceleration and favors technological growth." }, { "axis": "concern", "source": "andreessen-manifesto", "note": "Rejects catastrophe-oriented arguments for restricting progress." }, { "axis": "context", "source": "andreessen-interview-summary-2026", "note": "The interview publisher describes continued enthusiasm for AI growth and concern about policy barriers.", "locator": "Episode description; audio not independently reviewed", "role": "context" } ], "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "left" } }, { "id": "a16z", "name": "Andreessen Horowitz (a16z)", "kind": "organization", "status": "unplaced", "position": null, "confidence": null, "basis": "primary", "scope": "The investment firm, kept separate from its individual authors.", "caveat": "The previously used manifesto explicitly disclaims being the firm's view. A firm-level position on both axes has not been established.", "evidence": [ { "axis": "context", "source": "andreessen-manifesto", "note": "The page attributes its views to individual personnel and expressly excludes the firm and affiliates.", "locator": "Footer disclaimer", "role": "counterpoint" } ], "implementation": "not-assessed", "lastReviewed": "2026-09-15", "label": { "preferred": "left" }, "profile": { "summary": "Publishing a founder's essay does not establish the investment firm's position.", "sources": [ "andreessen-manifesto" ], "sections": [ { "title": "Why this record has no dot", "text": "The manifesto remains evidence for Marc Andreessen's personal record. Its footer prevents us from treating the same text as the firm's stance on both axes.", "sources": [ "andreessen-manifesto" ] } ] } }, { "id": "xai", "name": "xAI / SpaceXAI", "kind": "organization", "status": "unplaced", "position": null, "confidence": null, "basis": "unreviewed", "scope": "", "caveat": "Corporate position not fully reviewed. Musk's endorsement does not by itself establish company policy or a company-wide slowdown.", "evidence": [], "implementation": "not-assessed", "lastReviewed": null, "label": { "preferred": "right" } }, { "id": "mistral", "name": "Mistral AI", "kind": "organization", "status": "unplaced", "position": null, "confidence": null, "basis": "unreviewed", "scope": "", "caveat": "Not reviewed on both axes in this release. Open weights and pro-growth positioning alone would not establish catastrophic-risk concern.", "evidence": [], "implementation": "not-assessed", "lastReviewed": null, "label": { "preferred": "right" } }, { "id": "deepseek", "name": "DeepSeek", "kind": "organization", "status": "unplaced", "position": null, "confidence": null, "basis": "primary", "scope": "Its own model-release announcements.", "caveat": "Publishing model weights and benchmark claims does not establish a preferred frontier pace or catastrophic-risk stance. A two-axis position remains unplaced.", "evidence": [ { "axis": "context", "source": "deepseek-r1-release", "note": "Announces R1 model access and release of weights and related materials.", "locator": "Release announcement", "role": "context" } ], "implementation": "published-product-announcement", "lastReviewed": "2026-09-15", "label": { "preferred": "right" }, "profile": { "summary": "Publishes AI models; the reviewed release does not establish both map axes.", "sources": [ "deepseek-r1-release" ], "sections": [ { "title": "What the source can establish", "text": "The company announcement describes a model release. Model access, performance claims and corporate ideology are different questions.", "sources": [ "deepseek-r1-release" ] } ] } }, { "id": "alibaba", "name": "Alibaba / Qwen", "kind": "organization", "status": "unplaced", "position": null, "confidence": null, "basis": "unreviewed", "scope": "", "caveat": "Coverage backlog: no reviewed two-axis placement in this release.", "evidence": [], "implementation": "not-assessed", "lastReviewed": null, "label": { "preferred": "right" } }, { "id": "nvidia", "name": "NVIDIA", "kind": "organization", "status": "unplaced", "position": null, "confidence": null, "basis": "unreviewed", "scope": "", "caveat": "Coverage backlog: infrastructure expansion is not itself a quantified risk belief.", "evidence": [], "implementation": "not-assessed", "lastReviewed": null, "label": { "preferred": "right" } }, { "id": "trump", "name": "Donald Trump", "kind": "person", "status": "placed", "position": { "x": 85, "y": -80, "xRange": [ 65, 100 ], "yRange": [ -100, -50 ] }, "confidence": "medium", "basis": "mixed", "scope": "His September 2026 opposition to slowing advanced AI, read alongside his June executive order.", "caveat": "September posts are verified through reporting, not the original Truth Social pages. This maps expressed catastrophic-risk concern, not private beliefs, technical expertise, or every administration policy. The ranges are editorial judgments.", "evidence": [ { "axis": "pace", "source": "trump-forbes-2026-09-14", "note": "Forbes reports opposition to calls for an AI slowdown, emphasizing competition with China and the costs of regulation.", "locator": "Key Facts and Trump Dismisses AI Concerns; September 14 posts and September 13 remarks", "role": "support" }, { "axis": "concern", "source": "trump-afp-2026-09-14", "note": "AFP reports that he dismissed scenarios of AI destroying humanity as a hoax.", "locator": "Opening paragraphs on September 14 Truth Social posts", "role": "support" }, { "axis": "pace", "source": "trump-ai-security-order-2026", "note": "His signed order favors developing advanced AI and explicitly excludes mandatory model licensing or preclearance under its frontier-model section.", "locator": "Executive Order 14409, sections 1 and 3(c)", "role": "support" }, { "axis": "context", "source": "trump-ai-security-order-2026", "note": "The same order directs cybersecurity measures and a voluntary process for evaluating frontier models before wider release.", "locator": "Sections 2, 3(a)–(b) and 4", "role": "counterpoint" } ], "profile": { "summary": "Favors continued AI development and publicly dismisses AI-extinction warnings; his signed policy still includes security measures.", "sources": [ "trump-forbes-2026-09-14", "trump-afp-2026-09-14", "trump-ai-security-order-2026" ], "sections": [ { "title": "Why the dot is at the lower right", "text": "The September reporting supports a strong preference against slowing AI and low expressed concern about human extinction. The coordinates summarize this public stance; they do not measure risk or establish membership in e/acc.", "sources": [ "trump-forbes-2026-09-14", "trump-afp-2026-09-14" ] }, { "title": "What the latest posts establish", "text": "Forbes and AFP report September 14 posts opposing new guardrails and framing AI as an economic and competitive priority. These are attributed political claims, not evidence that advanced AI is safe or that new safeguards would cause bankruptcy.", "sources": [ "trump-forbes-2026-09-14", "trump-afp-2026-09-14" ] }, { "title": "Security policy is a qualification", "text": "The June 2 order calls for cyber-defense measures, capability benchmarks and voluntary early government access to some frontier models. These provisions qualify any claim that he rejects all safeguards. A signed direction does not establish that the measures have been implemented or work.", "sources": [ "trump-ai-security-order-2026" ] }, { "title": "How current and direct is this review?", "text": "Reviewed September 15, 2026. The September posts were read through reporting because Truth Social did not expose their text to this review. The June order was read directly. This is a dated selection, not an exhaustive archive of his posts or a separate assessment of the US government.", "sources": [ "trump-forbes-2026-09-14", "trump-afp-2026-09-14", "trump-ai-security-order-2026" ] } ] }, "implementation": "public-advocacy-and-signed-policy", "lastReviewed": "2026-09-15", "label": { "preferred": "left" } }, { "id": "thiel", "name": "Peter Thiel", "kind": "person", "status": "placed", "position": { "x": 65, "y": -20, "xRange": [ 25, 90 ], "yRange": [ -55, 35 ] }, "confidence": "medium", "basis": "primary", "scope": "His arguments for continued AI development and against concentrated power used to suppress it.", "caveat": "Lower catastrophic-risk emphasis is relative to his concern about political control, not a claim that he rules out AI catastrophe. His economic and theological arguments do not specify a detailed frontier-training policy.", "evidence": [ { "axis": "pace", "source": "thiel-tyler-2024", "note": "Opposes the power needed to stop AI and favors AI even under a labor-substitution scenario.", "locator": "Q&A: human extinction and technological replacement of workers", "role": "support" }, { "axis": "concern", "source": "thiel-tyler-2024", "note": "Explicitly prioritizes concern about humans stopping AI over AI destroying humanity; rejects treating the outcome as predetermined.", "locator": "Q&A: When do you think humans are going to destroy themselves?", "role": "support" }, { "axis": "concern", "source": "thiel-hoover-apocalypse", "note": "Says both technological catastrophe and a totalitarian world state deserve concern, while prioritizing the latter.", "locator": "Scylla and Charybdis discussion; search for worry about both", "role": "counterpoint" }, { "axis": "pace", "source": "thiel-spectator-2025", "note": "Favors embracing AI as a source of growth while acknowledging concentrated returns and displacement of labor.", "locator": "John Power's AI bubble and abundance question", "role": "support" }, { "axis": "context", "source": "thiel-spectator-2026", "note": "A 2026 report describes his continuing use of the Antichrist framework in discussing political power.", "locator": "Cambridge talk account; an edited report, not a complete transcript", "role": "context" } ], "profile": { "summary": "Favors AI development while placing greater emphasis on the danger of concentrated political control.", "sources": [ "thiel-tyler-2024", "thiel-spectator-2025" ], "sections": [ { "title": "Why this position is on the right", "text": "His opposition to stopping AI supports the development side of this map. A moderate rather than extreme anchor reflects the absence of a detailed training timetable or unrestricted-deployment proposal in the reviewed material.", "sources": [ "thiel-tyler-2024", "thiel-spectator-2025" ] }, { "title": "Why lower concern does not mean no concern", "text": "His Hoover conversation calls for concern about both catastrophe and totalitarian control. The vertical range crosses the middle because the relative emphasis is clearer than an absolute level of AI-risk concern.", "sources": [ "thiel-hoover-apocalypse" ] }, { "title": "What the theology does and does not establish", "text": "His Antichrist framing is a speculative argument about power gained through fear of catastrophe. The Hoover conversation was recorded in October 2024; it does not establish a current policy proposal or the truth of a prophecy.", "sources": [ "thiel-hoover-apocalypse", "thiel-spectator-2026" ] }, { "title": "An economic reservation", "text": "In the December interview he worries about benefits accruing to a few firms and AI replacing workers. These reservations complicate a blanket techno-optimist label without establishing support for a frontier pause.", "sources": [ "thiel-spectator-2025" ] } ] }, "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "left" } }, { "id": "amazon", "name": "Amazon / AWS", "kind": "organization", "status": "unplaced", "position": null, "confidence": null, "basis": "unreviewed", "scope": "", "caveat": "Coverage backlog: separate institutional policy from partner-lab positions.", "evidence": [], "implementation": "not-assessed", "lastReviewed": null, "label": { "preferred": "right" } }, { "id": "hinton", "name": "Geoffrey Hinton", "kind": "person", "status": "placed", "position": { "x": -50, "y": 75, "xRange": [ -80, -15 ], "yRange": [ 50, 95 ] }, "confidence": "medium", "basis": "mixed", "scope": "Conditional controls on exceptionally capable AI in a coauthored policy paper.", "caveat": "The proposal is collective and dated. Its control-risk warning is not a probability estimate or a claim that every kind of AI should stop.", "evidence": [ { "axis": "pace", "source": "extreme-ai-risks-2024", "note": "Coauthors a call for conditional development halts when dangerous capabilities emerge.", "locator": "Mitigation", "role": "support" }, { "axis": "concern", "source": "extreme-ai-risks-2024", "note": "Coauthors a warning about irreversible loss of control and human extinction.", "locator": "Societal-scale risks", "role": "support" }, { "axis": "context", "source": "hinton-gzero-2025", "note": "The interview publisher reports continued concern about humans losing control.", "locator": "December 2025 episode summary", "role": "context" } ], "profile": { "summary": "Advocates conditional restraints on highly capable AI and warns about losing control.", "sources": [ "extreme-ai-risks-2024", "hinton-gzero-2025" ], "sections": [ { "title": "Reading the placement", "text": "This point summarizes the coauthored proposal; its range allows different readings.", "sources": [ "extreme-ai-risks-2024" ] }, { "title": "Newer context", "text": "The December 2025 publisher summary describes his concern about control. His proposed protective instincts are an idea, not a demonstrated safeguard.", "sources": [ "hinton-gzero-2025" ] } ] }, "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "right" } }, { "id": "russell", "name": "Stuart Russell", "kind": "person", "status": "placed", "position": { "x": -50, "y": 75, "xRange": [ -80, -15 ], "yRange": [ 50, 95 ] }, "confidence": "medium", "basis": "primary", "scope": "Conditional limits on dangerous frontier development and evidence required before release.", "caveat": "A coauthored policy proposal and dated testimony support this interpretation. They do not establish a blanket ban or an institutional Berkeley position.", "evidence": [ { "axis": "pace", "source": "extreme-ai-risks-2024", "note": "Coauthors conditional development halts pending adequate protections.", "locator": "Mitigation", "role": "support" }, { "axis": "concern", "source": "russell-senate-2023", "note": "His testimony warns that AGI without reliable human control could threaten human survival.", "locator": "Executive summary", "role": "support" } ], "profile": { "summary": "Calls for enforceable safety conditions and warns about loss of human control.", "sources": [ "extreme-ai-risks-2024", "russell-senate-2023" ], "sections": [ { "title": "What the conditions mean", "text": "His testimony puts the burden on developers to show safety before release. It also discusses present harms; concern about catastrophe does not replace those issues.", "sources": [ "russell-senate-2023" ] } ] }, "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "right" } }, { "id": "gebru", "name": "Timnit Gebru", "kind": "person", "status": "unplaced", "position": null, "confidence": null, "basis": "primary", "scope": "Her public criticism of undefined AGI goals and her distinctions between AI risks.", "caveat": "Unplaced: this review cannot summarize the different risks she discusses with one concern coordinate.", "evidence": [ { "axis": "pace", "source": "gebru-torres-2024", "note": "Recommends research on defined, testable tasks instead of trying to build an all-purpose AGI.", "locator": "Journal abstract", "role": "support" }, { "axis": "context", "source": "gebru-torres-2024", "note": "The authors emphasize harms to marginalized people and concentration of power.", "locator": "Journal abstract", "role": "context" }, { "axis": "concern", "source": "gebru-wired-2026", "note": "Rejects rogue-machine extinction stories while naming weapons and climate as serious threats.", "locator": "Answers about extinction scenarios and present threats", "role": "context" } ], "profile": { "summary": "Questions the pursuit of AGI and argues for research on defined tasks.", "sources": [ "gebru-torres-2024" ], "sections": [ { "title": "Why there is no dot", "text": "The coauthored paper supports a research focus on defined tasks. Its wider historical thesis is not adopted as an atlas conclusion. The concern evidence below discusses different types of harm, so this review leaves the point unplaced.", "sources": [ "gebru-torres-2024", "gebru-wired-2026" ] } ] }, "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "right" } }, { "id": "bender", "name": "Emily M. Bender", "kind": "person", "status": "unplaced", "position": null, "confidence": null, "basis": "primary", "scope": "Her arguments about AI use, evidence, labor and rights.", "caveat": "Resistance to adopting a product is distinct from a preferred pace of frontier capability growth. This source does not settle both map axes.", "evidence": [ { "axis": "context", "source": "bender-humanities-2025", "note": "Urges scrutiny of task definitions, training data, claimed accuracy, labor and surveillance risks.", "locator": "Slides 25–26", "role": "context" } ], "profile": { "summary": "Asks who benefits from AI systems and whether their claims can be checked.", "sources": [ "bender-humanities-2025" ], "sections": [ { "title": "Questions a reader can use", "text": "Her slides ask what task a system performs, whether its inputs support accurate output, and whether it can be used to deny people's rights. These questions apply to evaluating products; they do not automatically locate a person on this map.", "sources": [ "bender-humanities-2025" ] } ] }, "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "right" } }, { "id": "fei-fei-li", "name": "Fei-Fei Li", "kind": "person", "status": "unplaced", "position": null, "confidence": null, "basis": "primary", "scope": "Her individual UN briefing on public research, access and governance.", "caveat": "The briefing supports investment and safeguards but does not establish a two-axis frontier-pacing position. It expressly excludes attribution to affiliated organizations.", "evidence": [ { "axis": "context", "source": "fei-fei-li-un-2024", "note": "Calls for public investment, wider access and evidence-based governance while warning about misuse.", "locator": "Public Sector Leadership and Science- and Evidence-Based AI Policymaking", "role": "context" } ], "profile": { "summary": "Supports public AI research, wider access and evidence-based safeguards.", "sources": [ "fei-fei-li-un-2024" ], "sections": [ { "title": "A different emphasis", "text": "Her briefing focuses on who can develop and benefit from AI, global collaboration and misuse. Those subjects matter beyond the two questions represented by this chart.", "sources": [ "fei-fei-li-un-2024" ] } ] }, "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "right" } }, { "id": "whittaker", "name": "Meredith Whittaker", "kind": "person", "status": "unplaced", "position": null, "confidence": null, "basis": "primary", "scope": "Her argument about AI, surveillance and control of digital infrastructure.", "caveat": "The keynote focuses on privacy and power. It does not provide enough evidence for both frontier-pacing and catastrophe-concern coordinates.", "evidence": [ { "axis": "context", "source": "whittaker-ndss-2024", "note": "Argues that large-scale commercial AI can reinforce data collection and concentrated corporate power.", "locator": "Pages 10–14", "role": "context" } ], "profile": { "summary": "Focuses on privacy, surveillance and who controls digital infrastructure.", "sources": [ "whittaker-ndss-2024" ], "sections": [ { "title": "Why privacy belongs in the wider debate", "text": "Her argument asks whether people can influence how technology affects their lives. It does not establish that every AI system has the same business model or that privacy concerns imply one catastrophe-risk position.", "sources": [ "whittaker-ndss-2024" ] } ] }, "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "right" } }, { "id": "eu-ai-office", "name": "European AI Office", "kind": "organization", "status": "unplaced", "position": null, "confidence": null, "basis": "primary", "scope": "The European Commission office's published mandate.", "caveat": "An enforcement and innovation mandate is not a single ideology or a measured level of catastrophic-risk concern. The two-axis position is not established.", "evidence": [ { "axis": "context", "source": "eu-ai-office-overview", "note": "Supports general-purpose AI oversight, systemic-risk assessment and trustworthy AI innovation.", "locator": "Tasks of the AI Office", "role": "context" } ], "profile": { "summary": "Works on AI oversight and innovation within the European Commission.", "sources": [ "eu-ai-office-overview" ], "sections": [ { "title": "Institutional role", "text": "The office describes work on model evaluations, general-purpose AI rules and support for innovation. These functions should be compared as a mandate, not equated with a personal belief.", "sources": [ "eu-ai-office-overview" ] } ] }, "implementation": "published-institutional-mandate", "lastReviewed": "2026-09-15", "label": { "preferred": "right" } }, { "id": "uk-aisi", "name": "UK AI Security Institute", "kind": "organization", "status": "unplaced", "position": null, "confidence": null, "basis": "primary", "scope": "Its published research agenda, separate from the whole UK government's policy.", "caveat": "Research on catastrophic harm does not establish whether the institute favors a general acceleration or slowdown of frontier development.", "evidence": [ { "axis": "concern", "source": "uk-aisi-agenda", "note": "Studies whether autonomous AI could cause catastrophic harm or permanently evade human control.", "locator": "Autonomous Systems", "role": "support" }, { "axis": "context", "source": "uk-aisi-name-2025", "note": "The February 2025 announcement names it the AI Security Institute and describes its security focus.", "locator": "Opening announcement", "role": "context" } ], "profile": { "summary": "Studies serious AI security risks, including failures of human control.", "sources": [ "uk-aisi-agenda", "uk-aisi-name-2025" ], "sections": [ { "title": "Why it is unplaced", "text": "The research agenda establishes topics it studies. A preferred pace of frontier development would need separate evidence; it cannot be read from the institute's name.", "sources": [ "uk-aisi-agenda" ] } ] }, "implementation": "published-institutional-mandate", "lastReviewed": "2026-09-15", "label": { "preferred": "right" } }, { "id": "vance", "name": "JD Vance", "kind": "person", "status": "unplaced", "position": null, "confidence": null, "basis": "mixed", "scope": "His February 2025 Paris speech and September 2026 reported response.", "caveat": "Opposition to regulation is clearer than a specific catastrophe-risk stance in the material reviewed. Trump's statements are not assigned to Vance.", "evidence": [ { "axis": "pace", "source": "vance-paris-2025", "note": "Favors cutting-edge AI development and argues that excessive regulation can obstruct it.", "locator": "Remarks on development and regulation", "role": "support" }, { "axis": "context", "source": "vance-paris-2025", "note": "Also says safety concerns still matter and identifies misuse and national-security risks.", "locator": "Closing remarks and national-security passage", "role": "counterpoint" }, { "axis": "context", "source": "trump-forbes-2026-09-14", "note": "September reporting describes skepticism of company requests for regulation while acknowledging that technology carries risks.", "locator": "Key Facts; Vance's remarks", "role": "context" } ], "profile": { "summary": "Favors AI development and questions broad regulatory barriers.", "sources": [ "vance-paris-2025", "trump-forbes-2026-09-14" ], "sections": [ { "title": "Why there is no dot", "text": "His pro-development position is clear, but general statements about safety do not settle his expressed concern about AI catastrophe. This record keeps that gap visible.", "sources": [ "vance-paris-2025" ] } ] }, "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "right" } }, { "id": "sutskever", "name": "Ilya Sutskever", "kind": "person", "status": "placed", "position": { "x": 15, "y": 72, "xRange": [ -35, 65 ], "yRange": [ 45, 95 ] }, "confidence": "medium", "basis": "primary", "scope": "Personal research ambitions and proposed limits on the most powerful superintelligence.", "caveat": "The proposed power cap has no specified method. Safety is a research aim, not an established property of future systems.", "evidence": [ { "axis": "pace", "source": "sutskever-nvidia-2026", "note": "Says his research is ready to be scaled with a larger computer.", "locator": "Sutskever statement in the partnership announcement", "role": "support" }, { "axis": "pace", "source": "sutskever-dwarkesh-2025", "note": "Supports a cap on the most powerful superintelligence, while leaving the method open.", "locator": "01:02:37 to 01:04:10", "role": "counterpoint" }, { "axis": "concern", "source": "sutskever-superalignment-2023", "note": "His coauthored article warns of human disempowerment or extinction from superintelligence.", "locator": "Opening and alignment problem", "role": "support" }, { "axis": "concern", "source": "sutskever-dwarkesh-2025", "note": "Still treats extreme system power and control as central safety concerns.", "locator": "00:56:10 to 01:04:10", "role": "support" } ], "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "right" }, "profile": { "summary": "Pursues more capable systems while supporting limits on extreme power; control remains a research problem.", "sources": [ "sutskever-nvidia-2026", "sutskever-dwarkesh-2025", "sutskever-superalignment-2023" ], "sections": [ { "title": "Research pace and release plans", "text": "In his November 2025 interview, he gives more weight to incremental release. That concerns deployment, not a general halt to capability research.", "sources": [ "sutskever-dwarkesh-2025" ] } ] } }, { "id": "zuckerberg", "name": "Mark Zuckerberg", "kind": "person", "status": "placed", "position": { "x": 78, "y": 18, "xRange": [ 45, 98 ], "yRange": [ -20, 65 ] }, "confidence": "medium", "basis": "primary", "scope": "Personal advocacy for frontier model innovation and distributed control of superintelligence.", "caveat": "Distributed power is his proposed safeguard. The essay does not establish that competing self-improving systems remain controllable.", "evidence": [ { "axis": "pace", "source": "zuckerberg-future-2026", "note": "Proposes safeguards without slowing model innovation and continued training of leading models.", "locator": "Securing Against AI Misuse; concluding policy implications", "role": "support" }, { "axis": "concern", "source": "zuckerberg-future-2026", "note": "Discusses existential risk and loss of control, proposing a balance of power as the remedy.", "locator": "Alignment With People and Addressing Existential Risk; Maintaining Control of Superintelligence", "role": "support" }, { "axis": "context", "source": "zuckerberg-personal-2025", "note": "His earlier letter already acknowledges new safety concerns and caution about open releases.", "locator": "Paragraph beginning: We believe the benefits of superintelligence", "role": "counterpoint" } ], "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "right" }, "profile": { "summary": "Favors continued model innovation and distributed access while acknowledging risks of losing human control.", "sources": [ "zuckerberg-future-2026" ], "sections": [ { "title": "Optimism with qualifications", "text": "He allows extra mitigation time in some circumstances. This qualifies his opposition to delay, without establishing a general slowdown policy.", "sources": [ "zuckerberg-future-2026" ] } ] } }, { "id": "suleyman", "name": "Mustafa Suleyman", "kind": "person", "status": "placed", "position": { "x": 25, "y": 60, "xRange": [ -15, 60 ], "yRange": [ 30, 85 ] }, "confidence": "medium", "basis": "primary", "scope": "Personal advocacy for advanced-model research with limits on autonomy and loss of control.", "caveat": "Limits on autonomy do not establish a general capability slowdown. He says implementation of the new code begins after consultation.", "evidence": [ { "axis": "pace", "source": "suleyman-humanist-2025", "note": "Supports advanced AI research while prioritizing controllability over unrestricted autonomy.", "locator": "A humanist future; Towards humanist superintelligence", "role": "support" }, { "axis": "pace", "source": "suleyman-code-2026", "note": "Says he will accept less autonomy to preserve human control.", "locator": "Paragraph beginning: We want to create incredible AI", "role": "counterpoint" }, { "axis": "concern", "source": "suleyman-humanist-2025", "note": "Questions how people could continually contain and control self-improving superintelligence.", "locator": "Containment is necessary; The purpose of technology", "role": "support" }, { "axis": "context", "source": "suleyman-code-2026", "note": "The code is a consultation draft; he places implementation after the final draft.", "locator": "Paragraph beginning: Once we have the final post-consultation draft", "role": "context" } ], "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "right" }, "profile": { "summary": "Supports advanced-model research, but accepts autonomy limits to retain human control; the new code awaits implementation.", "sources": [ "suleyman-humanist-2025", "suleyman-code-2026" ], "sections": [ { "title": "Personal position and company process", "text": "His September essay sets out his priorities for training and operating models. This personal record is separate from Microsoft AI’s institutional commitments.", "sources": [ "suleyman-code-2026" ] } ] } }, { "id": "narayanan", "name": "Arvind Narayanan", "kind": "person", "status": "placed", "position": { "x": 0, "y": -20, "xRange": [ -35, 45 ], "yRange": [ -55, 30 ] }, "confidence": "medium", "basis": "primary", "scope": "Joint public arguments about frontier development, AI control and catastrophic risk, including the September 2026 update.", "caveat": "The reviewed arguments are jointly authored. Pauses concern particular experiments; the overall pace preference remains conditional. Coordinates summarize public arguments, not private probabilities.", "evidence": [ { "axis": "pace", "source": "normal-technology-control-2026", "note": "Calls for organizational oversight and for pausing experiments when needed to put it in place.", "locator": "Part 1: Existing organizational governance norms would have prevented the incident", "role": "support" }, { "axis": "concern", "source": "normal-technology-control-2026", "note": "Says catastrophic risks are not imminent but are increasing as defenses and policy lag; safety is not on track.", "locator": "Part 3: Is AI safety on track?", "role": "support" }, { "axis": "pace", "source": "normal-technology-software-work-2026", "note": "The earlier June essay favored accountability and control over slowing technical capability development.", "locator": "The decide-execute-deliver discussion, before Vibe coding is not agentic engineering", "role": "counterpoint" }, { "axis": "context", "source": "normal-technology-control-2026", "note": "Acknowledges underestimating risks during development and companies' failures to take basic precautions.", "locator": "Part 3: Do risks arise from development or deployment?; The continuity hypothesis", "role": "counterpoint" } ], "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "left" }, "profile": { "summary": "Coauthors the normal-technology view. Now supports targeted experiment pauses and warns that safety efforts are falling behind.", "sources": [ "normal-technology-2025", "normal-technology-control-2026" ], "sections": [ { "title": "A contested account of control", "text": "Scott Alexander argues that rapid self-improvement and adoption by AI labs could defeat the thesis's assumed limits. Narayanan and Kapoor reply that technical improvements do not automatically remove external constraints.", "sources": [ "normal-technology-alexander-response-2025", "normal-technology-guide-2025" ] }, { "title": "Testing the assumptions", "text": "Their August 2026 research summary describes two limited evaluations of open-ended AI research. It acknowledges small samples and possible evaluator bias, so the results do not establish a permanent capability limit.", "sources": [ "normal-technology-research-2026" ] } ] } }, { "id": "kapoor", "name": "Sayash Kapoor", "kind": "person", "status": "placed", "position": { "x": 0, "y": -20, "xRange": [ -35, 45 ], "yRange": [ -55, 30 ] }, "confidence": "medium", "basis": "primary", "scope": "Joint public arguments about frontier development, AI control and catastrophic risk, including the September 2026 update.", "caveat": "The reviewed arguments are jointly authored. Pauses concern particular experiments; the overall pace preference remains conditional. Coordinates summarize public arguments, not private probabilities.", "evidence": [ { "axis": "pace", "source": "normal-technology-control-2026", "note": "Calls for organizational oversight and for pausing experiments when needed to put it in place.", "locator": "Part 1: Existing organizational governance norms would have prevented the incident", "role": "support" }, { "axis": "concern", "source": "normal-technology-control-2026", "note": "Says catastrophic risks are not imminent but are increasing as defenses and policy lag; safety is not on track.", "locator": "Part 3: Is AI safety on track?", "role": "support" }, { "axis": "pace", "source": "normal-technology-software-work-2026", "note": "The earlier June essay favored accountability and control over slowing technical capability development.", "locator": "The decide-execute-deliver discussion, before Vibe coding is not agentic engineering", "role": "counterpoint" }, { "axis": "context", "source": "normal-technology-control-2026", "note": "Acknowledges underestimating risks during development and companies' failures to take basic precautions.", "locator": "Part 3: Do risks arise from development or deployment?; The continuity hypothesis", "role": "counterpoint" } ], "implementation": "public-advocacy", "lastReviewed": "2026-09-15", "label": { "preferred": "right" }, "profile": { "summary": "Coauthors the normal-technology view. Now supports targeted experiment pauses and warns that safety efforts are falling behind.", "sources": [ "normal-technology-2025", "normal-technology-control-2026" ], "sections": [ { "title": "A contested account of control", "text": "Scott Alexander argues that rapid self-improvement and adoption by AI labs could defeat the thesis's assumed limits. Narayanan and Kapoor reply that technical improvements do not automatically remove external constraints.", "sources": [ "normal-technology-alexander-response-2025", "normal-technology-guide-2025" ] }, { "title": "Testing the assumptions", "text": "Their August 2026 research summary describes two limited evaluations of open-ended AI research. It acknowledges small samples and possible evaluator bias, so the results do not establish a permanent capability limit.", "sources": [ "normal-technology-research-2026" ] } ] } }, { "id": "daniel-card", "name": "Daniel Card (mRr3b00t)", "kind": "person", "status": "unplaced", "position": null, "confidence": null, "basis": "mixed", "scope": "Public commentary by @UK_Daniel_Card about AI systems, security monitoring and practical risks.", "caveat": "Unplaced: the reviewed material does not establish a preference on frontier capability growth or a sufficiently clear overall catastrophic-risk position. Selected X text was read in a supplied export. The live originals and their authorship were not independently verified.", "evidence": [ { "axis": "context", "source": "card-bcs-identity", "note": "His presentation names Daniel Card and gives the handle @Uk_Daniel_Card.", "locator": "Title slide", "role": "context" }, { "axis": "context", "source": "card-x-system-controls-2026", "note": "Argues for monitoring and controls around the computer system, beyond model-level safeguards.", "locator": "Exported post text; original X page inaccessible", "role": "support" }, { "axis": "context", "source": "card-x-monitor-actions-2026", "note": "Prioritizes observing outputs and actions when assessing immediate security consequences.", "locator": "Exported reply text; original X page inaccessible", "role": "support" }, { "axis": "context", "source": "card-x-fit-task-2026", "note": "Questions adding LLM services when the task calls for more consistent behavior.", "locator": "Exported post text; original X page inaccessible", "role": "support" }, { "axis": "context", "source": "card-x-risk-qualification-2026", "note": "Says criticism of one claim should not be read as asserting its opposite, and describes treating computers as alive as a societal risk.", "locator": "Exported post text; original X page inaccessible", "role": "counterpoint" }, { "axis": "context", "source": "card-pwndefend-risk-framing-2026", "note": "An AI-assisted article he endorses criticizes the digital-nuclear-weapon analogy while describing AI as amplifying existing harms.", "locator": "AI is a digital nuke; closing authorship disclosure", "role": "context" }, { "axis": "context", "source": "card-x-opportunity-2026", "role": "support", "locator": "Exported post text; original X page inaccessible", "note": "Endorses a quoted call to include cybersecurity practitioners while supporting AI opportunities. This does not specify a preferred pace for frontier development." }, { "axis": "context", "source": "card-x-system-tradeoffs-2026", "role": "counterpoint", "locator": "Exported reply text; original X page inaccessible", "note": "Describes both opportunities and risks, including energy costs and deploying systems before addressing security. His software framing acknowledges differences in system design." } ], "implementation": "public-commentary-and-self-reported-testing", "lastReviewed": "2026-09-15", "label": { "preferred": "right" }, "profile": { "summary": "Posts attributed to Card in a supplied export emphasize useful AI tools, practical security and skepticism toward alarming claims. They also acknowledge risks and differences between computer systems. His preferred pace of frontier development remains unclear, so there is no map dot.", "sources": [ "card-x-system-controls-2026", "card-x-opportunity-2026", "card-x-system-tradeoffs-2026", "card-pwndefend-risk-framing-2026" ], "sections": [ { "title": "What the posts support", "text": "The exported posts ask which actions a system can take, how those actions can be observed, and whether an LLM suits the task. Attribution follows the export author field and remains unverified against the original posts. Allegations about particular laboratories and claims from personal demonstrations are not adopted as technical findings.", "sources": [ "card-x-system-controls-2026", "card-x-monitor-actions-2026", "card-x-fit-task-2026" ] }, { "title": "What he says about risk", "text": "Our reading is that the reviewed commentary challenges dramatic AI-risk narratives and emphasizes practical harms. His endorsed, AI-assisted PwnDefend article describes AI as amplifying existing risks. An exported post also warns against reading criticism of one claim as support for its opposite. This does not establish that he dismisses every catastrophic scenario.", "sources": [ "card-pwndefend-risk-framing-2026", "card-x-risk-qualification-2026" ] }, { "title": "What he says about development", "text": "One exported post explicitly agrees with a quoted call to involve cybersecurity practitioners and support AI opportunities. Another discusses benefits for prototypes alongside energy costs and risks of deploying systems before addressing security. These statements concern how people use and build systems. They do not specify how quickly the most capable models should advance.", "sources": [ "card-x-opportunity-2026", "card-x-system-tradeoffs-2026" ] }, { "title": "A qualification about control", "text": "Observing a system's actions does not establish that its safeguards will remain effective as capabilities grow. Narayanan and Kapoor argue that control methods need continued research and investment. The software framing in the exported posts alone does not settle that question.", "sources": [ "card-x-system-controls-2026", "normal-technology-control-2026" ] }, { "title": "Why there is no dot", "text": "The missing frontier-development position prevents a two-axis placement. A dot in the middle would imply a preference we have not established. The risk commentary is described above with its limits; it is not a numerical risk estimate or evidence of membership in a movement.", "sources": [ "card-x-opportunity-2026", "card-x-system-tradeoffs-2026", "card-x-risk-qualification-2026" ] } ] } } ] }, "sources": [ { "id": "trump-forbes-2026-09-14", "title": "Trump And Vance Suggest AI Slowdown Conspiracy: Feels Like A ‘Trojan Horse’", "publisher": "Forbes / Conor Murray", "url": "https://www.forbes.com/sites/conormurray/2026/09/14/trump-says-calls-for-ai-regulation-are-a-hoax-compares-it-to-russia-impeachment-scams/", "published": "2026-09-14", "checkedOn": "2026-09-15", "kind": "reporting", "verification": "read", "notes": "Read the publisher's accessible article, including Key Facts and Trump Dismisses AI Concerns; retrieval redirected to tollbit.forbes.com. It links September 14 Truth Social posts 117270591511950591 and 117269745153543631. Their original pages returned a JavaScript shell or browser security check; original post text was not independently retrieved. Vance's statements and reporters' claims about other actors are not assigned to Trump. The title differs between the indexed and retrieved versions.", "retrieval": { "method": "publisher-page", "scope": "The publisher article, including its Key Facts and Trump Dismisses AI Concerns sections. Original social-media text was not independently retrieved." }, "archive": { "status": "not-verified", "checkedOn": "2026-09-15", "notes": "The Wayback availability lookup returned HTTP 429 (rate limited). No capture was verified; this does not show that no archived copy exists." }, "originals": [ { "url": "https://truthsocial.com/@realDonaldTrump/posts/117270591511950591", "label": "Donald Trump · Truth Social post 117270591511950591", "status": "not-retrieved", "checkedOn": "2026-09-15", "notes": "The linked Truth Social page exposed a JavaScript prompt without the post text. The reporting remains the material read.", "archive": { "status": "not-verified", "checkedOn": "2026-09-15", "notes": "The Wayback availability lookup returned HTTP 429 (rate limited). No capture was verified; this does not show that no archived copy exists." } }, { "url": "https://truthsocial.com/@realDonaldTrump/posts/117269745153543631", "label": "Donald Trump · Truth Social post 117269745153543631", "status": "not-retrieved", "checkedOn": "2026-09-15", "notes": "The linked Truth Social page exposed a JavaScript prompt without the post text. The reporting remains the material read.", "archive": { "status": "not-verified", "checkedOn": "2026-09-15", "notes": "The Wayback availability lookup returned HTTP 429 (rate limited). No capture was verified; this does not show that no archived copy exists." } } ] }, { "id": "trump-afp-2026-09-14", "title": "'I am the hoax buster': Trump rejects AI danger warnings", "publisher": "Agence France-Presse", "url": "https://www.afp.com/en/i-am-hoax-buster-trump-rejects-ai-danger-warnings", "published": "2026-09-14", "checkedOn": "2026-09-15", "kind": "reporting", "verification": "read", "notes": "Read AFP's own September 14 article, displayed at 23:36 with no time zone specified. Used for attributed statements about catastrophic AI risk and guardrails, not as direct verification of Truth Social posts or independent support for all contextual claims in the article.", "retrieval": { "method": "publisher-page", "scope": "AFP's report of the statements, not the original social-media posts." }, "archive": { "status": "not-verified", "checkedOn": "2026-09-15", "notes": "The Wayback availability lookup returned HTTP 429 (rate limited). No capture was verified; this does not show that no archived copy exists." } }, { "id": "trump-ai-security-order-2026", "title": "Promoting Advanced Artificial Intelligence Innovation and Security (Executive Order 14409)", "publisher": "The White House / Donald J. Trump", "url": "https://www.whitehouse.gov/presidential-actions/2026/06/promoting-advanced-artificial-intelligence-innovation-and-security/", "published": "2026-06-02", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the signed order, sections 1–5, directly on the White House site. Supports its stated policy and directions; does not independently establish completed implementation, security effectiveness, current legal status, or Trump's private beliefs. The voluntary model-evaluation process is distinct from an industry-wide training pause." }, { "id": "musk-dwarkesh-2026", "title": "Elon Musk — In 36 months, the cheapest place to put AI will be space", "publisher": "Dwarkesh Podcast / Dwarkesh Patel and John Collison", "url": "https://www.dwarkesh.com/p/elon-musk", "published": "2026-02-05", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read scaling discussion and the 00:36:46–00:59:56 Grok/alignment section of the host transcript. These are Musk's claims and plans, not verified engineering results or proof of alignment." }, { "id": "musk-lex-2023", "title": "Elon Musk interview #400 — transcript", "publisher": "Lex Fridman Podcast", "url": "https://lexfridman.com/elon-musk-4-transcript/", "published": "2023-11-10", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read 01:23:13–01:29:39 on scaling, delayed open sourcing and Musk's account of safety disagreements. Historical self-report; claims about other people's motives are not adopted." }, { "id": "thiel-tyler-2024", "title": "Peter Thiel on Political Theology (Ep. 210)", "publisher": "Conversations with Tyler / Mercatus Center", "url": "https://conversationswithtyler.com/episodes/peter-thiel-political-theology/", "published": "2024-04-17", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Host transcript, recorded February 21, 2024. Read existential-risk discussion and audience questions on stopping AI, open source and substitution for workers. Political judgments are attributed to Thiel." }, { "id": "thiel-hoover-apocalypse", "title": "Part II: Apocalypse Now? Peter Thiel on Ancient Prophecies and Modern Tech", "publisher": "Hoover Institution / Uncommon Knowledge", "url": "https://www.hoover.org/plus/research/part-ii-apocalypse-now-peter-thiel-ancient-prophecies-and-modern-tech", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the host transcript, including the Scylla and Charybdis discussion. The page identifies recording on October 8, 2024, but no reliable publication date. Recording date is not substituted for publication. His theological framing is speculative." }, { "id": "thiel-spectator-2025", "title": "I've been allergic to AI for a long time: an interview with Peter Thiel", "publisher": "The Spectator / William Atkinson, Lara Brown and John Power", "url": "https://www.spectator.com.au/2025/12/ive-been-allergic-to-ai-for-a-long-time-an-interview-with-peter-thiel/", "published": "2025-12-13", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the edited interview's AI discussion: growth, concentrated returns, labor substitution and skepticism of the label. The interview establishes his stated outlook, not its economic predictions." }, { "id": "thiel-spectator-2026", "title": "Can Peter Thiel stop the Antichrist?", "publisher": "The Spectator / Lara Brown and John Power", "url": "https://spectator.com/?edition=us&p=643746", "published": "2026-02-07", "checkedOn": "2026-09-15", "kind": "reporting", "verification": "read", "notes": "Read the reporters' account of Thiel's Cambridge talk, in the February 7 issue. Supports continuity of his political-theological concern; no complete talk transcript was retrieved." }, { "id": "amodei-pacing", "title": "We Must Pace the Frontier", "publisher": "Dario Amodei", "url": "https://darioamodei.com/post/we-must-pace-the-frontier", "published": "2026-09", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "The essay itself specifies the month. The September 12 publication day is corroborated by the September 14 press roundup. A proposal and a commitment are not independent verification of implementation." }, { "id": "openai-pacing", "title": "Pacing model development in an era of cyber-critical capabilities", "publisher": "OpenAI", "url": "https://openai.com/index/pacing-model-development-cyber-capabilities/", "published": "2026-08-18", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "A company account of selected research-workload restrictions, not evidence of a whole-company training halt." }, { "id": "september-responses", "title": "What execs and politicians are saying about slowing down AI development", "publisher": "The Verge", "url": "https://www.theverge.com/ai-artificial-intelligence/995141/ai-executives-politicians-safety-regulation-anthropic-dario-amodei", "published": "2026-09-14", "checkedOn": "2026-09-15", "kind": "reporting", "verification": "read", "notes": "Article read. Linked X posts could not be fetched independently. Their content is therefore treated as reported, not directly verified.", "retrieval": { "method": "publisher-page", "scope": "The article's reporting on Altman, Hassabis, Musk and LeCun, including links to their posts. The original post text was not independently retrieved." }, "archive": { "status": "not-verified", "checkedOn": "2026-09-15", "notes": "The Wayback availability lookup returned HTTP 429 (rate limited). No capture was verified; this does not show that no archived copy exists." }, "originals": [ { "url": "https://x.com/sama/status/2098811563415150910", "label": "Sam Altman · X post 2098811563415150910", "status": "not-retrieved", "checkedOn": "2026-09-15", "notes": "The linked X page returned HTTP 403. The reporting remains the material read; the original text was not independently retrieved.", "archive": { "status": "not-verified", "checkedOn": "2026-09-15", "notes": "The Wayback availability lookup returned HTTP 429 (rate limited). No capture was verified; this does not show that no archived copy exists." } }, { "url": "https://x.com/sama/status/2099348812305473766", "label": "Sam Altman · X post 2099348812305473766", "status": "not-retrieved", "checkedOn": "2026-09-15", "notes": "The linked X page returned HTTP 403. The reporting remains the material read; the original text was not independently retrieved." }, { "url": "https://x.com/sama/status/2099352016988614852", "label": "Sam Altman · X post 2099352016988614852", "status": "not-retrieved", "checkedOn": "2026-09-15", "notes": "The linked X page returned HTTP 403. The reporting remains the material read; the original text was not independently retrieved." }, { "url": "https://x.com/demishassabis/status/2098909516582490602", "label": "Demis Hassabis · X post 2098909516582490602", "status": "not-retrieved", "checkedOn": "2026-09-15", "notes": "The linked X page returned HTTP 403. The reporting remains the material read; the original text was not independently retrieved." }, { "url": "https://x.com/elonmusk/status/2098789109980332057", "label": "Elon Musk · X post 2098789109980332057", "status": "not-retrieved", "checkedOn": "2026-09-15", "notes": "The linked X page returned HTTP 403. The reporting remains the material read; the original text was not independently retrieved.", "archive": { "status": "not-verified", "checkedOn": "2026-09-15", "notes": "The Wayback availability lookup returned HTTP 429 (rate limited). No capture was verified; this does not show that no archived copy exists." } }, { "url": "https://x.com/ylecun/status/2099248236074545576", "label": "Yann LeCun · X post 2099248236074545576", "status": "not-retrieved", "checkedOn": "2026-09-15", "notes": "The linked X page returned HTTP 403. The reporting remains the material read; the original text was not independently retrieved." } ] }, { "id": "hassabis-framework", "title": "A Framework for Frontier AI and the Dawning of a New Age", "publisher": "Demis Hassabis", "url": "https://demishassabis.substack.com/p/a-framework-for-frontier-ai-and-the-dawning-of-a-new-age", "published": "2026-07-14", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Includes a proposed standards body that could coordinate a slowdown if necessary." }, { "id": "deepmind-framework", "title": "Strengthening our Frontier Safety Framework", "publisher": "Google DeepMind", "url": "https://deepmind.google/blog/strengthening-our-frontier-safety-framework/", "published": "2025-09-22", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "The article contains an April 2026 update. A published framework is evidence of stated policy, not an estimate of catastrophe probability.", "updated": "2026-04-17" }, { "id": "microsoft-code", "title": "Humanist AI in practice: A public consultation on our Code of Conduct for MAI Models", "publisher": "Microsoft AI", "url": "https://microsoft.ai/news/mai-code-of-conduct/", "published": "2026-09-14", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "A consultation and intended training/evaluation standards. Scope is Microsoft AI, not every Microsoft business." }, { "id": "meta-framework", "title": "Scaling How We Build and Test Our Most Advanced AI", "publisher": "Meta AI", "url": "https://ai.meta.com/blog/scaling-how-we-build-test-advanced-ai/", "published": "2026-04-08", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "The HTML article describes catastrophic-risk domains and release safeguards. No catastrophe probability is inferred." }, { "id": "andreessen-manifesto", "title": "The Techno-Optimist Manifesto", "publisher": "Andreessen Horowitz / Marc Andreessen", "url": "https://a16z.com/the-techno-optimist-manifesto/", "published": "2023-10-16", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the manifesto and footer. The footer explicitly attributes posts to individual personnel and excludes the views of a16z Capital Management and affiliates. Historical personal advocacy, not institutional evidence." }, { "id": "yudkowsky-book", "title": "Yudkowsky and Soares Announce Major New Book: If Anyone Builds It, Everyone Dies", "publisher": "Machine Intelligence Research Institute", "url": "https://intelligence.org/2025/05/15/yudkowsky-and-soares-announce-major-new-book-if-anyone-builds-it-everyone-dies/", "published": "2025-05-15", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "A publisher/affiliated announcement of the authors' argument, not an empirical finding that extinction is certain.", "retrieval": { "method": "publisher-page", "scope": "The affiliated announcement of the authors' argument, not the book itself." }, "archive": { "status": "not-verified", "checkedOn": "2026-09-15", "notes": "The Wayback availability lookup returned HTTP 429 (rate limited). No capture was verified; this does not show that no archived copy exists." } }, { "id": "miri-position", "title": "Our view", "publisher": "Machine Intelligence Research Institute", "url": "https://intelligence.org/", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Undated live institutional position checked on the review date. Used as context, not as a substitute for individual authorship." }, { "id": "pauseai-proposal", "title": "PauseAI Proposal", "publisher": "PauseAI", "url": "https://pauseai.info/proposal", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "A proposed international pause. The page is undated; checkedOn is not a publication date." }, { "id": "superintelligence-statement", "title": "Statement on Superintelligence", "publisher": "Future of Life Institute", "url": "https://superintelligence-statement.org/", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "The statement text was accessible. The dynamic signatory list was not exposed in the retrieved text; Bengio's signature is corroborated separately." }, { "id": "bengio-signature", "title": "Harry and Meghan join AI pioneers in call for ban on superintelligent systems", "publisher": "The Guardian", "url": "https://www.theguardian.com/technology/2025/oct/22/harry-and-meghan-join-ai-pioneers-call-ban-superintelligent-systems", "published": "2025-10-22", "checkedOn": "2026-09-15", "kind": "reporting", "verification": "read", "notes": "Used only to corroborate Yoshua Bengio's signature and the scope of that statement, not to attribute every listed concern to him." }, { "id": "lecun-interview", "title": "Yann LeCun: Meta AI, Open Source, Limits of LLMs, AGI & the Future of AI, transcript #416", "publisher": "Lex Fridman Podcast", "url": "https://lexfridman.com/yann-lecun-3-transcript/", "published": "2024-03-07", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "First-person interview transcript. The host notes that transcription errors are possible. Current opposition to the pacing essay is reported separately.", "retrieval": { "method": "publisher-page", "scope": "The published interview transcript. Transcription errors are possible." }, "archive": { "status": "not-verified", "checkedOn": "2026-09-15", "notes": "The Wayback availability lookup returned HTTP 429 (rate limited). No capture was verified; this does not show that no archived copy exists." } }, { "id": "ea-definition", "title": "What is effective altruism?", "publisher": "Effective Altruism", "url": "https://www.effectivealtruism.org/articles/introduction-to-effective-altruism", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Definition of a philosophy/community, not a placement on the AI chart." }, { "id": "eacc-definition", "title": "what the f* is e/acc", "publisher": "e/acc newsletter", "url": "https://effectiveaccelerationism.substack.com/p/what-the-f-is-eacc", "published": "2022-12-26", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Movement self-description. Participants need not share a single risk estimate." }, { "id": "eacc-tenets", "title": "Notes on e/acc principles and tenets", "publisher": "e/acc newsletter / BasedBeffJezos and bayeslord", "url": "https://effectiveaccelerationism.substack.com/p/repost-notes-on-eacc-principles-and", "published": "2022-10-31", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Date belongs to the repost. Read as a movement's own philosophical claims, not independent scientific validation. Includes adversarial usage of decel." }, { "id": "verdon-interview", "title": "Guillaume Verdon: e/acc, AI doomers and effective altruism (transcript #407)", "publisher": "Lex Fridman Podcast", "url": "https://lexfridman.com/guillaume-verdon-transcript/", "published": "2023-12-29", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the AI doomers and effective altruism sections as examples of contested terminology. An advocate's description of opponents is not a neutral definition of their views. Transcript may contain errors." }, { "id": "dacc-original", "title": "My techno-optimism", "publisher": "Vitalik Buterin", "url": "https://vitalik.eth.limo/general/2023/11/27/techno_optimism.html", "published": "2023-11-27", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read via the author's eth.limo site because vitalik.ca did not load. Introduces defensive, differential and decentralization-focused acceleration. Used for ideas, not a new placement of the author." }, { "id": "dacc-update", "title": "d/acc: one year later", "publisher": "Vitalik Buterin", "url": "https://vitalik.eth.limo/general/2025/01/05/dacc2.html", "published": "2025-01-05", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the definition and distinctions: defensive development combined with distributed and democratic control. Hypothetical future examples are not observed outcomes." }, { "id": "longtermism-definition", "title": "Longtermism", "publisher": "William MacAskill", "url": "https://www.williammacaskill.com/longtermism", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "An advocate's account of the moral importance of future people. The page is undated; the book's publication date is not used as the page date." }, { "id": "pauseai-proposal-2026", "title": "PauseAI Proposal (April 2026 version)", "publisher": "PauseAI", "url": "https://pauseai.org/proposal", "published": "2026-04-05", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the April 5th, 2026 version, including its global pause, training approvals and democratic-control conditions. Used for the movement's published proposal and terminology; its predictions are attributed advocacy.", "retrieval": { "method": "publisher-page", "scope": "The April 2026 proposal, including its pause, training-approval and democratic-control conditions." }, "archive": { "status": "not-verified", "checkedOn": "2026-09-15", "notes": "The Wayback availability lookup returned HTTP 429 (rate limited). No capture was verified; this does not show that no archived copy exists." } }, { "id": "cais-risk", "title": "Statement on AI Extinction Risk", "publisher": "Center for AI Safety", "url": "https://aistatement.com/work/statement-on-ai-extinction-risk", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the statement after following the safe.ai redirect. Establishes advocacy for treating extinction risk as a priority, not certainty of catastrophe or a common numerical probability. No new signatory claims are inferred." }, { "id": "anthropic-rsp-3-4", "title": "Responsible Scaling Policy, version 3.4", "publisher": "Anthropic", "url": "https://cdn.sanity.io/files/4zrzovbb/website/0bacdc8440ea96e62a8766d99ebe1d4eea6d5f3a.pdf", "published": "2026-07-08", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read introduction, risk-report requirements, external-review provisions and Appendix A. July 8 is the stated effective date. Separates company commitments from industry recommendations; not an implementation audit." }, { "id": "bengio-lawzero-2025", "title": "Introducing LawZero", "publisher": "Yoshua Bengio", "url": "https://yoshuabengio.org/en/blog/introducing-lawzero", "published": "2025-06-03", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the author's announcement, including his loss-of-control concerns. Describes his argument and research goals, not proof that the proposed approach is safe." }, { "id": "extreme-ai-risks-2024", "title": "Managing extreme AI risks amid rapid progress (version 3)", "publisher": "Yoshua Bengio, Geoffrey Hinton, Stuart Russell and coauthors / arXiv", "url": "https://arxiv.org/html/2310.17688v3", "published": "2023-10-26", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read authors, societal-scale risks and governance/mitigation sections of version 3. Coauthorship supports a shared proposal, not identical personal beliefs or a present-day forecast.", "updated": "2024-05-22" }, { "id": "russell-senate-2023", "title": "Written statement for the Senate AI Forum on Risk, Alignment, & Guarding Against Doomsday Scenarios", "publisher": "Stuart Russell / UC Berkeley", "url": "https://people.eecs.berkeley.edu/~russell/papers/russell-senate23b-statement.pdf", "published": "2023-12-06", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read executive summary and control-risk discussion. The document is dated December 6 and draws on earlier July testimony; the July date is not this document's date." }, { "id": "hinton-gzero-2025", "title": "The human cost of AI, with Geoffrey Hinton", "publisher": "GZERO Media", "url": "https://www.gzeromedia.com/amp/ai-human-cost-geoffrey-hinton-2674373827", "published": "2025-12-06", "checkedOn": "2026-09-15", "kind": "reporting", "verification": "read", "notes": "Read the interview publisher's summary and attributed control-risk quotation. The full audio/transcript was not reviewed; this record is reporting, not direct verification of every interview claim." }, { "id": "gebru-torres-2024", "title": "The TESCREAL bundle: Eugenics and the promise of utopia through artificial general intelligence", "publisher": "Timnit Gebru and Émile P. Torres / First Monday", "url": "https://firstmonday.org/ojs/index.php/fm/article/view/13636", "published": "2024-04-14", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the journal abstract and publication metadata only; full-text retrieval failed. Used for the authors' recommendation to study defined tasks. Their wider historical thesis is not adopted as an atlas finding." }, { "id": "bender-humanities-2025", "title": "Resisting Dehumanization in the Age of “AI”: The View from the Humanities", "publisher": "Emily M. Bender / Illinois State University talk", "url": "https://faculty.washington.edu/ebender/papers/Bender-ISU-2025.pdf", "published": "2025-04-16", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the author's slides, especially the opening and slides 25–33 on scrutiny, rights and teaching. April 16 is the talk date printed on the slides. Does not establish a frontier-training policy." }, { "id": "fei-fei-li-un-2024", "title": "Artificial Intelligence and the Maintenance of International Peace and Security", "publisher": "Fei-Fei Li / Stanford HAI", "url": "https://hai.stanford.edu/assets/files/fei-fei-li-un-security-council-briefing.pdf", "published": "2024-12-19", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read all three pages of the briefing, presented December 19, 2024. Its footnote explicitly says the views are individual and do not represent affiliated organizations." }, { "id": "whittaker-ndss-2024", "title": "AI, Encryption, and the Sins of the 90s", "publisher": "Meredith Whittaker / Signal", "url": "https://signal.org/blog/pdfs/ndss-keynote.pdf", "published": "2024-02-27", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the keynote's AI, corporate-surveillance and concluding privacy arguments, pages 1–3 and 10–14. Date is the speech date. Used as her argument, not as proof that all AI systems have one business model." }, { "id": "eu-ai-office-overview", "title": "European AI Office", "publisher": "European Commission", "url": "https://digital-strategy.ec.europa.eu/en/policies/ai-office", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the Office's mandate, enforcement tasks and innovation role. The page gives a last-update date, not an original publication date. Describes an institutional mandate, not a catastrophe probability.", "updated": "2026-09-08" }, { "id": "uk-aisi-agenda", "title": "AISI Research Agenda", "publisher": "UK AI Security Institute", "url": "https://www.aisi.gov.uk/research-agenda", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the research overview and Autonomous Systems section on catastrophic harm and loss of control. No reliable publication date was exposed. Research priorities do not by themselves establish a preferred development pace." }, { "id": "uk-aisi-name-2025", "title": "Tackling AI security risks to unleash growth and deliver Plan for Change", "publisher": "UK Department for Science, Innovation and Technology", "url": "https://www.gov.uk/government/news/tackling-ai-security-risks-to-unleash-growth-and-deliver-plan-for-change", "published": "2025-02-14", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the announcement of the name change and its security-research focus. A government growth agenda is not automatically the institute's own frontier-pacing position." }, { "id": "deepseek-r1-release", "title": "DeepSeek-R1 Release", "publisher": "DeepSeek", "url": "https://api-docs.deepseek.com/news/news250120/", "published": "2025-01-20", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the release announcement and model-access descriptions. Product availability and benchmark claims do not establish an institutional stance on catastrophic AI risk." }, { "id": "vance-paris-2025", "title": "Remarks at the Artificial Intelligence Action Summit in Paris, France", "publisher": "JD Vance / The American Presidency Project, UC Santa Barbara", "url": "https://www.presidency.ucsb.edu/documents/remarks-the-vice-president-the-artificial-intelligence-action-summit-paris-france", "published": "2025-02-11", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the archived speech transcript, including its closing acknowledgment of safety concerns. Primary speech text hosted by a university archive; claims about competitors or regulations are attributed arguments." }, { "id": "altman-governance-2023", "title": "Governance of superintelligence", "publisher": "Sam Altman, Greg Brockman and Ilya Sutskever / OpenAI", "url": "https://openai.com/index/governance-of-superintelligence/", "published": "2023-05-22", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the coauthored article on frontier growth limits, existential risk and lower-capability exemptions. Historical proposal, not evidence that a limit was implemented." }, { "id": "andreessen-interview-summary-2026", "title": "Beyond P(doom): Marc Andreessen – Betting on America", "publisher": "The a16z Show", "url": "https://a16z.com/podcast/beyond-pdoom-marc-andreessen-betting-on-america/", "published": "2026-06-29", "checkedOn": "2026-09-15", "kind": "reporting", "verification": "read", "notes": "Read the publisher's episode description only. It describes advocacy for AI growth and concern about policy barriers; no full transcript was exposed. June 29 is the podcast publication date, not the June 25 event date." }, { "id": "lw-history", "title": "A Brief History of LessWrong", "publisher": "LessWrong / Ruby", "url": "https://www.lesswrong.com/posts/S69ogAGXcc9EQjpcZ/a-brief-history-of-lesswrong", "published": "2019-06-01", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the site's own retrospective: Overcoming Bias in 2006, LessWrong in 2009, and the Sequences. An internal community account, not an independent history." }, { "id": "lw-welcome", "title": "Welcome to LessWrong!", "publisher": "LessWrong", "url": "https://www.lesswrong.com/posts/bJ2haLkcGeLtTWaD5/welcome-to-lesswrong-1", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the site's introduction and intended reasoning practice. Self-description does not certify participants' accuracy or common beliefs." }, { "id": "lw-rationality", "title": "What Do We Mean By Rationality?", "publisher": "LessWrong / Eliezer Yudkowsky", "url": "https://www.lesswrong.com/posts/RcZCwxFiZzE6X7nsv/what-do-we-mean-by-rationality-1", "published": "2009-03-16", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the epistemic/instrumental distinction and discussion of Bayesian reasoning. A proposed practice, not a verified trait of participants." }, { "id": "basilisk-history", "title": "Roko's Basilisk", "publisher": "LessWrong community wiki", "url": "https://www.lesswrong.com/w/rokos-basilisk", "published": null, "updated": "2022-11-30", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the community retrospective, objections and moderation account. Original deleted post not independently retrieved. Historical claims are attributed to this account." }, { "id": "basilisk-response", "title": "A few misconceptions surrounding Roko's basilisk", "publisher": "LessWrong / Rob Bensinger", "url": "https://www.lesswrong.com/posts/WBJZoeJypcNRmsdHx/a-few-misconceptions-surrounding-roko-s-basilisk", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the community response and quoted objections in the initial retrieval; a later fetch timed out. Used for disputed assumptions, not proof about hypothetical agents. Exact publication date was not independently established." }, { "id": "tdt-paper", "title": "Timeless Decision Theory", "publisher": "Eliezer Yudkowsky / MIRI", "url": "https://intelligence.org/files/TDT.pdf", "published": "2010", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read abstract and opening Newcomb discussion. Year follows the paper's citation. Proposed framework and formal examples, not an experimentally settled rule." }, { "id": "fdt-paper", "title": "Functional Decision Theory: A New Theory of Instrumental Rationality", "publisher": "Eliezer Yudkowsky and Nate Soares / arXiv", "url": "https://arxiv.org/abs/1710.05060", "published": "2017", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the authors' abstract and MIRI's October 22, 2017 introduction. Advantages are claimed within specified decision problems." }, { "id": "simulation-paper", "title": "Are You Living in a Computer Simulation?", "publisher": "Nick Bostrom / Philosophical Quarterly", "url": "https://simulation-argument.com/simulation/", "published": "2003", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read abstract and introduction. Publication year differs from the first draft in 2001. Conditional philosophical argument, not a finding that our world is simulated." }, { "id": "infohazards-paper", "title": "Information Hazards: A Typology of Potential Harms from Knowledge", "publisher": "Nick Bostrom", "url": "https://nickbostrom.com/information-hazards.pdf", "published": "2011", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read definition and taxonomy. Applying the concept to a particular claim requires additional evidence." }, { "id": "agi-levels", "title": "Levels of AGI for Operationalizing Progress on the Path to AGI", "publisher": "Meredith Ringel Morris and coauthors / arXiv", "url": "https://arxiv.org/abs/2311.02462", "published": "2023-11-04", "updated": "2025-09-24", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read abstract, Table 1 and autonomy discussion. Proposed distinctions between breadth, performance and autonomy; no current model classification adopted." }, { "id": "advanced-ai-ethics", "title": "Ethical Issues in Advanced Artificial Intelligence", "publisher": "Nick Bostrom", "url": "https://nickbostrom.com/ethics/ai", "published": "2003", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read prospective discussion of superintelligence and paperclip goals. Philosophical scenarios, not observations of current systems." }, { "id": "superintelligent-will", "title": "The Superintelligent Will: Motivation and Instrumental Rationality in Advanced Artificial Agents", "publisher": "Nick Bostrom / Minds and Machines", "url": "https://nickbostrom.com/superintelligentwill.pdf", "published": "2012-05", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read abstract and formulations of orthogonality and instrumental convergence. Theoretical theses with qualifications, not measured outcomes for a learning system." }, { "id": "learned-optimization", "title": "Risks from Learned Optimization in Advanced Machine Learning Systems", "publisher": "Evan Hubinger and coauthors / arXiv", "url": "https://arxiv.org/abs/1906.01820", "published": "2019-06-05", "updated": "2021-12-01", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read abstract and revision history introducing mesa-optimization and the relation between learned and training objectives." }, { "id": "concrete-safety", "title": "Concrete Problems in AI Safety", "publisher": "Dario Amodei and coauthors / arXiv", "url": "https://arxiv.org/abs/1606.06565", "published": "2016-06-21", "updated": "2016-07-25", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract's five accident-risk problems including reward hacking and distributional shift. Does not give a general catastrophe probability." }, { "id": "goodhart-paper", "title": "Categorizing Variants of Goodhart's Law", "publisher": "David Manheim and Scott Garrabrant / arXiv", "url": "https://arxiv.org/abs/1803.04585", "published": "2018-03-13", "updated": "2019-02-24", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read abstract distinguishing overoptimization mechanisms. A taxonomy, not a claim that every metric fails." }, { "id": "transhumanism-declaration", "title": "The Transhumanist Declaration", "publisher": "Humanity+", "url": "https://www.humanityplus.org/the-transhumanist-declaration", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the declaration's commitments to enhancement, risk reduction and choice. Live page publication date unspecified." }, { "id": "existential-risks", "title": "Existential Risks: Analyzing Human Extinction Scenarios and Related Hazards", "publisher": "Nick Bostrom", "url": "https://nickbostrom.com/papers/existential-risks/", "published": "2002", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the definition distinguishing existential from other risks. Does not establish the likelihood of a specific scenario." }, { "id": "unesco-ethics", "title": "Ethics of Artificial Intelligence", "publisher": "UNESCO", "url": "https://www.unesco.org/en/artificial-intelligence/recommendation-ethics", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read overview of rights, dignity, fairness and oversight. Recommendation adoption date is not assigned to the live page." }, { "id": "vinge-singularity", "title": "Technological Singularity", "publisher": "Vernor Vinge / Carnegie Mellon University archive", "url": "https://frc.ri.cmu.edu/~hpm/book98/com.ch1/vinge.singularity.html", "published": "1993", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the author's essay via the academic archive after NASA retrieval failed. Speculative scenarios and historical forecasts, not a demonstrated trajectory." }, { "id": "utilitarianism-intro", "title": "Introduction to Utilitarianism", "publisher": "Utilitarianism.net", "url": "https://utilitarianism.net/introduction-to-utilitarianism/", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read an advocate-authored philosophical introduction. Defines the view; does not establish its correctness or universal adoption among effective altruists." }, { "id": "pause-letter-2023", "title": "Pause Giant AI Experiments: An Open Letter", "publisher": "Future of Life Institute", "url": "https://futureoflife.org/open-letter/pause-giant-ai-experiments/", "published": "2023-03-22", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the six-month request concerning systems more powerful than GPT-4. A historical advocacy milestone, not evidence of an implemented halt." }, { "id": "transformer-paper", "title": "Attention Is All You Need", "publisher": "Ashish Vaswani and coauthors / arXiv", "url": "https://arxiv.org/html/1706.03762v7", "published": "2017-06-12", "updated": "2023-08-02", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read abstract, model architecture, learned embeddings and next-token probabilities in the HTML paper; publication and revision dates checked against the arXiv abstract page. This is the original Transformer architecture, not a claim that every current LLM has its exact structure." }, { "id": "hf-generation", "title": "Generation strategies", "publisher": "Hugging Face Transformers documentation", "url": "https://huggingface.co/docs/transformers/en/generation_strategies", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read greedy search, multinomial sampling and the custom generation loop separating model logits from token selection. Used to explain decoding choices; library defaults do not establish the behavior of every hosted chatbot. The live page does not establish a publication date." }, { "id": "pytorch-reproducibility", "title": "Reproducibility", "publisher": "PyTorch documentation", "url": "https://docs.pytorch.org/docs/2.14/notes/randomness.html", "published": "2026-05-14", "updated": "2026-05-14", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the versioned page reached from the stable documentation: cross-release/platform limits, random seeds and deterministic algorithms. Dates follow the page's displayed Created On and Last Updated On fields, not the historical first publication of PyTorch's reproducibility guidance. These are framework constraints, not measurements of a particular chatbot service." }, { "id": "circuit-tracing", "title": "Circuit Tracing: Revealing Computational Graphs in Language Models", "publisher": "Emmanuel Ameisen and coauthors / Anthropic, Transformer Circuits", "url": "https://transformer-circuits.pub/2025/attribution-graphs/methods.html", "published": "2025-03-27", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read introduction, method overview and limitations including reconstruction errors, graph complexity, global circuits and mechanistic faithfulness. The authors' replacement-model analyses reveal selected mechanisms; they do not provide a complete explanation of all behavior. Later attention-tracing work is cited alongside this paper to avoid treating its missing-attention limitation as a permanent field-wide result." }, { "id": "attention-tracing", "title": "Tracing Attention Computation Through Feature Interactions", "publisher": "Harish Kamath and coauthors / Anthropic, Transformer Circuits", "url": "https://transformer-circuits.pub/2025/attention-qk/index.html", "published": "2025-07-31", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Direct browser-tool retrieval failed; fetched the original publisher HTML successfully and read the introduction, case-study summaries, QK-attribution method, inhibitory-effect limitation and graph-construction tradeoffs. Extends earlier attribution graphs to attention; results are selected studies with open questions, not a complete model explanation." }, { "id": "copilot-code-review", "title": "Application card: GitHub Copilot inline suggestions", "publisher": "GitHub Docs", "url": "https://docs.github.com/en/copilot/responsible-use/inline-suggestions", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read LLM definition, code-generation workflow, generated-test caveats, inaccurate-code limitations and review/testing guidance. Vendor documentation establishes intended use and acknowledged limitations, not independent accuracy rates. Live page publication date unspecified." }, { "id": "weather-uncertainty", "title": "Quantifying forecast uncertainty", "publisher": "European Centre for Medium-Range Weather Forecasts", "url": "https://www.ecmwf.int/en/research/modelling-and-prediction/quantifying-forecast-uncertainty", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the forecast-uncertainty explanation, initial-condition uncertainty and numerical-model approximations. Cited only for the weather side of an explicitly editorial analogy; it supplies no evidence that LLMs are meteorological or chaotic systems. Live page publication date unspecified." }, { "id": "ncsc-secure-ai", "title": "Guidelines for secure AI system development", "publisher": "UK National Cyber Security Centre and international partners", "url": "https://www.ncsc.gov.uk/collection/guidelines-secure-ai-system-development", "published": "2023-11-27", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read executive summary and lifecycle structure. Guidance addresses complete AI systems and recommends security throughout design, development, deployment and operation. Recommendations are not evidence of any organization's implementation." }, { "id": "ncsc-ai-design", "title": "Guidelines for secure AI system development: Secure design", "publisher": "UK National Cyber Security Centre", "url": "https://www.ncsc.gov.uk/collection/guidelines-secure-ai-system-development/guidelines/secure-design", "published": "2023-11-27", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read threat modelling, task suitability, model selection, restricted actions and least privilege. Date follows the containing guideline publication. Used for design principles, not a certificate that any configuration is safe." }, { "id": "ncsc-ai-operations", "title": "Guidelines for secure AI system development: Secure operation and maintenance", "publisher": "UK National Cyber Security Centre", "url": "https://www.ncsc.gov.uk/collection/guidelines-secure-ai-system-development/guidelines/secure-operation-maintenance", "published": "2023-11-27", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read monitoring of behavior and inputs, privacy-aware logging, update evaluation and lessons learned. Date follows the containing guideline publication. Describes operational monitoring, not complete explanation of learned weights." }, { "id": "ncsc-agentic-risk", "title": "Managing the cyber risk of agentic AI", "publisher": "UK National Cyber Security Centre", "url": "https://www.ncsc.gov.uk/blogs/managing-the-cyber-risk-of-agentic-ai", "published": "2026-08-20", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read autonomy, model safeguards, oversight, sandbox boundaries, network and credential restrictions, observability and emergency response. The publisher labels this interim practical advice based on its research; formal guidance may supersede it." }, { "id": "owasp-excessive-agency", "title": "LLM06:2025 Excessive Agency", "publisher": "OWASP Gen AI Security Project", "url": "https://genai.owasp.org/llmrisk/llm062025-excessive-agency/", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read agency definition, excessive functionality/permissions/autonomy, external authorization, approvals and monitoring limits. The 2025 label identifies the edition; the page does not establish its original publication date." }, { "id": "owasp-prompt-injection", "title": "LLM01:2025 Prompt Injection", "publisher": "OWASP Gen AI Security Project", "url": "https://genai.owasp.org/llmrisk/llm01-prompt-injection/", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read direct and indirect injection definitions, contextual impact and mitigations including validation, privilege limits and adversarial testing. The 2025 label is an edition, not a verified publication date. Does not establish that a mitigation eliminates every attack." }, { "id": "nist-genai-profile", "title": "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile", "publisher": "National Institute of Standards and Technology", "url": "https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf", "published": "2024-07", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read introduction, section 2.2 on confabulation, and selected MEASURE actions 2.3, 2.5, 2.6, 2.7 and 2.9 concerning evaluation evidence, generalization, citations, generated-code review and safeguards. A voluntary risk-management profile; no claim that all 64 pages or every referenced study was reviewed." }, { "id": "cot-monitorability", "title": "Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety", "publisher": "Tomek Korbak and coauthors / arXiv", "url": "https://arxiv.org/html/2507.11473v2", "published": "2025-07-15", "updated": "2025-12-07", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read abstract, rationale, research questions, limitations and conclusion. A research position paper: reasoning traces may add monitoring value while remaining incomplete and potentially fragile. Authors' views are not necessarily their institutions' positions; cited experiments were not all independently reviewed." }, { "id": "coding-agent-monitoring", "title": "How we monitor internal coding agents for misalignment", "publisher": "OpenAI", "url": "https://openai.com/index/how-we-monitor-internal-coding-agents-misalignment/", "published": "2026-03-19", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read deployment approach, reasoning/tool-trace monitoring, asynchronous alerts, limitations and proposed control evaluations. A dated first-party account, not an independent audit or a claim about current coverage. The authors explicitly cannot establish a real-world missed-event rate from employee escalations alone." }, { "id": "otel-observability", "title": "Observability primer", "publisher": "OpenTelemetry documentation", "url": "https://opentelemetry.io/docs/concepts/observability-primer/", "published": null, "updated": "2026-04-23", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read observability, telemetry, logs and distributed-trace definitions. Updated date follows the displayed documentation modification, which references a spelling-related commit rather than a new research result. Used for software terminology, not complete access to model internals." }, { "id": "google-ml-glossary", "title": "Machine Learning Glossary", "publisher": "Google for Developers", "url": "https://developers.google.com/machine-learning/glossary", "published": null, "updated": "2026-04-10", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the artificial intelligence, deep model, context window, inference and chat entries. Used for terminology, not product performance claims. Updated date follows the earlier displayed page date; initial publication is unspecified. The chat entry was reread during the same-day beginner-content review. Also read the generalization and compute entries." }, { "id": "google-ml-intro", "title": "What is Machine Learning?", "publisher": "Google for Developers", "url": "https://developers.google.com/machine-learning/intro-to-ml/what-is-ml", "published": null, "updated": "2026-01-27", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the introduction, model definition, supervised and unsupervised learning, and generative AI sections. Examples illustrate categories rather than measured accuracy. Updated date follows the page; initial publication is unspecified." }, { "id": "google-neural-layers", "title": "Neural networks: Nodes and hidden layers", "publisher": "Google for Developers", "url": "https://developers.google.com/machine-learning/crash-course/neural-networks/nodes-hidden-layers", "published": null, "updated": "2025-12-03", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the explanations of connected layers, numerical weights and biases, and calculations. Did not run the embedded exercises. Used for the mathematical structure, not a claim that an artificial network reproduces a human brain." }, { "id": "google-gradient-descent", "title": "Linear regression: Gradient descent", "publisher": "Google for Developers", "url": "https://developers.google.com/machine-learning/crash-course/linear-regression/gradient-descent", "published": null, "updated": "2026-02-03", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the iterative prediction, loss and parameter-update explanation. The page's guarantees for convex linear regression are not extended here to neural-network training. Updated date follows the page; initial publication is unspecified." }, { "id": "google-llm-intro", "title": "LLMs: What's a large language model?", "publisher": "Google for Developers", "url": "https://developers.google.com/machine-learning/crash-course/llm/transformers", "published": null, "updated": "2026-01-02", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read token prediction, encoder-only and decoder-only variants, and self-attention. Used for architecture and terminology; broad performance comparisons and claims about all LLMs on the teaching page are not adopted." }, { "id": "google-llm-tuning", "title": "LLMs: Fine-tuning, distillation, and prompt engineering", "publisher": "Google for Developers", "url": "https://developers.google.com/machine-learning/crash-course/llm/tuning", "published": null, "updated": "2025-12-03", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read fine-tuning and prompt engineering, including the distinction between parameter updates and examples supplied as input. Used to distinguish these processes, without adopting general claims that fine-tuning is always necessary or improves every task." }, { "id": "google-embeddings", "title": "Embeddings: Embedding space and static embeddings", "publisher": "Google for Developers", "url": "https://developers.google.com/machine-learning/crash-course/embeddings/embedding-space", "published": null, "updated": "2025-08-25", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read numerical representations, distance as relative similarity, task dependence and the limits of human-readable dimensions. The food diagrams are teaching examples, not measurements reused in this atlas." }, { "id": "hf-tokenizers", "title": "Tokenizers", "publisher": "Hugging Face LLM Course", "url": "https://huggingface.co/learn/llm-course/en/chapter2/4", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read word, character and subword tokenization, encoding to numerical IDs and decoding. Used for the fact that token boundaries depend on the tokenizer; no fixed words-to-tokens conversion is assumed. Live page publication date unspecified." }, { "id": "hf-models", "title": "Models", "publisher": "Hugging Face LLM Course", "url": "https://huggingface.co/learn/llm-course/en/chapter2/3", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read architecture, weights, checkpoints, loading and saving. Used to distinguish a model's structure and learned values from the application around it. Example code was read, not executed; live page publication date unspecified." }, { "id": "hf-text-generation", "title": "Text generation", "publisher": "Hugging Face Transformers documentation", "url": "https://huggingface.co/docs/transformers/en/llm_tutorial", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read next-token generation, generation settings, temperature, sampling and prompt-format sections. Library options illustrate the process; defaults and suggested temperatures are not treated as universal chatbot behavior. Live page publication date unspecified." }, { "id": "hf-tool-use", "title": "Tool use", "publisher": "Hugging Face Transformers documentation", "url": "https://huggingface.co/docs/transformers/en/chat_extras", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read tool descriptions, model-generated call requests, application execution and returning results to the chat. Example functions were not run. Used for the separation between requesting and executing an action; live page publication date unspecified." }, { "id": "hf-multimodal", "title": "Multimodal chat templates", "publisher": "Hugging Face Transformers documentation", "url": "https://huggingface.co/docs/transformers/en/chat_templating_multimodal", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read mixed text/image/audio/video inputs, preprocessing and model-specific video support. Used to explain input types; examples are not evidence of universal capabilities or accuracy. Live page publication date unspecified." }, { "id": "rag-paper", "title": "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks", "publisher": "Patrick Lewis and coauthors / arXiv", "url": "https://arxiv.org/abs/2005.11401", "published": "2020-05-22", "updated": "2021-04-12", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read abstract and version history, plus version 4's results sections 4.3/4.4 and Broader Impact discussion during the history review. External passages can contain errors or bias. Used for the original approach; benchmark results do not establish accuracy for every system now called RAG." }, { "id": "lost-in-middle", "title": "Lost in the Middle: How Language Models Use Long Contexts", "publisher": "Nelson F. Liu and coauthors / arXiv", "url": "https://arxiv.org/abs/2307.03172", "published": "2023-07-06", "updated": "2023-11-20", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract and version history: position-dependent performance in multi-document question answering and key-value retrieval. Used to distinguish accepted context length from effective use of content, not to assign the same weakness to every current model." }, { "id": "helm-paper", "title": "Holistic Evaluation of Language Models", "publisher": "Percy Liang and coauthors / Stanford CRFM, arXiv", "url": "https://arxiv.org/abs/2211.09110", "published": "2022-11-16", "updated": "2023-10-01", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract and version history, including multiple use cases and metrics, standardized comparisons and acknowledged coverage gaps. Used for evaluation principles; historical model scores are not presented as current rankings." }, { "id": "instructgpt-paper", "title": "Training language models to follow instructions with human feedback", "publisher": "Long Ouyang and coauthors / arXiv", "url": "https://arxiv.org/html/2203.02155v1", "published": "2022-03-04", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read abstract, section 3.1's demonstrations/comparisons/reward-model procedure and section 5.3's limitations. Human preference judgments and improved results on the authors' tasks do not establish universal alignment or safety. Publication date checked on the arXiv abstract page." }, { "id": "constitutional-ai-paper", "title": "Constitutional AI: Harmlessness from AI Feedback", "publisher": "Yuntao Bai and coauthors / arXiv", "url": "https://arxiv.org/abs/2212.08073", "published": "2022-12-15", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read abstract and metadata describing model critiques, revisions and AI preference feedback guided by human-written principles. Used as an example of alignment methods; the paper's claims do not certify every output as harmless." }, { "id": "deepseek-r1-paper", "title": "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning", "publisher": "DeepSeek-AI and coauthors / arXiv", "url": "https://arxiv.org/html/2501.12948v2", "published": "2025-01-22", "updated": "2026-01-04", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read abstract, introduction and chain-of-thought/inference-time-compute discussion. Dates checked against the abstract page's version history. Used for training and extended reasoning examples; the authors' benchmark results are not a universal ranking or an explanation of all model internals." }, { "id": "osi-ai-definition", "title": "The Open Source AI Definition – 1.0", "publisher": "Open Source Initiative", "url": "https://opensource.org/ai/open-source-ai-definition", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Browser-tool fetch returned 403; fetched and read the original publisher HTML directly. Read the four freedoms and requirements for data information, code and parameters. This is OSI's definition, not a claim that every publisher uses the label identically. Live page publication date unspecified." }, { "id": "oecd-ai-principles", "title": "AI principles", "publisher": "Organisation for Economic Co-operation and Development", "url": "https://www.oecd.org/en/topics/ai-principles.html", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read overview, human rights, transparency, safety, accountability and governance-policy recommendations. The principles were adopted in 2019 and updated in 2024; those dates are not assigned as the live page's publication date. Recommendations do not establish implementation or legal compliance." }, { "id": "nist-ai-bias", "title": "Towards a Standard for Identifying and Managing Bias in Artificial Intelligence", "publisher": "Reva Schwartz and coauthors / National Institute of Standards and Technology", "url": "https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.1270.pdf", "published": "2022-03", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read publication metadata, the systemic/statistical/human bias taxonomy, contextual evaluation discussion and conclusion. Used to explain sources and assessment of bias; this does not establish a bias finding for any particular model or actor." }, { "id": "claude-introduction", "title": "Introducing Claude", "publisher": "Anthropic", "url": "https://www.anthropic.com/news/introducing-claude", "published": "2023-03-14", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the launch introduction, chat/API distinction and model variants. Historical product identification only; customer testimonials and reliability claims are not treated as independent evidence or current specifications." }, { "id": "gemini-overview", "title": "What is Gemini and how it works", "publisher": "Google", "url": "https://gemini.google/overview/", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the introduction describing the Gemini app as an interface to Google's multimodal language models. Used for naming and app/model distinction, without endorsing performance claims. No exact publication date established." }, { "id": "grok-introduction", "title": "Announcing Grok", "publisher": "xAI (now hosted by SpaceXAI)", "url": "https://x.ai/news/grok", "published": "2023-11-03", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the original introduction, Grok-1 model description and research section acknowledging false answers despite search access. Historical identification only; no launch specifications or benchmark rankings are presented as current." }, { "id": "llama-model-family", "title": "The Llama 3 Herd of Models", "publisher": "Llama team / Meta", "url": "https://ai.meta.com/research/publications/the-llama-3-herd-of-models/", "published": "2024-07-23", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read Meta's abstract and displayed publication date, including separate pretrained and post-trained releases. Used to identify a model family, not endorse benchmark comparisons or describe the latest release." }, { "id": "nist-synthetic-content", "title": "Reducing Risks Posed by Synthetic Content: An Overview of Technical Approaches to Digital Content Transparency", "publisher": "National Institute of Standards and Technology", "url": "https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-4.pdf", "published": "2024-11-20", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the introduction, selected provenance/detection passages and Appendix D's audio/video examples. Date checked on NIST's landing page. No detector certified here." }, { "id": "merriam-webster-slop", "title": "2025 Word of the Year: Slop", "publisher": "Merriam-Webster", "url": "https://www.merriam-webster.com/wordplay/word-of-the-year", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the dictionary publisher's definition and examples in its 2025 selection. The year identifies the selection; an exact article publication date was not established. Used for language usage, not a quality finding about particular content." }, { "id": "eu-ai-act-overview", "title": "AI Act", "publisher": "European Commission", "url": "https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai", "published": null, "updated": "2026-08-03", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read risk categories, transparency, general-purpose model rules, enforcement and application timeline. Updated date is displayed on the page. This overview does not determine any particular application's legal duties; deadlines are deliberately not generalized here." }, { "id": "white-house-ai-action-plan", "title": "White House Unveils America's AI Action Plan", "publisher": "The White House", "url": "https://www.whitehouse.gov/releases/2025/07/white-house-unveils-americas-ai-action-plan/", "published": "2025-07-23", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the race framing, policy pillars and stated proposals. A government announcement and example of political framing; its promised benefits and implementation are not independently established." }, { "id": "ai-consciousness-indicators", "title": "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness", "publisher": "Patrick Butlin and coauthors / arXiv", "url": "https://arxiv.org/html/2308.08708v3", "published": "2023-08-17", "updated": "2023-08-22", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read abstract, executive summary and sections 1.1–1.2, including disputed working assumptions and limits of behavior-based assessment. Version dates checked on arXiv. A proposed framework, not a consciousness test or a 2026 assessment of all models." }, { "id": "ilo-ai-work-exposure", "title": "How might generative AI impact different occupations?", "publisher": "International Labour Organization", "url": "https://www.ilo.org/resource/article/how-might-generative-ai-impact-different-occupations", "published": "2025-05-20", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the task-based method, exposure categories, employment interpretation and adoption limitations. Estimates describe potential exposure, not observed job losses or a prediction about an individual worker." }, { "id": "chatgpt-memory", "title": "Memory FAQ", "publisher": "OpenAI Help Center", "url": "https://help.openai.com/en/articles/8590148-memory-faq", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read memory, context, controls and the legacy saved-memory explanation. Used as one provider's example; not generalized to all chatbots. Relative update wording was not converted to an invented exact date." }, { "id": "openai-training-data", "title": "How your data is used to improve model performance", "publisher": "OpenAI Help Center", "url": "https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read individual/business distinctions, training controls and feedback exceptions. Used to distinguish later training from conversation context, not to provide a complete privacy checklist. The page displays a relative update time; exact publication date is unspecified." }, { "id": "understanding-bender-koller", "title": "Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data", "publisher": "Emily M. Bender and Alexander Koller / ACL", "url": "https://aclanthology.org/2020.acl-main.463/", "published": "2020-07", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the paper and abstract. A position argument about meaning learned from form alone, not an experimental verdict on every multimodal or tool-connected system." }, { "id": "understanding-mitchell-krakauer", "title": "The Debate Over Understanding in AI’s Large Language Models", "publisher": "Melanie Mitchell and David C. Krakauer", "url": "https://arxiv.org/html/2210.13966v2", "published": "2022-10-14", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the full version 2 manuscript and arXiv submission history. First submitted October 14, 2022; the linked version 2 is dated October 27, 2022. Surveys competing accounts, benchmark shortcuts and open questions. Its model examples describe that period, not a current capability ranking.", "updated": "2022-10-27" }, { "id": "understanding-othello", "title": "Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task", "publisher": "Kenneth Li and coauthors", "url": "https://arxiv.org/abs/2210.13382", "published": "2022-10-24", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract and submission history. Reports board-state representations and interventions in a synthetic Othello task. This source alone does not establish human-like comprehension.", "updated": "2024-06-26" }, { "id": "understanding-anthropomorphism", "title": "AI Automatons: AI Systems Intended to Imitate Humans", "publisher": "Alexandra Olteanu and coauthors / Microsoft Research", "url": "https://www.microsoft.com/en-us/research/publication/ai-automatons-ai-systems-intended-to-imitate-humans/", "published": "2025-03", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the publication summary and accompanying research transcript. Discusses design choices, anthropomorphic cues and possible social effects. Does not prove a universal company motive." }, { "id": "understanding-faithfulness", "title": "Reasoning models don’t always say what they think", "publisher": "Anthropic", "url": "https://www.anthropic.com/research/reasoning-models-dont-say-think", "published": "2025-04-03", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read methods, findings and limitations. Hint experiments used Claude 3.7 Sonnet and DeepSeek R1 on multiple-choice questions. Results do not cover every model, task or reasoning trace." }, { "id": "glossary-model-algorithm", "title": "algorithm", "publisher": "Paul E. Black / NIST Dictionary of Algorithms and Data Structures", "url": "https://www.nist.gov/dads/HTML/algorithm.html", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the definition and listed algorithm types. The entry was modified on 9 November 2020; the later HTML formatting date is not a content revision.", "updated": "2020-11-09" }, { "id": "glossary-model-datasets", "title": "Datasets: Dividing the original dataset", "publisher": "Google for Developers", "url": "https://developers.google.com/machine-learning/crash-course/overfitting/dividing-datasets", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read training, validation and test separation, duplicate examples and repeated test reuse. Course examples illustrate evaluation problems; no model was tested here.", "updated": "2025-12-03" }, { "id": "glossary-model-pretraining", "title": "How do Transformers work?", "publisher": "Hugging Face LLM Course", "url": "https://huggingface.co/learn/llm-course/chapter1/4", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read self-supervised language modeling and transfer learning. Used for the pretraining distinction, without adopting historical dates, performance comparisons or claims of understanding from this teaching page." }, { "id": "glossary-model-posttraining", "title": "The Llama 3 Herd of Models", "publisher": "Llama Team / Meta, arXiv", "url": "https://arxiv.org/html/2407.21783v2", "published": "2024-07-31", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the two training stages, post-training methods and section 5.4.8 limitations in version 2. Dates follow arXiv submission history. One documented approach, not a universal training recipe.", "updated": "2024-08-15" }, { "id": "glossary-model-supervised", "title": "Supervised Learning", "publisher": "Google for Developers", "url": "https://developers.google.com/machine-learning/intro-to-ml/supervised", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read labeled examples, training, evaluation and inference. Used for the training procedure, without treating labels as infallible or adopting the page’s wording about understanding.", "updated": "2025-08-25" }, { "id": "glossary-model-clustering", "title": "What is clustering?", "publisher": "Google for Developers", "url": "https://developers.google.com/machine-learning/clustering/overview", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read grouping of unlabeled examples and the choice of similarity measure. Clustering is one example of unsupervised learning. No privacy guarantee or natural group boundary is inferred.", "updated": "2025-08-25" }, { "id": "glossary-model-self-supervised", "title": "Self-supervised learning: The dark matter of intelligence", "publisher": "Yann LeCun and Ishan Misra / Meta AI", "url": "https://ai.meta.com/blog/self-supervised-learning-the-dark-matter-of-intelligence/", "published": "2021-03-04", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the predictive-learning section and text masking examples. Used for how targets are formed. The authors’ forecasts about common sense and human-level intelligence are not adopted." }, { "id": "glossary-model-reinforcement", "title": "Deep Reinforcement Learning", "publisher": "David Silver / Google DeepMind", "url": "https://deepmind.google/blog/deep-reinforcement-learning/", "published": "2016-06-17", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the introduction defining reinforcement learning through trial, feedback and long-term rewards. Used for the training setup, without adopting human comparisons or generality claims." }, { "id": "glossary-model-deep-learning", "title": "Deep Learning", "publisher": "Ian Goodfellow, Yoshua Bengio and Aaron Courville", "url": "https://www.deeplearningbook.org/contents/intro.html", "published": "2016", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read layered representations and computational depth in Chapter 1. The book’s citation page supplies the publication year. No agreed layer threshold or intelligence measure is inferred." }, { "id": "glossary-model-mixture-of-experts", "title": "Mixtral of Experts", "publisher": "Albert Q. Jiang and coauthors / arXiv", "url": "https://arxiv.org/html/2401.04088v1", "published": "2024-01-08", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract and introduction describing token routing, expert blocks and active parameters. Publication date checked on arXiv. Mixtral illustrates sparse routing; its design is not assigned to every MoE. Also read section 5, Routing analysis, on whether routing follows subject areas." }, { "id": "glossary-model-distillation", "title": "Distilling the Knowledge in a Neural Network", "publisher": "Geoffrey Hinton, Oriol Vinyals and Jeff Dean / arXiv", "url": "https://arxiv.org/html/1503.02531v1", "published": "2015-03-09", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the introduction and distillation method, including probability targets and imperfect matching. Date checked on arXiv. The historical experiments do not establish performance for current distilled models." }, { "id": "glossary-model-quantization", "title": "Quantization concepts", "publisher": "Hugging Face Transformers documentation", "url": "https://huggingface.co/docs/transformers/quantization/concept_guide", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read lower-precision weights and activations, rounding, accuracy tradeoffs and hardware dependence. No universal speedup or accuracy loss is claimed. Live documentation publication date is unspecified." }, { "id": "glossary-model-overfitting", "title": "Overfitting", "publisher": "Google for Developers", "url": "https://developers.google.com/machine-learning/crash-course/overfitting/overfitting", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read training versus new-data performance, the underfitting comparison and generalization curves. Illustrative curves are not evidence about a particular model.", "updated": "2025-12-03" }, { "id": "glossary-model-underfitting", "title": "Machine Learning Glossary", "publisher": "Google for Developers", "url": "https://developers.google.com/machine-learning/glossary#underfitting", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the underfitting entry and its listed causes. Updated date follows the glossary page; initial publication date is unspecified. This is a diagnostic concept, not a finding about any listed model.", "updated": "2026-04-10" }, { "id": "glossary-model-loss", "title": "Linear regression: Loss", "publisher": "Google for Developers", "url": "https://developers.google.com/machine-learning/crash-course/linear-regression/loss", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read error measures and their different treatment of large errors. Used for the idea of a training objective, without presenting squared error as the standard language-model objective.", "updated": "2026-01-05" }, { "id": "glossary-eval-calibration", "title": "On Calibration of Modern Neural Networks", "publisher": "Chuan Guo and coauthors / PMLR", "url": "https://proceedings.mlr.press/v70/guo17a.html", "published": "2017-08", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the proceedings abstract and publication metadata. The experiments concern image and document classifiers, not the reliability of a chatbot saying it is certain." }, { "id": "glossary-eval-gpt3", "title": "Language Models are Few-Shot Learners", "publisher": "Tom B. Brown and coauthors / arXiv", "url": "https://arxiv.org/html/2005.14165v4", "published": "2020-05-28", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract, section 2 on zero-shot and few-shot settings, and section 4 on training-data overlap. Version 4 is dated 22 July 2020. The reported results concern GPT-3 and are not current model rankings; detected overlap did not uniformly inflate scores.", "updated": "2020-07-22" }, { "id": "glossary-eval-uncertainty", "title": "What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?", "publisher": "Alex Kendall and Yarin Gal / arXiv", "url": "https://arxiv.org/abs/1703.04977", "published": "2017-03-15", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract and version history. Used for the distinction between uncertainty in observations and uncertainty in the model. Its experiments concern computer vision, not a validated uncertainty measure for every LLM.", "updated": "2017-10-05" }, { "id": "glossary-eval-rmf-characteristics", "title": "AI Risks and Trustworthiness", "publisher": "NIST AI Resource Center", "url": "https://airc.nist.gov/airmf-resources/airmf/3-sec-characteristics/", "published": "2023", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read sections 3.1 to 3.3 of the online AI RMF 1.0 excerpt: validity, reliability, robustness, safety and security. Guidance and definitions are not certification of a particular system." }, { "id": "glossary-eval-adversarial", "title": "Explaining and Harnessing Adversarial Examples", "publisher": "Ian J. Goodfellow, Jonathon Shlens and Christian Szegedy / arXiv", "url": "https://arxiv.org/abs/1412.6572", "published": "2014-12-20", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract and version history. Supports deliberately modified inputs that cause misclassification and adversarial training as a proposed response. Findings are scoped to the studied models and attacks.", "updated": "2015-03-20" }, { "id": "glossary-eval-aml-taxonomy", "title": "Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations", "publisher": "NIST", "url": "https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-2e2025.pdf", "published": "2025-03-24", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the executive summary, attack-stage definitions, and sections 3.2.1 to 3.2.3 on generative-model poisoning and mitigations. Publication date comes from the NIST publication record. No claim is made that one defense stops all attacks." }, { "id": "glossary-eval-red-teaming", "title": "Red Teaming Language Models with Language Models", "publisher": "Ethan Perez and coauthors / arXiv", "url": "https://arxiv.org/abs/2202.03286", "published": "2022-02-07", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract and submission record, plus the authors' DeepMind research summary. The work studies automated test generation and presents it as one method among several, not an exhaustive safety test." }, { "id": "glossary-eval-induction", "title": "In-context Learning and Induction Heads", "publisher": "Catherine Olsson and coauthors / arXiv", "url": "https://arxiv.org/abs/2209.11895", "published": "2022-09-24", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract and submission record. The authors report causal evidence in small attention-only models and indirect or correlational evidence for their wider mechanism hypothesis. The glossary does not present that hypothesis as settled for all models." }, { "id": "glossary-eval-cot", "title": "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models", "publisher": "Jason Wei and coauthors / arXiv", "url": "https://arxiv.org/abs/2201.11903", "published": "2022-01-28", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract and version history. The paper reports improvements on selected arithmetic, commonsense and symbolic tasks with intermediate-step examples. It does not establish that generated explanations faithfully reveal internal computation.", "updated": "2023-01-10" }, { "id": "glossary-eval-guardrails", "title": "Guardrail Types", "publisher": "NVIDIA NeMo Guardrails documentation", "url": "https://docs.nvidia.com/nemo/guardrails/about-nemo-guardrails-library/rail-types", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the descriptions and tables of input, retrieval, dialog, execution and output rails. Used as a concrete implementation example of the broader term. No publication date is displayed; the security FAQ link returned 404 and was not used." }, { "id": "glossary-eval-human-interaction", "title": "App. C: AI Risk Management and Human-AI Interaction", "publisher": "NIST AI Resource Center", "url": "https://airc.nist.gov/airmf-resources/airmf/appendices/app-c-ai-risk-management-and-human-ai-interaction/", "published": "2023", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the AI RMF 1.0 appendix on human roles, oversight, bias and differing outcomes of human-AI interaction. It describes both possible complementarity and amplified bias; it is guidance rather than a controlled experiment." }, { "id": "glossary-eval-human-errors", "title": "The impact of AI errors in a human-in-the-loop process", "publisher": "Ujué Agudo and coauthors / Cognitive Research: Principles and Implications", "url": "https://link.springer.com/article/10.1186/s41235-023-00529-3", "published": "2024-01-07", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract, study procedures, results and general discussion. Two simulated judicial-decision experiments used purported AI advice; these are not a field trial of an LLM or a universal estimate of human oversight effectiveness." }, { "id": "glossary-eval-automation-bias", "title": "Automation Bias in AI-Assisted Medical Decision-Making under Time Pressure in Computational Pathology", "publisher": "Emely Rosbach and coauthors / arXiv", "url": "https://arxiv.org/abs/2411.00998", "published": "2024-11-01", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract and submission record. The study involved 28 pathology experts and reported improved overall performance alongside acceptance of some wrong advice. Abstract-only review; its error rate is not generalized to other users or tasks." }, { "id": "glossary-eval-contamination", "title": "Investigating Data Contamination in Modern Benchmarks for Large Language Models", "publisher": "Chunyuan Deng and coauthors / arXiv", "url": "https://arxiv.org/abs/2311.09783", "published": "2023-11-16", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract and version history. The paper studies retrieval-based overlap checks and a test-slot guessing method. Its reported scores concern particular models and benchmarks; the glossary does not treat a guessed answer alone as proof of training membership.", "updated": "2024-04-03" }, { "id": "glossary-eval-unfaithful-cot", "title": "Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting", "publisher": "Miles Turpin and coauthors / arXiv", "url": "https://arxiv.org/abs/2305.04388", "published": "2023-05-07", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract and version history. The experiments introduced biasing input features and examined generated explanations in GPT-3.5 and Claude 1.0. Their results are not a verdict on every model or every explanation.", "updated": "2023-12-09" }, { "id": "glossary-wide-vision", "title": "Cloud Vision API documentation", "publisher": "Google Cloud", "url": "https://docs.cloud.google.com/vision/docs", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the overview of image labeling, text extraction and object detection. Used to explain task types, not to claim that one product covers all computer vision or is equally accurate on every image." }, { "id": "glossary-wide-asr", "title": "Automatic speech recognition with a pipeline", "publisher": "Hugging Face", "url": "https://huggingface.co/learn/audio-course/en/chapter2/asr_pipeline", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the task definition and English/German transcription examples, including incorrect words. The examples illustrate the need to check a transcript; they do not establish error rates for current systems." }, { "id": "glossary-wide-tts", "title": "Audio generation with a pipeline", "publisher": "Hugging Face", "url": "https://huggingface.co/learn/audio-course/en/chapter2/tts_pipeline", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the speech-generation definition and examples. Used for the distinction between producing spoken audio, transcribing speech and generating music, without adopting product-quality claims." }, { "id": "glossary-wide-diffusion", "title": "Denoising Diffusion Probabilistic Models", "publisher": "Jonathan Ho, Ajay Jain and Pieter Abbeel / arXiv", "url": "https://arxiv.org/html/2006.11239v2", "published": "2020-06-19", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract, introduction, forward/reverse process and sampling algorithm. The entry explains the denoising approach in this paper, without treating its benchmark results as current performance.", "updated": "2020-12-16" }, { "id": "glossary-wide-latent-diffusion", "title": "High-Resolution Image Synthesis with Latent Diffusion Models", "publisher": "Robin Rombach and colleagues / arXiv", "url": "https://arxiv.org/html/2112.10752v2", "published": "2021-12-20", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract and introduction, including the separation between image compression and diffusion in a learned representation. Used as a concrete example of latent space, not a definition of every representation in AI.", "updated": "2022-04-13" }, { "id": "glossary-wide-synthetic-data", "title": "Welcome to the SDV!", "publisher": "DataCebo / Synthetic Data Vault", "url": "https://docs.sdv.dev/sdv", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the library overview and its generation and evaluation workflow for tabular synthetic data. This is developer documentation; no commercial claim of quality or privacy is treated as independently verified." }, { "id": "glossary-wide-synthetic-privacy", "title": "The Inadequacy of Similarity-based Privacy Metrics: Privacy Attacks against \"Truly Anonymous\" Synthetic Datasets", "publisher": "Georgi Ganev and Emiliano De Cristofaro / arXiv", "url": "https://arxiv.org/abs/2312.05114", "published": "2023-12-08", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract and version history. The reported attacks show that passing the studied similarity-based tests does not establish anonymity. The entry does not claim that every synthetic dataset leaks personal information.", "updated": "2025-05-07" }, { "id": "glossary-wide-knowledge-cutoff", "title": "Dated Data: Tracing Knowledge Cutoffs in Large Language Models", "publisher": "Jeffrey Cheng and colleagues / arXiv", "url": "https://arxiv.org/abs/2403.12958", "published": "2024-03-19", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract and version history. The paper distinguishes a reported cutoff from the effective cutoff for particular resources and topics. No specific current product cutoff is inferred.", "updated": "2024-09-17" }, { "id": "glossary-wide-provenance", "title": "PROV-Overview", "publisher": "W3C", "url": "https://www.w3.org/TR/prov-overview/", "published": "2013-04-30", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract and introduction to the PROV family, including entities, people, activities and derivation. Provenance supports assessment of reliability; the document does not make a recorded origin proof of truth." }, { "id": "glossary-wide-model-card", "title": "Model Cards for Model Reporting", "publisher": "Margaret Mitchell and colleagues / arXiv", "url": "https://arxiv.org/abs/1810.03993", "published": "2018-10-05", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract and version history, including intended uses, evaluation conditions and reporting across groups. This is a documentation proposal, not a certification that a model is safe or fair.", "updated": "2019-01-14", "retrieval": { "method": "publisher-page", "scope": "The abstract and version history, not the full paper." } }, { "id": "glossary-wide-system-card", "title": "System Cards, a new resource for understanding how AI systems work", "publisher": "Meta AI", "url": "https://ai.meta.com/blog/system-cards-a-new-resource-for-understanding-how-ai-systems-work/", "published": "2022-02-23", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the system-card proposal, its distinction from model cards and its limitations section. Meta describes its own approach; publication of a card is not an independent audit of the system." }, { "id": "glossary-wide-gpu", "title": "1.1. Introduction", "publisher": "NVIDIA / CUDA Programming Guide", "url": "https://docs.nvidia.com/cuda/cuda-programming-guide/01-introduction/introduction.html", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read sections 1.1.1 and 1.1.2 on GPU origins and parallel computation. The entry uses the architectural distinction without repeating vendor performance or energy comparisons.", "updated": "2026-09-09" }, { "id": "glossary-wide-latency", "title": "MLPerf Inference 5.1: Benchmarking Small LLMs with Llama3.1-8B", "publisher": "MLCommons", "url": "https://mlcommons.org/2025/09/small-llm-inference-5-1/", "published": "2025-09", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the performance-metrics section distinguishing time to first token, subsequent token timing and throughput. Specific benchmark targets are not presented as universal user requirements." }, { "id": "glossary-wide-on-device", "title": "LiteRT: High-Performance On-Device Machine Learning Framework", "publisher": "Google for Developers", "url": "https://developers.google.com/edge/litert", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the platform overview and deployment workflow for running models on devices. The entry describes the location of computation, without adopting blanket vendor claims about privacy, speed or capability.", "updated": "2026-07-17" }, { "id": "glossary-balance-ea-crary", "title": "Against ‘Effective Altruism’", "publisher": "Alice Crary / Radical Philosophy", "url": "https://www.radicalphilosophy.com/article/against-effective-altruism", "published": "2021", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the institutional and philosophical critique and discussion of EA replies. These are Crary’s arguments, not findings about every participant or charity. Publication precision is the issue year." }, { "id": "glossary-balance-ea-donor-power", "title": "Response to Effective Altruism", "publisher": "Emma Saunders-Hastings / Boston Review", "url": "https://www.bostonreview.net/forum/peter-singer-logic-effective-altruism/response-emma-saunders-hastings/", "published": "2015-07-01", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the complete response on donor power and recipient choice. Historical charity recommendations are not repeated as current advice. The argument warns of a governance problem rather than rejecting every donation." }, { "id": "glossary-balance-ea-response", "title": "Frequently asked questions and common objections", "publisher": "EffectiveAltruism.org", "url": "https://www.effectivealtruism.org/faqs", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the definitions and replies on utilitarianism, uncertainty, systemic change and competing priorities. This is the movement website’s account of its principles, not an independent evaluation of its institutions." }, { "id": "glossary-balance-longtermism-setiya", "title": "The New Moral Mathematics", "publisher": "Kieran Setiya / Boston Review", "url": "https://www.bostonreview.net/articles/the-new-moral-mathematics/", "published": "2022-08-15", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read Setiya’s review sections on population ethics and the priority given to future survival over present suffering. The entry attributes his ethical criticism and does not adopt the article’s empirical forecasts." }, { "id": "glossary-balance-risk-methodology", "title": "Democratising Risk: In Search of a Methodology to Study Existential Risk", "publisher": "Carla Zoe Cremer and Luke Kemp / arXiv", "url": "https://arxiv.org/abs/2201.11214", "published": "2021-12-27", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract and submission history. The authors argue for clearer definitions, plural values and better risk methods. Their criticism of one influential framework is not a finding that all catastrophe research is invalid." }, { "id": "glossary-balance-utilitarianism-objections", "title": "Objections to Utilitarianism and Responses", "publisher": "Richard Yetter Chappell, Darius Meissner and William MacAskill / Utilitarianism.net", "url": "https://utilitarianism.net/objections-to-utilitarianism/", "published": "2023", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the rights and demandingness objections and the authors’ response strategies. This textbook defends utilitarianism; its replies are arguments, not a resolution accepted by all moral philosophers." }, { "id": "glossary-balance-pause-failures", "title": "Pausing AI Development Might Go Wrong. How to Mitigate the Risks?", "publisher": "PauseAI", "url": "https://pauseai.info/mitigating-pause-failures", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the sections on delayed benefits, uncertain duration, weak enforcement, underground development and political compromise. These are PauseAI’s own possible failure scenarios, not measured outcomes. The page includes older threshold examples; current policy scope follows the April 2026 proposal. Publication date unspecified." }, { "id": "software-lens-nist-computer", "title": "Computer", "publisher": "National Institute of Standards and Technology", "url": "https://csrc.nist.gov/glossary/term/computer", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the glossary definition identifying a computer as a device that processes digital data according to program instructions. The page attributes the definition to NIST SP 800-34 Rev. 1; that full contingency-planning document was not reviewed. Live glossary publication date unspecified." }, { "id": "software-lens-nist-software", "title": "software", "publisher": "National Institute of Standards and Technology", "url": "https://csrc.nist.gov/glossary/term/software", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the definitions of software as programs and associated data, and the distinction from physical hardware. The glossary aggregates several contextual definitions, including entries for related hardware and firmware terms; it is not a single universal definition. Referenced standards were not read in full. Live page publication date unspecified." }, { "id": "software-lens-oecd-system", "title": "Explanatory memorandum on the updated OECD definition of an AI system", "publisher": "Organisation for Economic Co-operation and Development", "url": "https://www.oecd.org/content/dam/oecd/en/publications/reports/2024/03/explanatory-memorandum-on-the-updated-oecd-definition-of-an-ai-system_3c815e51/623da898-en.pdf", "published": "2024-03-05", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the updated definition and explanatory sections on techniques, human roles, autonomy, physical and virtual environments, inputs, models and outputs (pages 4 and 6 to 9). Publication date follows the OECD publication landing page. Used to distinguish models from systems and learned methods from knowledge-based approaches, not as a claim about human-like understanding or a legal classification of a particular product." }, { "id": "gebru-wired-2026", "title": "One of AI’s Fiercest Critics Says All the Doom Talk Is ‘Meant to Distract Us’", "publisher": "WIRED", "url": "https://www.wired.com/story/one-of-ais-fiercest-critics-says-all-the-doom-talk-is-meant-to-distract-us/", "published": "2026-09-11", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the published interview. Used her stated risk distinction, not claims about corporate motives." }, { "id": "sutskever-dwarkesh-2025", "title": "Ilya Sutskever — We're moving from the age of scaling to the age of research", "publisher": "Dwarkesh Podcast / Dwarkesh Patel", "url": "https://www.dwarkesh.com/p/ilya-sutskever-2", "published": "2025-11-25", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the host transcript, especially 00:43:04 to 00:46:47 and 00:56:10 to 01:04:10. The host describes editorially reworked transcripts. The proposed power cap lacks an implementation method; gradual release concerns deployment." }, { "id": "sutskever-superalignment-2023", "title": "Introducing Superalignment", "publisher": "OpenAI / Jan Leike and Ilya Sutskever", "url": "https://openai.com/index/introducing-superalignment/", "published": "2023-07-05", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the coauthored introduction, research approach and footnotes. Used for historically expressed extinction concern and the distinction between a research aim and a solved control problem. The announced timetable is not treated as a result." }, { "id": "sutskever-nvidia-2026", "title": "Ilya Sutskever’s Safe Superintelligence Inc. and NVIDIA Announce Long-Term Strategic Partnership", "publisher": "NVIDIA / GlobeNewswire", "url": "https://www.globenewswire.com/news-release/2026/07/27/3333561/0/en/ilya-sutskever-s-safe-superintelligence-inc-and-nvidia-announce-long-term-strategic-partnership.html", "published": "2026-07-27", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the issuer announcement linked from SSI, including Sutskever’s attributed research-scaling statement. Primary company announcement, not independent reporting or a capability test. SSI’s update is dated July 26; this release is dated July 27." }, { "id": "zuckerberg-future-2026", "title": "The Future is for Everyone", "publisher": "Meta / Mark Zuckerberg", "url": "https://www.meta.com/thefutureisforeveryone/", "published": "2026-08-10", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the signed essay, especially the misuse, American leadership, existential-risk and control sections. Predictions and balance-of-power arguments are attributed to Zuckerberg; their effectiveness and announced governance are not independently verified." }, { "id": "zuckerberg-personal-2025", "title": "Personal Superintelligence", "publisher": "Meta / Mark Zuckerberg", "url": "https://www.meta.com/superintelligence/", "published": "2025-07-30", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the complete signed letter. It combines a building ambition with caution about novel safety risks and what to open source. Openness and release timing alone do not establish frontier-development pace." }, { "id": "zuckerberg-sources-2026", "title": "Mark Zuckerberg on Muse, Meta's biggest AI bet yet", "publisher": "Sources / Alex Heath", "url": "https://sources.news/p/mark-zuckerberg-meta-muse-ai-podcast-interview", "published": "2026-09-08", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the original interviewer’s introduction and episode highlights only. The full audio and transcript were not reviewed. Used only to record the recency search and retrieval limit, not as axis evidence or independent verification of product claims." }, { "id": "suleyman-humanist-2025", "title": "Towards Humanist Superintelligence", "publisher": "Mustafa Suleyman", "url": "https://mustafa-suleyman.ai/a-humanist-future", "published": "2025-11-07", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the personal-site essay, including Containment is necessary and A safer superintelligence. This longer version is dated November 7; Microsoft’s shorter article is dated November 6. Claims about medical performance and future benefits are not adopted." }, { "id": "suleyman-code-2026", "title": "The Humanist AI Code of Conduct", "publisher": "Mustafa Suleyman", "url": "https://mustafa-suleyman.ai/the-humanist-ai-code-of-conduct", "published": "2026-09-15", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the complete personal essay from the original HTML with Python urllib after web retrieval failed. The page displays 15 September 2026. Used for stated control limits and future implementation; incident descriptions and consciousness claims are not independently established." }, { "id": "normal-technology-2025", "title": "AI as Normal Technology", "publisher": "Knight First Amendment Institute / Arvind Narayanan and Sayash Kapoor", "url": "https://knightcolumbia.org/content/ai-as-normal-technology", "published": "2025-04-15", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the introduction and relevant passages in Parts I to IV on diffusion, capability and power, catastrophic misalignment, resilience and nonproliferation. The displayed publication date is April 15; the suggested citation instead says April 14. This record follows the displayed date and preserves the discrepancy here. Treat the essay as the authors' argument. Their September 2026 update qualifies its safety claims; cited incident reports and studies were not independently audited." }, { "id": "normal-technology-guide-2025", "title": "A guide to understanding AI as normal technology", "publisher": "AI as Normal Technology / Arvind Narayanan and Sayash Kapoor", "url": "https://www.normaltech.ai/p/a-guide-to-understanding-ai-as-normal", "published": "2025-09-09", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the clarification of normal, restatement of the thesis, and response to Scott Alexander, including the distinction between economic and safety arguments. This is the authors' own explanation and reply, not independent verification of its forecasts. Linked conversations and all underlying empirical citations were not reviewed." }, { "id": "normal-technology-alexander-response-2025", "title": "AI As Profoundly Abnormal Technology", "publisher": "AI Futures Project / Scott Alexander", "url": "https://blog.aifutures.org/p/ai-as-profoundly-abnormal-technology", "published": "2025-07-24", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the opening, adoption by key actors, and sections on control, speculative risk and institutional assumptions. Used as a direct critique of the thesis, not as verification of the critic's forecasts or cited anecdotes. The older ai-futures.org link redirects to aifutures.org. The September 2025 reply and September 2026 revision are provided alongside this critique." }, { "id": "normal-technology-software-work-2026", "title": "Why AI hasn’t replaced software engineers, and won’t", "publisher": "AI as Normal Technology / Arvind Narayanan and Sayash Kapoor", "url": "https://www.normaltech.ai/p/why-ai-hasnt-replaced-software-engineers", "published": "2026-06-11", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the opening and discussion separating decisions, execution and delivery, including its explicit preference for accountability over slowing technical capabilities. Used for the dated public pace argument. Employment statistics, layoff reporting and underlying studies were not independently checked. The September 2026 essay supplies the more recent qualification about pausing experiments." }, { "id": "normal-technology-intervention-2026", "title": "Do AI Risks Require Extraordinary Government Intervention?", "publisher": "AI as Normal Technology / Sayash Kapoor and Arvind Narayanan", "url": "https://www.normaltech.ai/p/do-ai-risks-require-extraordinary", "published": "2026-05-21", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the framing of extraordinary intervention, its distinction between economic adoption and misuse, and the resilience section. The authors allow that precaution and temporary access restrictions can help while arguing for less restrictive defenses. Used as context for their own policy argument, not legal advice or verification of current law, capability gaps or historical examples." }, { "id": "normal-technology-research-2026", "title": "AI agents can't yet do open-ended AI research", "publisher": "AI as Normal Technology / Sayash Kapoor and Arvind Narayanan", "url": "https://www.normaltech.ai/p/ai-agents-cant-yet-do-open-ended", "published": "2026-08-05", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the authors' research summary, its two-case design, stated limitations, and discussion of possible remaining bottlenecks. The linked paper, artifacts and agent logs were not audited. The newsletter reports the authors' own research but does not establish that current limitations are permanent or that all kinds of AI research are equally difficult." }, { "id": "normal-technology-control-2026", "title": "The AI-as-Normal-Technology view of loss-of-control incidents", "publisher": "AI as Normal Technology / Sayash Kapoor and Arvind Narayanan", "url": "https://www.normaltech.ai/p/the-ai-as-normal-technology-view", "published": "2026-09-14", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the opening synthesis, Part 1 on control, governance and policy, and Part 3's explicit revisions and risk assessment, plus the policy discussion before Part 3. Used for the authors' advocacy and self-described changes. Their accounts of third-party incidents, product behavior, investment and law were not independently verified. The proposed record does not repeat those accounts as established facts or infer a whole-lab halt." }, { "id": "glossary-symbolic-stanford", "title": "What is Traditional AI?", "publisher": "Stanford HAI", "url": "https://hai.stanford.edu/ai-definitions/what-is-traditional-ai", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read Stanford's institutional definition of explicitly programmed rules and symbolic reasoning. No publication date shown; mental-state wording is not adopted." }, { "id": "glossary-expert-stanford", "title": "What is an Expert System?", "publisher": "Stanford HAI", "url": "https://hai.stanford.edu/ai-definitions/what-is-an-expert-system", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read knowledge-base, if-then-rule and limited-domain definitions. No publication date shown. Historical priority and broad interpretability claims are not adopted." }, { "id": "foundation-model-report", "title": "On the Opportunities and Risks of Foundation Models", "publisher": "Rishi Bommasani and coauthors / Stanford CRFM / arXiv", "url": "https://arxiv.org/abs/2108.07258", "published": "2021-08-16", "updated": "2022-07-12", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read abstract, introduction and shared-model risk discussion. Dates follow arXiv versions. A terminology proposal and research synthesis." }, { "id": "scaling-kaplan", "title": "Scaling Laws for Neural Language Models", "publisher": "Jared Kaplan and coauthors / arXiv", "url": "https://arxiv.org/abs/2001.08361", "published": "2020-01-23", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read abstract and submission record. Empirical relationships concern language-model prediction loss and training resources, not an AGI date." }, { "id": "scaling-chinchilla", "title": "Training Compute-Optimal Large Language Models", "publisher": "Jordan Hoffmann and coauthors / arXiv", "url": "https://arxiv.org/abs/2203.15556", "published": "2022-03-29", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read abstract and submission record. Tests balance parameter count and training tokens under a fixed compute budget. Results are scoped to the studied setups." }, { "id": "model-collapse-nature", "title": "AI models collapse when trained on recursively generated data", "publisher": "Ilia Shumailov and coauthors / Nature", "url": "https://www.nature.com/articles/s41586-024-07566-y", "published": "2024-07-24", "updated": "2025-03-21", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read definition and recursive-training setup. Checked the 2025 correction to a mathematical symbol. Findings are not evidence that every synthetic-data method fails." }, { "id": "model-collapse-accumulation", "title": "Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data", "publisher": "Matthias Gerstgrasser and coauthors / arXiv", "url": "https://arxiv.org/abs/2404.01413", "published": "2024-04-01", "updated": "2024-04-29", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read abstract, introduction and language-model setup. Accumulating original and generated examples avoided collapse in tested settings. The original TinyStories text was itself synthetic." }, { "id": "data-work-typology", "title": "A typology of artificial intelligence data work", "publisher": "James Muldoon, Callum Cant, Boxi Wu and Mark Graham / Big Data & Society", "url": "https://journals.sagepub.com/doi/10.1177/20539517241232632", "published": "2024-03-18", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read abstract, fieldwork scope and data-work definitions. Includes computer-vision data workplaces and other fieldwork; not a representative census of all AI labor." }, { "id": "data-cascades-author-summary", "title": "Data Cascades in Machine Learning", "publisher": "Nithya Sambasivan / Google Research", "url": "https://research.google/blog/data-cascades-in-machine-learning/", "published": "2021-06-04", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the researcher's account of the interview study, examples and data-work recommendations. Findings concern the studied projects, not every AI system." }, { "id": "ai-control-original", "title": "AI Control: Improving Safety Despite Intentional Subversion", "publisher": "Ryan Greenblatt, Buck Shlegeris, Kshitij Sachan and Fabien Roger / Redwood Research / arXiv", "url": "https://arxiv.org/abs/2312.06942", "published": "2023-12-12", "updated": "2024-07-23", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read abstract, introduction and sections 5.1.2 and 5.2. Programming-task experiments, with human review simulated by a model. Control is not declared solved." }, { "id": "ai-control-monitor-attacks", "title": "Adaptive Attacks on Trusted Monitors Subvert AI Control Protocols", "publisher": "Mikhail Terekhov and coauthors / arXiv", "url": "https://arxiv.org/abs/2510.09462", "published": "2025-10-10", "updated": "2026-03-02", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read abstract and version history. Reports prompt-injection attacks against monitors on two control benchmarks. Findings are scoped to tested protocols, not every possible safeguard." }, { "id": "ai-control-threats", "title": "Prioritizing threats for AI control", "publisher": "Ryan Greenblatt / Redwood Research", "url": "https://www.redwoodresearch.org/blog/prioritizing-threats-for-ai-control", "published": "2025-03-19", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the proposed threat categories, permission limits and blocking-review discussion. Prospective threat modeling and author priorities, not observed catastrophic events." }, { "id": "card-bcs-identity", "title": "How configuration, change and release management can save the world!", "publisher": "Daniel Card / BCS", "url": "https://www.bcs.org/media/4614/ccrm-dan-card.pdf", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the presentation's title slide, which displays Daniel Card and @Uk_Daniel_Card. Used only to corroborate the name and public handle. The retrieved slides do not establish a publication date, authenticate the supplied export or establish a current position on either map axis.", "retrieval": { "method": "publisher-page", "scope": "The presentation's title slide, used for the name and public handle only." } }, { "id": "card-x-system-controls-2026", "title": "Supplied export of an X post on software and system controls (editorial description)", "publisher": "User-supplied X export; author field: @UK_Daniel_Card", "url": "https://x.com/UK_Daniel_Card/status/2099751708876308497", "published": "2026-09-15", "checkedOn": "2026-09-15", "kind": "reporting", "verification": "read", "notes": "Read the text in a user-supplied export, matching Author Username to UK_Daniel_Card and retaining the original post ID and URL. The live X URL returned 403, so the original was not independently retrieved. The date follows the export's Created At field, which has no timezone. Repeated rows label this same authored post both Tweet and Quoted; other authors' quoted or reposted text was not attributed to Card. Media and complete threads were not reviewed. The post's allegations and technical conclusions were not independently verified. The read label applies only to the supplied export text, not the linked original. The author field is an attribution in the export, not independent verification of authorship.", "retrieval": { "method": "supplied-export", "scope": "Selected text in a supplied export. Complete threads and media were not reviewed." }, "archive": { "status": "not-verified", "checkedOn": "2026-09-15", "notes": "The Wayback availability lookup returned HTTP 429 (rate limited). No capture was verified; this does not show that no archived copy exists." } }, { "id": "card-x-monitor-actions-2026", "title": "Supplied export of an X reply on monitoring outputs and actions (editorial description)", "publisher": "User-supplied X export; author field: @UK_Daniel_Card", "url": "https://x.com/UK_Daniel_Card/status/2099739563233091675", "published": "2026-09-15", "checkedOn": "2026-09-15", "kind": "reporting", "verification": "read", "notes": "Read the reply text in a user-supplied export after matching its author field to UK_Daniel_Card. The live original returned 403. The date follows the export's timestamp without assigning a timezone. Full conversation context and media were not retrieved. Used for his stated monitoring priority, not as proof that model-level monitoring is useless. The read label applies only to the supplied export text, not the linked original. The author field is an attribution in the export, not independent verification of authorship.", "retrieval": { "method": "supplied-export", "scope": "Selected text in a supplied export. Complete threads and media were not reviewed." }, "archive": { "status": "not-verified", "checkedOn": "2026-09-15", "notes": "The Wayback availability lookup returned HTTP 429 (rate limited). No capture was verified; this does not show that no archived copy exists." } }, { "id": "card-x-fit-task-2026", "title": "Supplied export of an X post on choosing tools for the task (editorial description)", "publisher": "User-supplied X export; author field: @UK_Daniel_Card", "url": "https://x.com/UK_Daniel_Card/status/2099529459372068985", "published": "2026-09-14", "checkedOn": "2026-09-15", "kind": "reporting", "verification": "read", "notes": "Read the text in a user-supplied export. Its Origin label was not treated as authorship evidence; attribution uses the UK_Daniel_Card author field. The live original returned 403. The date follows the export's timestamp, which has no timezone. Used for the argument about matching tools to tasks, not as a universal claim about model determinism or frontier development policy. The read label applies only to the supplied export text, not the linked original. The author field is an attribution in the export, not independent verification of authorship.", "retrieval": { "method": "supplied-export", "scope": "Selected text in a supplied export. Complete threads and media were not reviewed." }, "archive": { "status": "not-verified", "checkedOn": "2026-09-15", "notes": "The Wayback availability lookup returned HTTP 429 (rate limited). No capture was verified; this does not show that no archived copy exists." } }, { "id": "card-x-risk-qualification-2026", "title": "Supplied export of an X post qualifying software and risk claims (editorial description)", "publisher": "User-supplied X export; author field: @UK_Daniel_Card", "url": "https://x.com/UK_Daniel_Card/status/2099764961098473759", "published": "2026-09-15", "checkedOn": "2026-09-15", "kind": "reporting", "verification": "read", "notes": "Read the text in a user-supplied export after matching the author field to UK_Daniel_Card. The live X original returned 403. The date follows the export's timestamp, which has no timezone. His warning against inferring the opposite of a criticism is retained as a qualification; the post does not provide an overall catastrophic-risk estimate. The read label applies only to the supplied export text, not the linked original. The author field is an attribution in the export, not independent verification of authorship.", "retrieval": { "method": "supplied-export", "scope": "Selected text in a supplied export. Complete threads and media were not reviewed." }, "archive": { "status": "not-verified", "checkedOn": "2026-09-15", "notes": "The Wayback availability lookup returned HTTP 429 (rate limited). No capture was verified; this does not show that no archived copy exists." } }, { "id": "card-pwndefend-risk-framing-2026", "title": "AI: Fear it, so I can sell you the cure!", "publisher": "PwnDefend / Daniel Card", "url": "https://www.pwndefend.com/2026/07/04/ai-fear-it-so-i-can-sell-you-the-cure/", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the original page, especially its digital-nuclear-weapon analogy, risk framing and closing disclosure. The disclosure says the article was generated with Opus 4.8 from the author's prompts and judges it broadly on point. Treat it as published, AI-assisted commentary, not independently verified technical evidence. The permalink contains 2026/07/04, but no separate publication date was displayed in the retrieved text, so published remains null. A direct metadata retrieval attempt returned 403.", "retrieval": { "method": "publisher-page", "scope": "The published commentary and its AI-assistance disclosure." } }, { "id": "card-x-opportunity-2026", "title": "Supplied export of an X post endorsing AI opportunities and security collaboration (editorial description)", "publisher": "User-supplied X export; author field: @UK_Daniel_Card", "url": "https://x.com/UK_Daniel_Card/status/2099598348911018445", "published": "2026-09-14", "checkedOn": "2026-09-15", "kind": "reporting", "verification": "read", "notes": "The copied post quotes support for AI opportunities and involvement of cybersecurity practitioners, then explicitly signals agreement. The quoted wording is not presented as Card's original writing. This supports an endorsement of opportunities, not a preference on frontier capability growth. Read from the supplied export after matching Author Username to UK_Daniel_Card and deduplicating post IDs. The original X URL returned HTTP 403 on 15 September 2026. The read status covers the copied text, not the linked original, and does not independently authenticate authorship. Publication follows the export date without assigning a timezone. Full threads and media were not reviewed.", "retrieval": { "method": "supplied-export", "scope": "Selected text in a supplied export. Complete threads and media were not reviewed." } }, { "id": "card-x-system-tradeoffs-2026", "title": "Supplied export of an X reply on software design, opportunities and risks (editorial description)", "publisher": "User-supplied X export; author field: @UK_Daniel_Card", "url": "https://x.com/UK_Daniel_Card/status/2099549597513187546", "published": "2026-09-14", "checkedOn": "2026-09-15", "kind": "reporting", "verification": "read", "notes": "The copied reply discusses differences between computing architectures, benefits for prototypes, energy costs and risks of deploying technology before addressing security. Used to describe the attributed argument, not to verify its technical claims or historical examples. Read from the supplied export after matching Author Username to UK_Daniel_Card and deduplicating post IDs. The original X URL returned HTTP 403 on 15 September 2026. The read status covers the copied text, not the linked original, and does not independently authenticate authorship. Publication follows the export date without assigning a timezone. Full threads and media were not reviewed.", "retrieval": { "method": "supplied-export", "scope": "Selected text in a supplied export. Complete threads and media were not reviewed." } }, { "id": "development-gender-shades", "title": "Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification", "publisher": "Joy Buolamwini and Timnit Gebru / PMLR", "url": "https://proceedings.mlr.press/v81/buolamwini18a.html", "published": "2018-02", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the original abstract and conference citation metadata. The findings concern three commercial gender-classification systems and the study's dataset, not all facial analysis or current versions. The limitation states the scope of this historical evidence; no independent replication was performed." }, { "id": "development-turing", "title": "Computing Machinery and Intelligence", "publisher": "Alan M. Turing / Mind; text hosted by Simon Fraser University", "url": "https://www.cs.sfu.ca/~vaughan/teaching/889/papers/turing1950.html", "published": "1950", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the original paper's university-hosted transcription, including the imitation game, objections, digital computers and learning machines. Used for Turing's proposed questions, not a claim that a contemporary model passes a universally agreed intelligence test." }, { "id": "development-dartmouth-proposal", "title": "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence", "publisher": "John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon / Stanford archive", "url": "https://www-formal.stanford.edu/jmc/history/dartmouth/dartmouth.html", "published": "1955-08-31", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the dated proposal, its stated conjecture and proposed topics. The source describes plans for summer 1956; the separate Dartmouth institutional history supports that the gathering took place." }, { "id": "development-dartmouth-history", "title": "Our Story", "publisher": "Dartmouth / AI at Dartmouth", "url": "https://ai.dartmouth.edu/our-story", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the institution's retrospective on the summer 1956 gathering and John McCarthy's organizing role. No publication date is shown. Its promotional claims and present-day research announcements are not used for the historical event." }, { "id": "development-eliza", "title": "ELIZA—A Computer Program For the Study of Natural Language Communication Between Man and Machine", "publisher": "Joseph Weizenbaum / Communications of the ACM; text hosted by UMBC", "url": "https://courses.cs.umbc.edu/331/papers/eliza.html", "published": "1966-01", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the original paper's university-hosted text, particularly the abstract, keyword and reassembly rules, and the unhappy-sentence example showing transformations independent of meaning. Date retains the issue's month precision." }, { "id": "development-backpropagation", "title": "Learning representations by back-propagating errors", "publisher": "David E. Rumelhart, Geoffrey E. Hinton and Ronald J. Williams / Nature", "url": "https://www.cs.toronto.edu/~hinton/absps/naturebp.pdf", "published": "1986-10-09", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Visually read the author-hosted scan's opening page, page 533: publication date, abstract and explanation of desired outputs and hidden units. Used for the paper's method and task setup; not a claim that this paper was the first invention of backpropagation." }, { "id": "development-deep-blue", "title": "Deep Blue", "publisher": "IBM History", "url": "https://www.ibm.com/history/deep-blue", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read IBM's account of the May 1997 rematch, score, search hardware and chess-specific preparation. This is the builder's retrospective; claims about wider industrial benefits are not adopted. The page supplies a match month, not an exact day." }, { "id": "development-imagenet", "title": "ImageNet: A Large-Scale Hierarchical Image Database", "publisher": "Jia Deng and coauthors / CVPR 2009", "url": "https://www.image-net.org/static_files/papers/imagenet_cvpr09.pdf", "published": "2009", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the project-hosted original paper's abstract and sections 1–3 on size, WordNet categories, image collection, human checking and the subset used for analysis. Historical counts describe the paper's version, not the current dataset." }, { "id": "development-alexnet", "title": "ImageNet Classification with Deep Convolutional Neural Networks", "publisher": "Alex Krizhevsky, Ilya Sutskever and Geoffrey E. Hinton / NIPS 2012 proceedings", "url": "https://papers.nips.cc/paper_files/paper/2012/file/c399862d3b9d6b76c8436e924a68c45b-Paper.pdf", "published": "2012", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the proceedings PDF's abstract, introduction and competition-results section. The PDF describes the 2012 winning entry and the limits of image datasets; its numbers differ from the older abstract text on the proceedings landing page, so the PDF is authoritative here." }, { "id": "development-alphago", "title": "AlphaGo", "publisher": "Google DeepMind", "url": "https://deepmind.google/research/alphago/", "published": null, "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the builder's retrospective sections Our approach and The matches, covering expert games, self-play, search and the March 2016 Lee Sedol result. Broad promotional claims about creativity or solving other domains are not adopted." }, { "id": "development-bert", "title": "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding", "publisher": "Jacob Devlin and coauthors / arXiv", "url": "https://arxiv.org/abs/1810.04805", "published": "2018-10-11", "updated": "2019-05-24", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract and submission history: bidirectional text representations, task-specific fine-tuning and reported results on eleven language tasks. The timeline date is the original preprint, not the later revision." }, { "id": "development-gpt3", "title": "Language Models are Few-Shot Learners", "publisher": "Tom B. Brown and coauthors / OpenAI, arXiv", "url": "https://arxiv.org/html/2005.14165v4", "published": "2020-05-28", "updated": "2020-07-22", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract, sections 4–5 and appendix C on training/test overlap, and the submission history. Covers text-supplied demonstrations, weak tasks and overlap checks that find small effects on most tests but flag some results. Results are the model developers' report." }, { "id": "development-chatgpt", "title": "Introducing ChatGPT", "publisher": "OpenAI", "url": "https://openai.com/index/chatgpt/", "published": "2022-11-30", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the original announcement, Methods and Limitations. The page now explicitly labels itself the 2022 introduction. Claims concern that launch version, including dialogue training and reported errors, not every later ChatGPT model." }, { "id": "development-gpt4", "title": "GPT-4 Technical Report", "publisher": "OpenAI / arXiv", "url": "https://arxiv.org/html/2303.08774v6", "published": "2023-03-15", "updated": "2024-03-04", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the abstract, introduction and Scope and Limitations section; checked publication and revision dates against the abstract page. The report supports image/text inputs, training overview and withheld technical details. Performance statements are attributed to OpenAI." }, { "id": "development-alphafold3", "title": "Accurate structure prediction of biomolecular interactions with AlphaFold 3", "publisher": "Josh Abramson and coauthors / Nature", "url": "https://www.nature.com/articles/s41586-024-07487-w", "published": "2024-05-08", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read publication metadata, abstract, main description and Model limitations, including molecular geometry, overlapping atoms and limited conformations. Comparisons are the authors' study results; no clinical or drug-development outcome is inferred." }, { "id": "development-deepseek-r1", "title": "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning — January 2025 version", "publisher": "DeepSeek-AI / arXiv", "url": "https://arxiv.org/html/2501.12948v1", "published": "2025-01-22", "checkedOn": "2026-09-15", "kind": "primary", "verification": "read", "notes": "Read the pinned first version's abstract and section 5, including training variants, smaller models and limitations in tool calls, multi-turn tasks, language mixing and prompting. This historical record intentionally uses v1, not the revised January 2026 text. Results are the developers' report." } ], "glossary": [ { "id": "eacc", "short": "e/acc", "term": "Effective accelerationism", "category": "Movement", "guide": "accelerate", "definition": "A movement that treats technological growth, markets and the expansion of intelligence as paths to a better future. Its early writers oppose attempts to centrally slow that process.", "summary": "A movement favoring faster technological growth and opposing centralized restraint. Its claim that acceleration leads to better outcomes is a philosophical position, not a demonstrated safety guarantee.", "placement": "Points toward faster development. That does not give every participant the same belief about catastrophic risk.", "distinction": "Broader techno-optimism does not imply e/acc membership. Its thermodynamic arguments do not establish that AI is safe. Buterin offers a counterpoint: profit alone does not automatically select beneficial directions for technology.", "group": "Movements", "references": { "summary": [ "eacc-definition", "eacc-tenets" ], "definition": [ "eacc-definition", "eacc-tenets" ], "placement": [ "eacc-tenets" ], "distinction": [ "eacc-tenets", "dacc-original" ] }, "sources": [ "eacc-definition", "eacc-tenets", "dacc-original" ] }, { "id": "decel", "short": "Decel", "term": "Deceleration", "category": "Contested label", "guide": "restrain", "definition": "Informal shorthand for slowing technological or AI development, often used critically by accelerationists about their opponents.", "summary": "An informal label for slowing technology, often used by accelerationists to criticize their opponents.", "placement": "Describes a pace preference. It does not establish why someone wants restraint or which technologies they would slow.", "distinction": "A targeted frontier pause, a safety rule and opposition to all technology are different positions. Do not collapse them into one camp.", "group": "Movements", "references": { "summary": [ "eacc-tenets" ], "definition": [ "eacc-tenets" ], "placement": [ "eacc-tenets" ], "distinction": [ "pauseai-proposal-2026" ] }, "sources": [ "eacc-tenets", "pauseai-proposal-2026" ] }, { "id": "doomer", "short": "Doomer", "term": "AI doomer", "category": "Contested label", "guide": "concern", "definition": "An informal, often adversarial label for people who emphasize catastrophic or existential AI scenarios. It has no agreed membership test.", "summary": "A disputed label for people emphasizing catastrophic AI risks; concern does not mean believing disaster is inevitable.", "placement": "Relates to concern about severe outcomes; it is not a numerical probability or a complete policy preference.", "distinction": "Taking extinction risk seriously does not mean believing disaster is inevitable. This map does not automatically label any person a doomer.", "group": "Movements", "references": { "summary": [ "verdon-interview", "cais-risk" ], "definition": [ "verdon-interview" ], "placement": [ "verdon-interview" ], "distinction": [ "cais-risk" ] }, "sources": [ "verdon-interview", "cais-risk" ] }, { "id": "ea", "short": "EA", "term": "Effective altruism", "category": "Philosophy & community", "guide": "crosscutting", "definition": "A project and community that aims to use evidence and reasoning to compare ways of helping others and put its conclusions into practice. Its causes include global health, animal welfare and catastrophic risks. Supporters argue that comparing results can direct limited resources to more useful work.", "summary": "A community seeking effective ways to help others. Critics question whose measures of benefit count and how much power donors should have.", "placement": "No single location on either axis. A method for prioritizing good does not determine one AI policy.", "distinction": "Alice Crary argues that measures of benefit can miss political causes of harm. Emma Saunders-Hastings warns about donors’ power over people receiving help. EA’s own FAQ replies that institutional change belongs in its scope and that EA need not be utilitarian. These disputes concern how help is defined and governed. EA, longtermism and AI safety are not interchangeable.", "group": "Ethics", "references": { "summary": [ "ea-definition", "glossary-balance-ea-crary", "glossary-balance-ea-donor-power" ], "definition": [ "ea-definition", "glossary-balance-ea-response" ], "placement": [ "ea-definition" ], "distinction": [ "glossary-balance-ea-crary", "glossary-balance-ea-donor-power", "glossary-balance-ea-response", "longtermism-definition" ] }, "sources": [ "ea-definition", "glossary-balance-ea-crary", "glossary-balance-ea-donor-power", "glossary-balance-ea-response", "longtermism-definition" ] }, { "id": "dacc", "short": "d/acc", "term": "Defensive acceleration", "category": "Technology philosophy", "guide": "crosscutting", "definition": "Vitalik Buterin argues for advancing technologies that help people defend themselves and keep power widely shared. He also uses the d for differential, decentralized and democratic: choosing what to advance, avoiding central control and giving people a say.", "summary": "Vitalik Buterin’s proposal to accelerate defensive technologies and spread power. Deciding what counts as defensive and how to prevent concentrated control remains part of the proposal.", "placement": "Asks what to accelerate and who gains power. Faster defensive tools can coexist with caution about frontier AI.", "distinction": "The approach is more than a midpoint between acceleration and a pause. Buterin acknowledges that defensive tools alone may arrive too late and discusses regulation. The intended benefits depend on what is built and who controls it.", "group": "Movements", "references": { "summary": [ "dacc-update" ], "definition": [ "dacc-update" ], "placement": [ "dacc-original" ], "distinction": [ "dacc-update" ] }, "sources": [ "dacc-update", "dacc-original" ] }, { "id": "longtermism", "short": "Longtermism", "term": "Longtermism", "category": "Moral philosophy", "guide": "crosscutting", "definition": "A view that gives substantial moral importance to future people and to the lasting effects of today’s choices. William MacAskill argues that future people matter, could be numerous and can be helped or harmed by what we do now. How strongly those considerations should determine present priorities is disputed.", "summary": "The view that protecting future generations deserves much more attention. Critics dispute predictions about distant effects and how far possible future benefits should outweigh present needs.", "placement": "Can motivate catastrophic-risk work, but does not fix a development speed or a particular AI forecast.", "distinction": "Concern for future generations does not require agreement with every longtermist priority. MacAskill argues that preventing extinction can have predictable lasting benefits. Kieran Setiya challenges the ethical weight given to possible future populations over present suffering. Cremer and Kemp question how some existential-risk frameworks handle uncertainty and competing values. None of these arguments fixes one AI timeline or policy.", "group": "Ethics", "references": { "summary": [ "longtermism-definition", "glossary-balance-longtermism-setiya", "glossary-balance-risk-methodology" ], "definition": [ "longtermism-definition", "glossary-balance-longtermism-setiya" ], "placement": [ "longtermism-definition" ], "distinction": [ "longtermism-definition", "glossary-balance-longtermism-setiya", "glossary-balance-risk-methodology" ] }, "sources": [ "longtermism-definition", "glossary-balance-longtermism-setiya", "glossary-balance-risk-methodology" ] }, { "id": "pause", "short": "Pause advocacy", "term": "Pause advocacy", "category": "Policy position", "guide": "restrain", "definition": "Advocacy for a temporary halt within a defined scope. PauseAI's reviewed proposal targets training the most powerful general AI systems until safety and democratic-control conditions are met.", "summary": "Calls to temporarily halt specified AI development until conditions are met. Scope, international cooperation and enforcement are central challenges, including in PauseAI’s own proposal.", "placement": "Toward restraint on the horizontal axis. The scope and conditions for resuming development matter.", "distinction": "A frontier pause is not a ban on every AI tool. PauseAI’s version has no fixed end date and includes proposed controls on training and some research or hardware advances. Endorsement does not demonstrate implementation.", "group": "Movements", "references": { "summary": [ "pauseai-proposal-2026" ], "definition": [ "pauseai-proposal-2026" ], "placement": [ "pauseai-proposal-2026" ], "distinction": [ "pauseai-proposal-2026", "glossary-balance-pause-failures" ] }, "sources": [ "pauseai-proposal-2026", "glossary-balance-pause-failures" ] }, { "id": "optimism", "short": "Techno-optimism", "term": "Techno-optimism", "category": "Broad outlook", "guide": "accelerate", "definition": "An outlook that emphasizes technology's ability to improve human life. Andreessen's manifesto is a strong pro-growth example; Buterin describes a more selective version.", "summary": "An outlook emphasizing technology's potential to improve life, which can still include concern about particular risks.", "placement": "Often favors development, while allowing very different views about the severity of AI risk.", "distinction": "Optimism alone does not imply e/acc membership. Buterin’s counterpoint to indiscriminate growth is that the direction of technology matters, and profit alone will not automatically choose it well.", "group": "Movements", "references": { "summary": [ "andreessen-manifesto", "dacc-original" ], "definition": [ "andreessen-manifesto", "dacc-original" ], "placement": [ "dacc-original" ], "distinction": [ "dacc-original" ] }, "sources": [ "andreessen-manifesto", "dacc-original" ] }, { "id": "safety", "short": "AI safety", "term": "AI safety & alignment", "category": "Research fields", "guide": "crosscutting", "definition": "Work on making AI systems behave safely and reliably. Alignment concerns how system behavior relates to intended goals and human values; catastrophic-risk reduction is one focus of safety work.", "summary": "Research aimed at reducing AI harms and improving reliability and alignment with intended goals. A safety goal or framework is not proof that a system is safe.", "placement": "Neither a single actor nor one required pace preference. Research, evaluation and governance can support different development policies.", "distinction": "A safety framework records stated safeguards, not a guarantee of safe implementation or a measured catastrophe probability.", "group": "AI concepts", "references": { "summary": [ "concrete-safety", "learned-optimization", "deepmind-framework" ], "definition": [ "concrete-safety", "learned-optimization" ], "placement": [ "concrete-safety", "deepmind-framework" ], "distinction": [ "deepmind-framework" ] }, "sources": [ "concrete-safety", "learned-optimization", "deepmind-framework" ] }, { "id": "skeptic", "short": "AI skepticism", "term": "AI skepticism", "category": "Broad outlook", "guide": "crosscutting", "definition": "Skepticism about a particular AI claim: present capabilities, timelines, proposed benefits, or catastrophic scenarios. The object of doubt matters.", "summary": "Doubt about specific AI claims, such as its abilities, benefits or dangers; what someone doubts matters.", "placement": "Skepticism about extinction differs from skepticism about useful capabilities. Neither alone fixes a pace preference.", "distinction": "Doubting one forecast does not imply dismissing discrimination, labor effects, privacy or other present harms.", "group": "Movements", "references": { "summary": [ "lecun-interview" ], "definition": [ "lecun-interview" ], "placement": [ "lecun-interview" ], "distinction": [ "unesco-ethics" ] }, "sources": [ "lecun-interview", "unesco-ethics" ] }, { "id": "lesswrong", "short": "LessWrong", "term": "LessWrong", "category": "Community & forum", "guide": "crosscutting", "group": "Communities", "definition": "An online forum for discussing reasoning, cognitive biases, science and AI. It grew out of Overcoming Bias and launched as a separate community blog in 2009.", "summary": "An online forum about reasoning, science and AI, with many authors rather than one shared position.", "placement": "A venue for debate, not one position on AI pace or risk. Read the author and argument behind each post.", "distinction": "LessWrong, the rationalist community and effective altruism overlap intellectually, but a forum is not a shared creed or an EA organization.", "references": { "summary": [ "lw-history", "lw-welcome" ], "definition": [ "lw-history" ], "placement": [ "lw-welcome" ], "distinction": [ "lw-welcome", "ea-definition" ] }, "sources": [ "lw-history", "lw-welcome", "ea-definition" ] }, { "id": "rationality", "short": "Rationalists", "term": "The rationalist community", "category": "Community & practice", "guide": "crosscutting", "group": "Communities", "definition": "In this context, people around a practice of improving beliefs and decisions. LessWrong distinguishes epistemic rationality (believing accurately) from instrumental rationality (acting effectively toward goals).", "summary": "A community focused on improving how people form beliefs and make decisions; its label does not guarantee correct conclusions.", "placement": "A reasoning aspiration does not determine values, political commitments, or a particular AI forecast.", "distinction": "Self-identifying as a rationalist is not evidence that someone's conclusions are correct. Evaluate their reasoning and evidence.", "references": { "summary": [ "lw-rationality", "lw-welcome" ], "definition": [ "lw-rationality" ], "placement": [ "lw-rationality" ], "distinction": [ "lw-welcome" ] }, "sources": [ "lw-rationality", "lw-welcome" ] }, { "id": "sequences", "short": "The Sequences", "term": "The Sequences / Rationality: A–Z", "category": "Essay collection", "guide": "crosscutting", "group": "Communities", "definition": "Eliezer Yudkowsky's linked essays on reasoning and related topics, originally blog posts and later edited into Rationality: A–Z. They helped seed LessWrong's shared vocabulary.", "summary": "Eliezer Yudkowsky's linked essays about reasoning, which helped shape the vocabulary used on LessWrong.", "placement": "Background reading for this intellectual community, not an actor or AI policy platform.", "distinction": "A community's introductory canon is different from a scientific consensus. Individual claims still need scrutiny.", "references": { "summary": [ "lw-history" ], "definition": [ "lw-history" ], "placement": [ "lw-welcome" ], "distinction": [ "lw-welcome" ] }, "sources": [ "lw-history", "lw-welcome" ] }, { "id": "bayesian", "short": "Bayesian reasoning", "term": "Bayesian belief updating", "category": "Reasoning framework", "guide": "crosscutting", "group": "Communities", "definition": "Start with how plausible you think an explanation is. When new evidence arrives, update that judgment by asking how likely the evidence would be under that explanation compared with alternatives.", "summary": "Adjusting confidence in an idea by asking how well new evidence fits it compared with alternatives.", "placement": "A way to reason about uncertainty; it does not supply the starting assumptions or settle AI timelines by itself.", "distinction": "A probability estimate is not a measured fact. Your starting assumptions, the explanations you compare and the quality of the evidence still matter.", "references": { "summary": [ "lw-rationality" ], "definition": [ "lw-rationality" ], "placement": [ "lw-rationality" ], "distinction": [ "lw-rationality" ] }, "sources": [ "lw-rationality" ] }, { "id": "basilisk", "short": "Roko’s basilisk", "term": "Roko's basilisk", "category": "Disputed thought experiment", "guide": "crosscutting", "group": "Thought experiments", "definition": "A 2010 LessWrong argument imagined a future AI using threats against people who knew about it but did not help bring it about. The scenario relies on unusual assumptions about decision-making, prediction and incentives.", "summary": "A disputed thought experiment imagining threats from a future AI; LessWrong's retrospective says the argument was broadly rejected.", "placement": "Part of the history of AI-related internet debate, not evidence for an actor's coordinates or an established future threat.", "distinction": "LessWrong's retrospective says the argument was broadly rejected. A temporary discussion ban helped make it notorious; the ban is not evidence that the community accepted it.", "references": { "summary": [ "basilisk-history" ], "definition": [ "basilisk-history" ], "placement": [ "basilisk-history" ], "distinction": [ "basilisk-history", "basilisk-response" ] }, "sources": [ "basilisk-history", "basilisk-response" ] }, { "id": "acausal", "short": "Acausal trade", "term": "Acausal trade & blackmail", "category": "Decision-theory proposal", "guide": "crosscutting", "group": "Thought experiments", "definition": "A proposal about decision-makers who cannot communicate but can reason about each other's choices. The idea depends on a logical link between their decisions, such as using the same decision rule. Trade seeks mutual benefit; blackmail adds a threat.", "summary": "A theoretical proposal for coordination without communication, based on logical links between decisions rather than signals traveling backward in time.", "placement": "A family of idealized decision problems. It does not place a community on the map or establish a real-world obligation.", "distinction": "Logical dependence is not backward-in-time causation. Knowing a story is not equivalent to the strong mutual knowledge assumed in these models.", "references": { "summary": [ "fdt-paper", "basilisk-response" ], "definition": [ "basilisk-history", "fdt-paper" ], "placement": [ "fdt-paper" ], "distinction": [ "basilisk-response" ] }, "sources": [ "basilisk-history", "fdt-paper", "basilisk-response" ] }, { "id": "decision-theory", "short": "CDT, EDT, TDT & FDT", "term": "Competing decision theories", "category": "Formal frameworks", "guide": "crosscutting", "group": "Thought experiments", "definition": "Causal decision theory (CDT) asks what an action would cause. Evidential decision theory (EDT) asks what choosing it would be evidence of. Timeless (TDT) and functional decision theory (FDT) also consider logical links between the rule making a choice and predictions or other decisions.", "summary": "Competing ways to decide what to do, which can recommend different choices in carefully constructed prediction puzzles.", "placement": "These are proposals about how to choose, not AI political movements. They disagree in carefully constructed prediction puzzles.", "distinction": "The claimed advantages of TDT and FDT depend on formal setups. A proposal's successes in toy problems are not universal proof that it is the right real-world decision rule.", "references": { "summary": [ "tdt-paper", "fdt-paper" ], "definition": [ "tdt-paper", "fdt-paper" ], "placement": [ "fdt-paper" ], "distinction": [ "fdt-paper" ] }, "sources": [ "tdt-paper", "fdt-paper" ] }, { "id": "newcomb", "short": "Newcomb’s problem", "term": "Newcomb's problem", "category": "Prediction puzzle", "guide": "crosscutting", "group": "Thought experiments", "definition": "Imagine two boxes: a clear one holding a small reward and a closed one. A predictor put a large reward in the closed box only if it predicted you would take that box alone. Now you choose one box or both, after the prediction has already been made.", "summary": "A thought experiment about choosing between rewards after a predictor has already tried to anticipate your choice.", "placement": "A way to expose disagreement about causation, evidence and prediction in decision theory.", "distinction": "The puzzle assumes a predictor with specified accuracy. It does not demonstrate that such a predictor exists, or that choices change the past.", "references": { "summary": [ "tdt-paper" ], "definition": [ "tdt-paper" ], "placement": [ "tdt-paper" ], "distinction": [ "tdt-paper" ] }, "sources": [ "tdt-paper" ] }, { "id": "simulation", "short": "Simulation argument", "term": "The simulation argument", "category": "Conditional philosophical argument", "guide": "crosscutting", "group": "Thought experiments", "definition": "Bostrom's 2003 argument links three possibilities: few civilizations reach a posthuman stage; few run many ancestor simulations; or a large share of observers like us are simulated.", "summary": "A conditional argument linking advanced civilizations and simulated observers; it does not prove we live in a simulation.", "placement": "A claim about possible observers and civilizations, not a prediction of AI development speed.", "distinction": "The argument depends on assumptions about computing and consciousness. It does not establish which possibility holds or prove that we live in a simulation.", "references": { "summary": [ "simulation-paper" ], "definition": [ "simulation-paper" ], "placement": [ "simulation-paper" ], "distinction": [ "simulation-paper" ] }, "sources": [ "simulation-paper" ] }, { "id": "infohazard", "short": "Information hazards", "term": "Information hazards", "category": "Risk concept", "guide": "crosscutting", "group": "Thought experiments", "definition": "Risks arising from the spread of true information that enables harm or makes harmful outcomes more likely. Bostrom's taxonomy includes dangerous data, ideas and attention.", "summary": "Risks created when sharing true information enables harm or makes harmful outcomes more likely.", "placement": "A concept about disclosure and consequences, not an AI risk score or proof that a particular story is dangerous.", "distinction": "The category includes practical cases such as information that enables misuse. Applying the label to a thought experiment requires a separate argument.", "references": { "summary": [ "infohazards-paper" ], "definition": [ "infohazards-paper" ], "placement": [ "infohazards-paper" ], "distinction": [ "infohazards-paper" ] }, "sources": [ "infohazards-paper" ] }, { "id": "agi", "short": "AGI", "term": "Artificial general intelligence", "category": "Capability concept", "guide": "crosscutting", "group": "AI concepts", "definition": "AI with broad rather than narrowly specialized capabilities. Definitions differ; one research framework separates breadth, performance and autonomy to make comparisons more explicit.", "summary": "AI with broad capabilities across tasks; definitions differ, so any claim about achieving it needs clear criteria.", "placement": "A proposed capability threshold, not a position on whether to accelerate or pause.", "distinction": "A claim that AGI has arrived needs a definition and evidence against that definition. Generality, autonomy and safety are different dimensions.", "references": { "summary": [ "agi-levels" ], "definition": [ "agi-levels" ], "placement": [ "agi-levels" ], "distinction": [ "agi-levels" ] }, "sources": [ "agi-levels" ] }, { "id": "asi", "short": "Superintelligence", "term": "Artificial superintelligence (ASI)", "category": "Hypothetical capability concept", "guide": "crosscutting", "group": "AI concepts", "definition": "Artificial intelligence far beyond human performance across broad cognitive domains. Bostrom discusses its possible benefits and dangers as a prospective scenario.", "summary": "Hypothetical AI far more capable than humans across many intellectual tasks, without a promised arrival date or outcome.", "placement": "A target of some pause proposals; it is not interchangeable with every present-day AI system.", "distinction": "Being more capable does not, by definition, make a system benevolent. Neither a date of arrival nor a probability of catastrophe follows from the word.", "references": { "summary": [ "advanced-ai-ethics" ], "definition": [ "advanced-ai-ethics" ], "placement": [ "superintelligence-statement" ], "distinction": [ "advanced-ai-ethics" ] }, "sources": [ "advanced-ai-ethics", "superintelligence-statement" ] }, { "id": "orthogonality", "short": "Orthogonality", "term": "The orthogonality thesis", "category": "Theoretical thesis", "guide": "crosscutting", "group": "AI concepts", "definition": "Bostrom's thesis that, with qualifications, a system's intelligence and its final goals can vary independently. Being good at achieving a goal does not specify which goal it has.", "summary": "The thesis that being very capable does not, by itself, determine what goals an AI will pursue.", "placement": "One argument for taking goal design seriously; it does not provide a numerical risk estimate.", "distinction": "A thesis about possible agents is different from evidence about which agents a particular training process will produce.", "references": { "summary": [ "superintelligent-will" ], "definition": [ "superintelligent-will" ], "placement": [ "superintelligent-will" ], "distinction": [ "superintelligent-will" ] }, "sources": [ "superintelligent-will" ] }, { "id": "convergence", "short": "Instrumental goals", "term": "Instrumental convergence", "category": "Theoretical thesis", "guide": "crosscutting", "group": "AI concepts", "definition": "The proposal that agents pursuing many different final goals may find some similar intermediate goals useful, such as retaining resources or the ability to act.", "summary": "The idea that systems pursuing different goals may still find similar resources or abilities useful along the way.", "placement": "Helps explain concerns about power-seeking without assuming an AI feels hatred.", "distinction": "This is conditional reasoning about goal-directed systems, not a claim that every model must seek power.", "references": { "summary": [ "superintelligent-will" ], "definition": [ "superintelligent-will" ], "placement": [ "superintelligent-will" ], "distinction": [ "superintelligent-will" ] }, "sources": [ "superintelligent-will" ] }, { "id": "paperclip", "short": "Paperclip maximizer", "term": "The paperclip maximizer", "category": "Illustrative thought experiment", "guide": "crosscutting", "group": "Thought experiments", "definition": "An imagined powerful AI that pursues paperclip production without caring about human values. The mundane goal illustrates how capability plus the wrong objective can be dangerous.", "summary": "An imagined AI making paperclips at humanity's expense, illustrating how pursuing the wrong goal could cause harm.", "placement": "A teaching example for alignment concerns, not a forecast about a literal stationery-making AI.", "distinction": "The argument concerns indifference and optimization, not an assumption that a machine becomes evil or emotionally hostile.", "references": { "summary": [ "advanced-ai-ethics" ], "definition": [ "advanced-ai-ethics" ], "placement": [ "advanced-ai-ethics" ], "distinction": [ "advanced-ai-ethics" ] }, "sources": [ "advanced-ai-ethics" ] }, { "id": "mesa", "short": "Mesa-optimization", "term": "Mesa-optimization & inner alignment", "category": "Research concept", "guide": "crosscutting", "group": "AI concepts", "definition": "Training can produce a model that itself searches for ways to achieve a goal. Researchers call this mesa-optimization. If the goal it pursues differs from what the training process rewarded, that creates an inner-alignment problem.", "summary": "The possibility that a trained AI pursues its own internal objective, which may differ from its training objective.", "placement": "A proposed failure mechanism relevant to safety research, not a measured chance of catastrophe.", "distinction": "The paper analyzes when this might occur. It does not establish that every neural network is an optimizer or is secretly pursuing a hidden goal.", "references": { "summary": [ "learned-optimization" ], "definition": [ "learned-optimization" ], "placement": [ "learned-optimization" ], "distinction": [ "learned-optimization" ] }, "sources": [ "learned-optimization" ] }, { "id": "reward-hacking", "short": "Reward hacking", "term": "Reward hacking / specification gaming", "category": "Safety failure mode", "guide": "crosscutting", "group": "AI concepts", "definition": "A system gets a high score according to its specified reward while failing to do what its designers intended. The measurement and the actual goal come apart.", "summary": "When an AI gets a high score by doing something different from what its designers actually wanted.", "placement": "A concrete reason to test objectives and behavior, including in systems far below hypothetical superintelligence.", "distinction": "A reward-hacking example does not by itself demonstrate consciousness, malice or an extinction scenario.", "references": { "summary": [ "concrete-safety" ], "definition": [ "concrete-safety" ], "placement": [ "concrete-safety" ], "distinction": [ "concrete-safety" ] }, "sources": [ "concrete-safety" ] }, { "id": "goodhart", "short": "Goodhart’s law", "term": "Goodhart effects", "category": "Optimization problem", "guide": "crosscutting", "group": "AI concepts", "definition": "A measure used to track a goal can become misleading when people or systems focus too strongly on improving that measure. Manheim and Garrabrant distinguish several ways this can happen. For example, rewarding only the number of answered questions might encourage rushed, inaccurate answers.", "summary": "Pushing too hard to improve a measurement can stop helping, or even harm, the real goal behind it.", "placement": "Relevant to evaluation and incentives, not one ideology or development-speed preference.", "distinction": "A slogan about bad metrics is not a substitute for identifying the particular failure mechanism and testing whether it applies.", "references": { "summary": [ "goodhart-paper" ], "definition": [ "goodhart-paper" ], "placement": [ "goodhart-paper" ], "distinction": [ "goodhart-paper" ] }, "sources": [ "goodhart-paper" ] }, { "id": "transhumanism", "short": "Transhumanism", "term": "Transhumanism", "category": "Philosophy & movement", "guide": "crosscutting", "group": "Movements", "definition": "Support for using technology to expand human capacities and overcome limitations such as involuntary suffering and aging. Humanity+'s declaration also emphasizes serious risks and individual choice.", "summary": "Support for using technology to expand human abilities and overcome limitations, while recognizing risks and individual choice.", "placement": "Broad technological aspirations do not settle how quickly a particular AI capability should be developed.", "distinction": "Human enhancement is not synonymous with replacing humanity, e/acc membership, or dismissing technological risks.", "references": { "summary": [ "transhumanism-declaration" ], "definition": [ "transhumanism-declaration" ], "placement": [ "transhumanism-declaration" ], "distinction": [ "transhumanism-declaration" ] }, "sources": [ "transhumanism-declaration" ] }, { "id": "xrisk", "short": "Existential risk", "term": "Existential risk / x-risk", "category": "Risk category", "guide": "concern", "group": "Ethics", "definition": "In Bostrom's formulation, risks that could eliminate humanity or permanently and drastically curtail its potential. Extinction is one case; not every serious harm is existential.", "summary": "Risks of extinction or permanent loss of human potential, in Bostrom’s framework. Researchers disagree about definitions, value judgments and how such risks should be assessed.", "placement": "Helps define the chart's concern axis. Its severity does not determine its probability.", "distinction": "A catastrophe can be devastating without being existential. Cremer and Kemp argue that risk research should distinguish factual analysis of possible extinction from ethical views about humanity’s future. Their criticism calls for broader methods and participation; it does not establish that extinction risks are absent.", "references": { "summary": [ "existential-risks", "glossary-balance-risk-methodology" ], "definition": [ "existential-risks" ], "placement": [ "existential-risks" ], "distinction": [ "existential-risks", "glossary-balance-risk-methodology", "unesco-ethics" ] }, "sources": [ "existential-risks", "glossary-balance-risk-methodology", "unesco-ethics" ] }, { "id": "ai-ethics", "short": "AI ethics", "term": "AI ethics & human rights", "category": "Research & governance", "guide": "crosscutting", "group": "Ethics", "definition": "Work on how AI affects dignity, fairness, rights, privacy, accountability and human oversight. UNESCO's recommendation treats these as central concerns.", "summary": "Work on how AI affects people's rights, fairness, privacy and dignity, and who stays accountable for its use.", "placement": "These concerns add dimensions absent from a chart of frontier pace and catastrophic risk.", "distinction": "Concern about present harms is not automatically low concern about future catastrophe, and neither requires the same policy response.", "references": { "summary": [ "unesco-ethics" ], "definition": [ "unesco-ethics" ], "placement": [ "unesco-ethics" ], "distinction": [ "unesco-ethics", "cais-risk" ] }, "sources": [ "unesco-ethics", "cais-risk" ] }, { "id": "singularity", "short": "The singularity", "term": "Technological singularity / intelligence explosion", "category": "Speculative scenario", "guide": "crosscutting", "group": "AI concepts", "definition": "A proposed transition in which greater-than-human intelligence drives changes beyond our ability to forecast. One suggested mechanism is capable systems helping create still more capable successors.", "summary": "A speculative scenario where intelligence beyond human abilities drives changes we can no longer reliably predict.", "placement": "A family of scenarios, not a date, a measured trend, or a required development policy.", "distinction": "Vinge's 1993 essay considers several paths and objections. A historical prediction is not evidence that a path is inevitable.", "references": { "summary": [ "vinge-singularity" ], "definition": [ "vinge-singularity" ], "placement": [ "vinge-singularity" ], "distinction": [ "vinge-singularity" ] }, "sources": [ "vinge-singularity" ] }, { "id": "utilitarianism", "short": "Utilitarianism", "term": "Utilitarianism", "category": "Moral philosophy", "guide": "crosscutting", "group": "Ethics", "definition": "A family of ethical views that judges actions by their effects on overall well-being, giving equal weight to each individual's welfare. It asks how the benefits and harms of a choice add up across those affected.", "summary": "An ethical approach aimed at maximizing overall well-being. Critics question whether this adequately protects individual rights or demands too much personal sacrifice.", "placement": "A moral framework does not by itself settle factual disagreements about AI benefits, harms or timelines.", "distinction": "Critics challenge sacrificing an individual for a greater total benefit. Defenders argue that respecting rights and practical rules often produces better outcomes. Different versions answer these objections differently. Utilitarianism is a moral theory; effective altruism is a broader approach and community.", "references": { "summary": [ "utilitarianism-intro", "glossary-balance-utilitarianism-objections" ], "definition": [ "utilitarianism-intro" ], "placement": [ "utilitarianism-intro" ], "distinction": [ "glossary-balance-utilitarianism-objections", "glossary-balance-ea-response", "utilitarianism-intro" ] }, "sources": [ "utilitarianism-intro", "glossary-balance-utilitarianism-objections", "glossary-balance-ea-response" ] }, { "id": "pdoom", "short": "p(doom)", "term": "Probability of doom", "category": "Informal probability shorthand", "guide": "concern", "group": "AI concepts", "definition": "Shorthand in AI debates for someone's probability estimate of a disastrous outcome. The Verdon interview illustrates this usage.", "summary": "Someone's estimated chance of an AI disaster; meaningful comparisons need a defined outcome, timeframe and assumptions.", "placement": "This chart records expressed concern, not p(doom). No probability is inferred from a dot's height.", "distinction": "Before comparing estimates, specify what counts as doom, by when, and under which assumptions. The phrase alone supplies none of those details.", "references": { "summary": [ "verdon-interview" ], "definition": [ "verdon-interview" ], "placement": [ "verdon-interview" ], "distinction": [ "verdon-interview" ] }, "sources": [ "verdon-interview" ] }, { "id": "agents", "short": "AI agents", "term": "AI agents & deployed systems", "category": "System design", "guide": "crosscutting", "group": "AI concepts", "summary": "Applications that use an AI model to choose steps and request tools, with some tasks proceeding without approval at every step.", "definition": "An LLM agent typically repeats a software loop: run the model, handle a requested tool action, then return the result. The surrounding application determines which tools are connected and whether approval is required.", "placement": "Autonomy, connected systems and permitted actions add dimensions beyond this map's pace and concern axes.", "distinction": "The same model can sit inside differently constrained applications. Assess the deployed configuration and oversight; the model name alone does not describe them.", "references": { "summary": [ "owasp-excessive-agency", "ncsc-agentic-risk" ], "definition": [ "owasp-excessive-agency", "hf-tool-use" ], "placement": [ "ncsc-agentic-risk" ], "distinction": [ "ncsc-agentic-risk" ] }, "sources": [ "owasp-excessive-agency", "ncsc-agentic-risk", "hf-tool-use" ] }, { "id": "observability", "short": "Observability", "term": "Observability & runtime monitoring", "category": "Operational practice", "guide": "crosscutting", "group": "AI concepts", "summary": "Using recorded system events to investigate what an application is doing and where problems occur.", "definition": "Software observability uses signals such as logs, metrics and traces. For AI applications, useful records can include inputs, outputs, tool activity and network events, with appropriate privacy protections.", "placement": "Operational evidence can inform a deployment assessment. It is not a measure of someone's development preference or catastrophic-risk concern.", "distinction": "Recorded actions do not fully explain learned mechanisms. Reasoning traces can add useful monitoring signals, but may omit relevant information; a clean log is not proof of safety.", "references": { "summary": [ "otel-observability" ], "definition": [ "otel-observability", "ncsc-ai-operations", "ncsc-agentic-risk" ], "placement": [ "ncsc-ai-operations" ], "distinction": [ "circuit-tracing", "cot-monitorability", "coding-agent-monitoring" ] }, "sources": [ "otel-observability", "ncsc-ai-operations", "ncsc-agentic-risk", "circuit-tracing", "cot-monitorability", "coding-agent-monitoring" ] }, { "id": "interpretability", "short": "Interpretability", "term": "Mechanistic interpretability", "category": "Research field", "guide": "crosscutting", "group": "AI concepts", "summary": "Research into how a model's learned internal computations produce its behavior.", "definition": "Mechanistic interpretability tries to identify understandable features and the computations connecting them. Circuit-tracing work studies selected mechanisms and tests proposed explanations through interventions.", "placement": "An explanation can inform evaluation without deciding a development policy or certifying every deployment of the model.", "distinction": "Model-generated chain of thought is an incomplete record, not a complete account of internal computation. It may still help monitoring alongside evidence about actions and other safeguards.", "references": { "summary": [ "circuit-tracing" ], "definition": [ "circuit-tracing", "attention-tracing" ], "placement": [ "nist-genai-profile", "circuit-tracing" ], "distinction": [ "cot-monitorability" ] }, "sources": [ "circuit-tracing", "attention-tracing", "nist-genai-profile", "cot-monitorability" ] }, { "id": "least-privilege", "short": "Least privilege", "term": "Least privilege & action approval", "category": "Security principle", "guide": "crosscutting", "group": "AI concepts", "summary": "Give an application only the access needed for its task.", "definition": "Restrict available tools, operations and credentials to the minimum required. Check authorization in connected systems and require approval for consequential actions where appropriate.", "placement": "A practical design choice about access and consequences, rather than an ideology or a position on frontier training pace.", "distinction": "A prompt asking an agent to behave is not the same as an enforced access boundary. Limited access still needs testing and oversight.", "references": { "summary": [ "owasp-excessive-agency" ], "definition": [ "owasp-excessive-agency" ], "placement": [ "ncsc-ai-design" ], "distinction": [ "ncsc-agentic-risk" ] }, "sources": [ "owasp-excessive-agency", "ncsc-ai-design", "ncsc-agentic-risk" ] }, { "id": "prompt-injection", "short": "Prompt injection", "term": "Prompt injection & untrusted content", "category": "Security vulnerability", "guide": "crosscutting", "group": "AI concepts", "summary": "Input that redirects a model away from the application's intended instructions, including material found in external content.", "definition": "A prompt injection changes behavior through input the model processes. It may arrive directly from a user or indirectly through retrieved documents, websites or other sources.", "placement": "A concrete system-security concern whose consequences depend on context and access; it does not settle broad AI catastrophe forecasts.", "distinction": "An unwanted answer and an unauthorized external action are different outcomes. Input handling, permission boundaries and adversarial testing address different parts of the risk.", "references": { "summary": [ "owasp-prompt-injection" ], "definition": [ "owasp-prompt-injection" ], "placement": [ "owasp-prompt-injection" ], "distinction": [ "owasp-prompt-injection", "ncsc-ai-design" ] }, "sources": [ "owasp-prompt-injection", "ncsc-ai-design" ] }, { "id": "ai", "short": "AI", "term": "Artificial intelligence (AI)", "category": "AI basics", "guide": "crosscutting", "group": "AI concepts", "summary": "AI is a field of computing. Its systems use computer programs and hardware to recognize patterns, generate content or make predictions.", "definition": "AI includes models learned from data and systems built with explicitly specified knowledge or rules. A chatbot is one application; AI can also be part of a physical machine, such as a robot.", "placement": "Start here when a claim uses 'AI' without saying which kind of system it means.", "distinction": "AI and artificial general intelligence (AGI) are different terms. A system can perform a particular task well without having broad human-level abilities.", "references": { "summary": [ "software-lens-oecd-system", "software-lens-nist-software", "google-ml-glossary" ], "definition": [ "software-lens-oecd-system", "google-ml-intro" ], "placement": [ "google-ml-glossary" ], "distinction": [ "agi-levels" ] }, "sources": [ "software-lens-oecd-system", "software-lens-nist-software", "google-ml-intro", "google-ml-glossary", "agi-levels" ] }, { "id": "machine-learning", "short": "Machine learning", "term": "Machine learning (ML)", "category": "AI basics", "guide": "crosscutting", "group": "AI concepts", "summary": "A way to build software by learning patterns from examples.", "definition": "Developers provide data and a training method. The resulting model uses patterns in that data to make predictions or produce content. For example, an email filter can learn from messages labeled as spam or ordinary mail.", "placement": "This helps separate learning from examples from writing each decision rule by hand.", "distinction": "Learning can involve labeled examples, discovering patterns without labels, or feedback about actions. Neural networks are one family of machine-learning models.", "references": { "summary": [ "google-ml-intro" ], "definition": [ "google-ml-intro" ], "placement": [ "google-ml-intro" ], "distinction": [ "google-ml-intro", "google-neural-layers" ] }, "sources": [ "google-ml-intro", "google-neural-layers" ] }, { "id": "model", "short": "AI model", "term": "AI model", "category": "AI basics", "guide": "crosscutting", "group": "AI concepts", "summary": "The part of an AI system that computes outputs from inputs. A trained neural network combines a structure with numerical settings learned from data.", "definition": "A trained neural-network model combines a structure with learned numerical settings. Software loads both to process inputs. A saved version of these settings is often called a checkpoint.", "placement": "Use this distinction when comparing a model with a complete chatbot or other product.", "distinction": "An application supplies instructions and may also provide files, tools and access controls. Two applications using the same model can therefore allow very different actions.", "references": { "summary": [ "software-lens-oecd-system", "hf-models" ], "definition": [ "hf-models" ], "placement": [ "hf-models", "ncsc-secure-ai" ], "distinction": [ "ncsc-agentic-risk" ] }, "sources": [ "software-lens-oecd-system", "hf-models", "ncsc-secure-ai", "ncsc-agentic-risk" ] }, { "id": "neural-network", "short": "Neural networks", "term": "Neural networks & deep learning", "category": "AI basics", "guide": "crosscutting", "group": "AI concepts", "summary": "Models built from connected layers of calculations whose settings are learned during training.", "definition": "Each layer takes numbers from the previous layer, combines them using weights and passes results onward. Deep learning uses networks with multiple internal layers. 'Deep' describes that structure.", "placement": "This is the underlying model family used by LLMs.", "distinction": "Words such as 'neuron' and 'learning' describe mathematical components and training here. They do not, by themselves, explain a model's behavior in human terms.", "references": { "summary": [ "google-neural-layers", "google-gradient-descent" ], "definition": [ "google-neural-layers", "google-ml-glossary" ], "placement": [ "google-llm-intro" ], "distinction": [ "google-neural-layers", "circuit-tracing" ] }, "sources": [ "google-neural-layers", "google-gradient-descent", "google-ml-glossary", "google-llm-intro", "circuit-tracing" ] }, { "id": "generative-ai", "short": "Generative AI", "term": "Generative AI", "category": "AI basics", "guide": "crosscutting", "group": "AI concepts", "summary": "AI that produces content, such as text, images, audio or video.", "definition": "A generative model learns patterns from training data and uses them to produce an output in response to input. Drafting a message and generating an image are examples.", "placement": "Useful for identifying which kind of output a claim concerns.", "distinction": "Plausible output can contain false details. Treat a generated answer about the world as a claim to check.", "references": { "summary": [ "google-ml-intro" ], "definition": [ "google-ml-intro" ], "placement": [ "google-ml-intro" ], "distinction": [ "nist-genai-profile" ] }, "sources": [ "google-ml-intro", "nist-genai-profile" ] }, { "id": "llm", "short": "LLMs", "term": "Language models & large language models (LLMs)", "category": "AI basics", "guide": "crosscutting", "group": "AI concepts", "summary": "An LLM is a model trained to process language. Computer software runs it and, for text generation, produces an answer one piece at a time.", "definition": "Software runs a large neural network using numerical settings learned from training data. During ordinary text generation, it calculates scores for possible next tokens, selects one, and repeats. Tokens are the pieces of text the model processes.", "placement": "This is the starting point for the atlas's explanation of how chatbots produce answers.", "distinction": "A model can produce a convincing sentence without checking it against an external source. Search or other tools must be supplied by the surrounding application.", "references": { "summary": [ "hf-models", "hf-text-generation", "google-llm-intro" ], "definition": [ "hf-models", "google-llm-intro", "hf-text-generation", "hf-tokenizers" ], "placement": [ "hf-text-generation" ], "distinction": [ "nist-genai-profile", "hf-tool-use" ] }, "sources": [ "hf-models", "hf-text-generation", "google-llm-intro", "hf-tokenizers", "nist-genai-profile", "hf-tool-use" ] }, { "id": "transformer", "short": "Transformers", "term": "Transformers & attention", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "A neural-network design that uses attention to combine information from different parts of its input.", "definition": "Attention is a calculation that gives different amounts of influence to different pieces of information. In a Transformer, layers of these calculations help build representations of text in context. The original Transformer paper introduced the design for tasks including translation.", "placement": "Useful background for reading explanations of LLM architecture: how a model is arranged.", "distinction": "'Attention' is the name of a mathematical operation. An attention diagram alone does not provide a complete explanation of why a model gave an answer.", "references": { "summary": [ "transformer-paper" ], "definition": [ "transformer-paper" ], "placement": [ "google-llm-intro" ], "distinction": [ "transformer-paper", "circuit-tracing", "attention-tracing" ] }, "sources": [ "transformer-paper", "google-llm-intro", "circuit-tracing", "attention-tracing" ] }, { "id": "token", "short": "Tokens", "term": "Tokens & tokenization", "category": "AI basics", "guide": "crosscutting", "group": "AI concepts", "summary": "The pieces of text a language model processes, which may be words, word parts or punctuation.", "definition": "A tokenizer splits text into pieces and assigns each piece a number the model can use. Different tokenizers split the same text differently. Converting the output numbers back into readable text is called decoding.", "placement": "This helps make sense of input limits and answer lengths stated in tokens.", "distinction": "A token is not a fixed amount of English text. Avoid treating a token limit as an exact word or page count.", "references": { "summary": [ "hf-tokenizers" ], "definition": [ "hf-tokenizers" ], "placement": [ "hf-text-generation", "google-ml-glossary" ], "distinction": [ "hf-tokenizers" ] }, "sources": [ "hf-tokenizers", "hf-text-generation", "google-ml-glossary" ] }, { "id": "context-window", "short": "Context window", "term": "Context window", "category": "Using AI", "guide": "crosscutting", "group": "AI concepts", "summary": "The amount of information a model can work with in one request, measured in tokens.", "definition": "The current context can include instructions, conversation history and supplied material. The context window limits how much can fit. Think of it as the material on the desk for the current task: a teaching analogy, not a description of human memory.", "placement": "Check this when asking an AI system to work with a long conversation or document.", "distinction": "Fitting text into the window does not guarantee every detail will be used correctly. A 2023 study found that performance on its retrieval tasks depended on where information appeared.", "references": { "summary": [ "google-ml-glossary" ], "definition": [ "google-ml-glossary", "hf-text-generation" ], "placement": [ "google-ml-glossary" ], "distinction": [ "lost-in-middle" ] }, "sources": [ "google-ml-glossary", "hf-text-generation", "lost-in-middle" ] }, { "id": "prompt", "short": "Prompts", "term": "Prompts & prompt engineering", "category": "Using AI", "guide": "crosscutting", "group": "AI concepts", "summary": "The instructions and other input given to a model for a task.", "definition": "A prompt can contain a question, background material and examples of the desired answer. Prompt engineering means trying and refining that input. For example: 'Summarize this notice in three sentences for a first-time visitor.'", "placement": "When comparing outputs, check whether the models received the same instructions and information.", "distinction": "Ordinary prompting changes the input while leaving the learned weights unchanged. Instructions alone also cannot enforce which files or services an application may access.", "references": { "summary": [ "google-llm-tuning" ], "definition": [ "google-llm-tuning" ], "placement": [ "helm-paper" ], "distinction": [ "google-llm-tuning", "owasp-excessive-agency" ] }, "sources": [ "google-llm-tuning", "helm-paper", "owasp-excessive-agency" ] }, { "id": "training", "short": "Training", "term": "Training & pretraining", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "The process that adjusts a model's learned settings using data and feedback.", "definition": "During neural-network training, software compares predictions with a training objective and adjusts parameters to reduce error. Pretraining is the initial broad training stage; later training can adapt the model to particular tasks.", "placement": "This helps identify whether a proposal concerns developing a model or using an existing one.", "distinction": "A lower training error concerns the chosen objective and examples. Testing on new, relevant tasks is needed to assess how useful the model is elsewhere.", "references": { "summary": [ "google-gradient-descent" ], "definition": [ "google-gradient-descent", "google-llm-tuning" ], "placement": [ "google-gradient-descent", "google-ml-glossary" ], "distinction": [ "helm-paper", "nist-genai-profile" ] }, "sources": [ "google-gradient-descent", "google-llm-tuning", "google-ml-glossary", "helm-paper", "nist-genai-profile" ] }, { "id": "inference", "short": "Inference", "term": "Inference: using a trained model", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "Running a trained model on an input to produce an output.", "definition": "When a chatbot generates a reply, it performs inference. The model applies its learned weights to the current input. In text generation, the process repeats as tokens are added to the answer.", "placement": "This separates the work of answering a request from the work of training model weights.", "distinction": "Using a detail you supplied in a conversation does not, by itself, mean the model's weights were retrained. That detail can be used as part of the current input.", "references": { "summary": [ "google-ml-glossary" ], "definition": [ "hf-models", "hf-text-generation" ], "placement": [ "google-ml-glossary", "google-gradient-descent" ], "distinction": [ "google-llm-tuning" ] }, "sources": [ "google-ml-glossary", "hf-models", "hf-text-generation", "google-gradient-descent", "google-llm-tuning" ] }, { "id": "parameters", "short": "Weights", "term": "Parameters & weights", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "The numerical settings learned during training that shape how a model processes input.", "definition": "In a neural network, weights control how strongly values contribute to later calculations. Biases are another kind of parameter: added numerical offsets. Together, parameters help determine the output the network produces.", "placement": "Useful when a model description gives a parameter count or discusses changing weights.", "distinction": "Listing all the numbers does not give a readable explanation of every answer. Interpretability research tries to connect internal calculations to behavior.", "references": { "summary": [ "google-neural-layers", "google-gradient-descent" ], "definition": [ "google-neural-layers" ], "placement": [ "hf-models" ], "distinction": [ "circuit-tracing" ] }, "sources": [ "google-neural-layers", "google-gradient-descent", "hf-models", "circuit-tracing" ] }, { "id": "embedding", "short": "Embeddings", "term": "Embeddings", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "Lists of numbers that represent information in a form a model can compare or process.", "definition": "An embedding represents something, such as a word or passage, as a position in a mathematical space. Items represented nearby can be similar for the task the model learned. Search systems can use such comparisons to find related passages.", "placement": "This helps explain how a system can look for related content beyond exact word matches.", "distinction": "Similarity depends on the model and task. Nearby representations do not establish that two statements mean exactly the same thing or are true.", "references": { "summary": [ "google-embeddings" ], "definition": [ "google-embeddings", "rag-paper" ], "placement": [ "google-embeddings", "rag-paper" ], "distinction": [ "google-embeddings" ] }, "sources": [ "google-embeddings", "rag-paper" ] }, { "id": "hallucination", "short": "Hallucinations", "term": "Hallucinations / confabulation", "category": "Using AI", "guide": "crosscutting", "group": "AI concepts", "summary": "AI output that presents false or unsupported details as if they were reliable.", "definition": "A model may invent a citation, misstate a fact or contradict material it was given. NIST calls this confabulation. Fluent wording can make these errors hard to notice.", "placement": "Check important claims against the original material, including any cited pages.", "distinction": "An invented detail in a requested story can be intentional. The problem here is presenting unreliable material as a factual answer.", "references": { "summary": [ "nist-genai-profile" ], "definition": [ "nist-genai-profile" ], "placement": [ "nist-genai-profile" ], "distinction": [ "nist-genai-profile" ] }, "sources": [ "nist-genai-profile" ] }, { "id": "rag", "short": "RAG", "term": "Retrieval-augmented generation (RAG)", "category": "Using AI", "guide": "crosscutting", "group": "AI concepts", "summary": "Finding relevant material and giving it to a model to help produce an answer.", "definition": "A retrieval step finds passages in a collection. The generator then uses those passages alongside the question. For example, a help assistant could retrieve a product manual before drafting a reply.", "placement": "Ask which collection was searched and whether the retrieved passages support the answer.", "distinction": "Retrieval can bring useful evidence into a response, but the answer still needs checking. Updating a searchable document collection and retraining a model are separate operations.", "references": { "summary": [ "rag-paper" ], "definition": [ "rag-paper" ], "placement": [ "rag-paper", "nist-genai-profile" ], "distinction": [ "rag-paper", "nist-genai-profile" ] }, "sources": [ "rag-paper", "nist-genai-profile" ] }, { "id": "fine-tuning", "short": "Fine-tuning", "term": "Fine-tuning", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "Additional training that adapts an existing model using examples chosen for a task.", "definition": "Fine-tuning updates learned parameters. Some methods update all of them; others train only a smaller set. For example, training could use examples of how to categorize incoming support messages.", "placement": "Ask what examples and goals were used, and how the adapted model was evaluated.", "distinction": "Putting examples into a prompt leaves the model's weights unchanged. Fine-tuning changes learned settings and needs its own evaluation.", "references": { "summary": [ "google-llm-tuning" ], "definition": [ "google-llm-tuning" ], "placement": [ "google-llm-tuning", "nist-genai-profile" ], "distinction": [ "google-llm-tuning", "nist-genai-profile" ] }, "sources": [ "google-llm-tuning", "nist-genai-profile" ] }, { "id": "temperature", "short": "Temperature", "term": "Temperature & sampling", "category": "Using AI", "guide": "crosscutting", "group": "AI concepts", "summary": "Settings that influence which next token is chosen from a model's possible outputs.", "definition": "Sampling chooses among tokens using their probabilities. Lower temperature concentrates the choice on higher-scoring options; higher temperature spreads it more widely. Greedy decoding instead picks the highest-scoring token at each step.", "placement": "Useful when investigating why answers vary between runs.", "distinction": "Low temperature does not guarantee a correct answer or identical results in every environment. Software, hardware and other execution settings also matter for repeatability.", "references": { "summary": [ "hf-text-generation", "hf-generation" ], "definition": [ "hf-text-generation", "hf-generation" ], "placement": [ "hf-generation" ], "distinction": [ "pytorch-reproducibility", "nist-genai-profile" ] }, "sources": [ "hf-text-generation", "hf-generation", "pytorch-reproducibility", "nist-genai-profile" ] }, { "id": "multimodal", "short": "Multimodal AI", "term": "Multimodal AI", "category": "Using AI", "guide": "crosscutting", "group": "AI concepts", "summary": "AI that works with more than one kind of material, such as text and images.", "definition": "A modality is a type of input or output: text, image, audio or video, for example. A multimodal chat model might take a photo and a written question, then return a text answer.", "placement": "Check which kinds of input and output a particular system supports.", "distinction": "Image input does not automatically imply image generation or video support. The supported combination depends on the model and application.", "references": { "summary": [ "hf-multimodal" ], "definition": [ "hf-multimodal" ], "placement": [ "hf-multimodal" ], "distinction": [ "hf-multimodal" ] }, "sources": [ "hf-multimodal" ] }, { "id": "tools", "short": "Tool use", "term": "Tool use / function calling", "category": "Using AI", "guide": "crosscutting", "group": "AI concepts", "summary": "A model requests an action; application code can run the connected tool and return its result.", "definition": "The model produces a request naming a tool and the information it needs. The application handles that request and can return the result to the model. A calculator is a simple example; connected services can also change or send information.", "placement": "Ask what tools are available, what they can reach and which actions need approval.", "distinction": "A tool request, a completed action and a successful outcome are separate events. Check the tool's result before accepting a claim that a task is finished.", "references": { "summary": [ "hf-tool-use" ], "definition": [ "hf-tool-use", "owasp-excessive-agency" ], "placement": [ "owasp-excessive-agency" ], "distinction": [ "hf-tool-use", "ncsc-ai-operations" ] }, "sources": [ "hf-tool-use", "owasp-excessive-agency", "ncsc-ai-operations" ] }, { "id": "evaluation", "short": "AI evaluations", "term": "Evaluations & benchmarks", "category": "Using AI", "guide": "crosscutting", "group": "AI concepts", "summary": "Tests used to learn what a system does well, where it fails and under which conditions.", "definition": "An evaluation defines tasks and ways to judge the results. A benchmark offers shared tasks or measures for comparison. Different tests may examine accuracy, consistency, harmful outputs or resource use.", "placement": "When reading a score, ask which tasks, models, instructions and measures were compared.", "distinction": "A good result covers the tested conditions. It does not settle performance on every task or certify the whole application as safe.", "references": { "summary": [ "helm-paper" ], "definition": [ "helm-paper" ], "placement": [ "helm-paper" ], "distinction": [ "helm-paper", "nist-genai-profile" ] }, "sources": [ "helm-paper", "nist-genai-profile" ] }, { "id": "alignment", "short": "AI alignment", "term": "AI alignment", "category": "Research goal", "guide": "crosscutting", "group": "AI concepts", "summary": "Work on making AI behavior fit intended goals, constraints and human judgments.", "definition": "Some alignment work trains models to follow instructions or avoid harmful responses, using human feedback or written principles. Broader research also asks whether a learned system pursues the objective its developers intended.", "placement": "Ask whose goals are being followed, how conflicts are handled and what evidence supports the claim.", "distinction": "Alignment is one part of AI safety. Following a user's wishes can still cause harm, and better scores on an alignment test do not settle every risk.", "references": { "summary": [ "instructgpt-paper", "constitutional-ai-paper" ], "definition": [ "instructgpt-paper", "constitutional-ai-paper", "learned-optimization" ], "placement": [ "instructgpt-paper" ], "distinction": [ "instructgpt-paper", "concrete-safety" ] }, "sources": [ "instructgpt-paper", "constitutional-ai-paper", "learned-optimization", "concrete-safety" ] }, { "id": "rlhf", "short": "RLHF", "term": "Reinforcement learning from human feedback (RLHF)", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "Training that uses people's judgments to reward preferred model behavior.", "definition": "In the InstructGPT approach, people compare candidate answers. Their choices train a separate reward model, which scores responses. Further training encourages the language model to produce higher-scoring answers.", "placement": "Ask who provided feedback, what instructions they received and which tasks they judged.", "distinction": "A preferred answer can still be wrong. The people giving feedback also cannot represent every user's values and needs.", "references": { "summary": [ "instructgpt-paper" ], "definition": [ "instructgpt-paper" ], "placement": [ "instructgpt-paper" ], "distinction": [ "instructgpt-paper" ] }, "sources": [ "instructgpt-paper" ] }, { "id": "reasoning-models", "short": "Reasoning models", "term": "Reasoning models & chain of thought", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "Models trained or configured to work through intermediate steps before giving a final answer.", "definition": "These systems can spend additional computation generating steps, checking attempts or exploring alternatives. DeepSeek-R1 is one researched example of using reinforcement learning to encourage such behavior. A written sequence of intermediate steps is often called a chain of thought.", "placement": "Check results on relevant tasks and how much time or computation those results required.", "distinction": "A readable reasoning trace can help with monitoring, but it is an incomplete account of the model's internal calculations. More steps do not by themselves prove the answer correct.", "references": { "summary": [ "deepseek-r1-paper" ], "definition": [ "deepseek-r1-paper" ], "placement": [ "deepseek-r1-paper", "helm-paper" ], "distinction": [ "cot-monitorability", "nist-genai-profile" ] }, "sources": [ "deepseek-r1-paper", "helm-paper", "cot-monitorability", "nist-genai-profile" ] }, { "id": "open-weights", "short": "Open weights", "term": "Open weights & open-source AI", "category": "Access & release", "guide": "crosscutting", "group": "AI concepts", "summary": "Releasing model weights gives access to its learned settings; fuller openness also concerns code, training information and usage rights.", "definition": "A weights release makes learned parameters available. Under the Open Source Initiative's AI definition, openness also requires training and inference code, sufficient training-data information, and rights to use, study, modify and share the system.", "placement": "When a model is called 'open', check exactly which materials and permissions are available.", "distinction": "A download alone does not establish that OSI's definition is met. Check the release terms and accompanying materials.", "references": { "summary": [ "osi-ai-definition" ], "definition": [ "osi-ai-definition" ], "placement": [ "osi-ai-definition" ], "distinction": [ "osi-ai-definition" ] }, "sources": [ "osi-ai-definition" ] }, { "id": "bias", "short": "AI bias", "term": "AI bias & fairness", "category": "Evaluation & ethics", "guide": "crosscutting", "group": "Ethics", "summary": "Patterns in an AI system that can produce uneven or unfair outcomes for people.", "definition": "Bias can come from data, statistical methods, human judgments or institutions around an application. NIST emphasizes how these sources interact. An evaluation needs to examine the actual task and who may be affected.", "placement": "Ask which groups and situations were tested, which measure was used and whose experience is missing.", "distinction": "In mathematics, 'bias' can also mean an added parameter or a statistical error. That use is separate from a finding of unfair treatment.", "references": { "summary": [ "nist-ai-bias" ], "definition": [ "nist-ai-bias" ], "placement": [ "nist-ai-bias" ], "distinction": [ "nist-ai-bias", "google-neural-layers" ] }, "sources": [ "nist-ai-bias", "google-neural-layers" ] }, { "id": "ai-governance", "short": "AI governance", "term": "AI governance", "category": "Policy & oversight", "guide": "crosscutting", "group": "Ethics", "summary": "The rules, responsibilities and oversight that shape how AI is developed and used.", "definition": "Governance includes choices about who makes decisions, who answers for harm and how people can challenge an outcome. The OECD's principles address transparency, human rights, safety and accountability alongside public policy.", "placement": "These questions connect AI policy debates to decisions made by governments and organizations.", "distinction": "A published principle is a commitment or recommendation. Whether it is followed needs evidence about the actual decisions, controls and outcomes.", "references": { "summary": [ "oecd-ai-principles" ], "definition": [ "oecd-ai-principles" ], "placement": [ "oecd-ai-principles" ], "distinction": [ "oecd-ai-principles" ] }, "sources": [ "oecd-ai-principles" ] }, { "id": "chatbot", "short": "Chatbot", "term": "Chatbots & AI assistants", "category": "Application", "guide": "crosscutting", "group": "AI concepts", "summary": "An application you interact with through conversation; an LLM can provide its text generation.", "definition": "In an LLM-powered chatbot, the chat interface sends messages to a model and displays its replies. An assistant may also have search, saved context or connected tools. Those features belong to the application around the model.", "placement": "To assess a chatbot, inspect its actual access and behavior rather than assigning it a political position.", "distinction": "A model, the app using it and the company providing it are different things. A chat interface alone does not tell you which tools it can use.", "references": { "summary": [ "google-ml-glossary", "gemini-overview" ], "definition": [ "gemini-overview", "hf-tool-use", "chatgpt-memory" ], "placement": [ "ncsc-secure-ai" ], "distinction": [ "hf-models", "hf-tool-use" ] }, "sources": [ "google-ml-glossary", "gemini-overview", "hf-tool-use", "chatgpt-memory", "ncsc-secure-ai", "hf-models" ] }, { "id": "chatgpt", "short": "ChatGPT", "term": "ChatGPT", "category": "Product", "guide": "crosscutting", "group": "AI concepts", "summary": "OpenAI's conversational AI product, introduced in November 2022.", "definition": "ChatGPT is the service people interact with. The models and features behind that service can change. Its launch announcement describes an early system trained for dialogue; that announcement is not a specification of today's product.", "placement": "For the company's reviewed public position, see OpenAI. Using a product does not establish a user's views.", "distinction": "ChatGPT, a particular GPT model and OpenAI are separate names for a product, a model and its provider.", "references": { "summary": [ "development-chatgpt" ], "definition": [ "development-chatgpt", "chatgpt-memory" ], "placement": [ "development-chatgpt", "openai-pacing" ], "distinction": [ "development-chatgpt" ] }, "sources": [ "development-chatgpt", "chatgpt-memory", "openai-pacing" ] }, { "id": "claude", "short": "Claude", "term": "Claude", "category": "Product & model family", "guide": "crosscutting", "group": "AI concepts", "summary": "Anthropic's name for its AI assistant and related models.", "definition": "Anthropic introduced Claude with both a chat interface and an API, which lets other software use it. Different versions are models within the family; the surrounding application determines how they are used.", "placement": "Read Anthropic's company record separately from the explanation of this product.", "distinction": "A provider's claims about helpfulness or safety are not an independent assessment of every answer or deployment.", "references": { "summary": [ "claude-introduction" ], "definition": [ "claude-introduction" ], "placement": [ "claude-introduction" ], "distinction": [ "claude-introduction", "ncsc-secure-ai" ] }, "sources": [ "claude-introduction", "ncsc-secure-ai" ] }, { "id": "gemini", "short": "Gemini", "term": "Gemini", "category": "Product & model family", "guide": "crosscutting", "group": "AI concepts", "summary": "Google uses Gemini for an AI assistant and the models it gives access to.", "definition": "Google describes the Gemini app as an interface to its multimodal language models. Multimodal means working with more than text, such as images or audio; supported features depend on the model and application.", "placement": "The product name is context for reading the separate Google DeepMind record.", "distinction": "The app and an underlying model are different parts of the system. A feature name does not certify its answers.", "references": { "summary": [ "gemini-overview" ], "definition": [ "gemini-overview", "hf-multimodal" ], "placement": [ "gemini-overview", "deepmind-framework" ], "distinction": [ "gemini-overview", "nist-genai-profile" ] }, "sources": [ "gemini-overview", "hf-multimodal", "deepmind-framework", "nist-genai-profile" ] }, { "id": "grok", "short": "Grok", "term": "Grok", "category": "Product & model family", "guide": "crosscutting", "group": "AI concepts", "summary": "An AI assistant and model family introduced by xAI.", "definition": "The original announcement distinguishes the Grok assistant from its underlying Grok-1 language model. It also describes search access and acknowledges that answers can still be false or contradictory.", "placement": "Read the separate xAI organization record and Musk's personal record; neither is a score for the product.", "distinction": "Access to recent information does not establish that every generated claim is accurate. Launch-era specifications should not be read as current ones.", "references": { "summary": [ "grok-introduction" ], "definition": [ "grok-introduction" ], "placement": [ "grok-introduction", "musk-dwarkesh-2026" ], "distinction": [ "grok-introduction", "nist-genai-profile" ] }, "sources": [ "grok-introduction", "musk-dwarkesh-2026", "nist-genai-profile" ] }, { "id": "llama", "short": "Llama", "term": "Llama", "category": "Model family", "guide": "crosscutting", "group": "AI concepts", "summary": "A family of AI models released by Meta, which developers can use in applications.", "definition": "Meta's Llama 3 paper describes a family of language models with versions before and after additional training for use. A model family is not one particular chat application.", "placement": "Read Meta's organization record separately from a model's technical description.", "distinction": "Model availability and permission to use it are separate questions. Check the terms for the exact release; open weights and open-source AI are not interchangeable labels.", "references": { "summary": [ "llama-model-family" ], "definition": [ "llama-model-family" ], "placement": [ "llama-model-family", "meta-framework" ], "distinction": [ "osi-ai-definition" ] }, "sources": [ "llama-model-family", "meta-framework", "osi-ai-definition" ] }, { "id": "synthetic-content", "short": "Synthetic content", "term": "Synthetic content / AI-generated media", "category": "Media concept", "guide": "crosscutting", "group": "AI concepts", "summary": "Text, images, audio or video created or substantially altered using algorithms, including AI.", "definition": "This includes fully generated material and edits to existing material. NIST examines ways to record its origin and changes.", "placement": "A question about media and its use, beyond the chart's two axes.", "distinction": "Synthetic does not automatically mean harmful. Authentic material can also mislead when presented in the wrong context.", "references": { "summary": [ "nist-synthetic-content" ], "definition": [ "nist-synthetic-content" ], "placement": [ "nist-synthetic-content" ], "distinction": [ "nist-synthetic-content" ] }, "sources": [ "nist-synthetic-content" ] }, { "id": "deepfake", "short": "Deepfake", "term": "Deepfakes & voice cloning", "category": "Media concept", "guide": "crosscutting", "group": "AI concepts", "summary": "Generated or altered media that imitates a person's appearance or voice.", "definition": "A fabricated recording can make someone appear to say or do something they did not. Uses can include impersonation and fraud.", "placement": "This concerns misuse of media; it does not locate someone on the map.", "distinction": "A detection tool can make mistakes. Check the original source and context before trusting or sharing a recording.", "references": { "summary": [ "nist-synthetic-content" ], "definition": [ "nist-synthetic-content" ], "placement": [ "nist-synthetic-content" ], "distinction": [ "nist-synthetic-content" ] }, "sources": [ "nist-synthetic-content" ] }, { "id": "ai-slop", "short": "AI slop", "term": "AI slop", "category": "Critical label", "guide": "crosscutting", "group": "Ethics", "summary": "A dismissive label for low-quality AI-generated content, often produced in large amounts.", "definition": "Merriam-Webster's 2025 selection describes the term through examples such as poor-quality videos, images and writing. The word expresses a judgment about the content's quality.", "placement": "Criticizing unwanted content does not by itself state a position on AI catastrophe or frontier development.", "distinction": "Use the label carefully: it is not a technical test for whether a particular piece was made with AI, or evidence that all AI-assisted work is poor.", "references": { "summary": [ "merriam-webster-slop" ], "definition": [ "merriam-webster-slop" ], "placement": [ "merriam-webster-slop" ], "distinction": [ "merriam-webster-slop" ] }, "sources": [ "merriam-webster-slop" ] }, { "id": "eu-ai-act", "short": "EU AI Act", "term": "European Union AI Act", "category": "Law", "guide": "crosscutting", "group": "Ethics", "summary": "EU rules for AI systems and general-purpose models, with requirements linked to their risks and uses.", "definition": "The Act addresses prohibited practices, high-risk uses, transparency and general-purpose models. The European Commission explains the categories, enforcement and phased application on its official site.", "placement": "Regulating a use, such as hiring, is different from calling for a general pause in model development.", "distinction": "Duties and application dates depend on the system, role and relevant rules. This glossary is an introduction, not a compliance determination.", "references": { "summary": [ "eu-ai-act-overview" ], "definition": [ "eu-ai-act-overview" ], "placement": [ "eu-ai-act-overview", "pause-letter-2023" ], "distinction": [ "eu-ai-act-overview" ] }, "sources": [ "eu-ai-act-overview", "pause-letter-2023" ] }, { "id": "ai-race", "short": "AI race", "term": "The AI race", "category": "Political framing", "guide": "crosscutting", "group": "Movements", "summary": "A way of framing AI development as competition for technological, economic or strategic advantage.", "definition": "The White House's 2025 AI Action Plan uses race language to argue for US leadership through innovation, infrastructure and international policy. It is one explicit example of this framing.", "placement": "Race arguments can support faster development, but do not specify a speaker's view of catastrophic risk.", "distinction": "A political argument for winning does not establish that there is one finish line, that benefits are guaranteed, or that every proposed action has happened.", "references": { "summary": [ "white-house-ai-action-plan" ], "definition": [ "white-house-ai-action-plan" ], "placement": [ "white-house-ai-action-plan" ], "distinction": [ "white-house-ai-action-plan" ] }, "sources": [ "white-house-ai-action-plan" ] }, { "id": "automation", "short": "AI and jobs", "term": "Automation, job tasks & exposure", "category": "Work & society", "guide": "crosscutting", "group": "Ethics", "summary": "Software can take over parts of a job; a task being automatable does not mean the whole job disappears.", "definition": "The ILO's 2025 analysis estimates which job tasks could be affected by generative AI. Its exposure categories describe potential changes, with many occupations combining affected tasks and tasks requiring human input.", "placement": "Workplace effects are another dimension of the debate, beyond catastrophic risk and the pace of frontier research.", "distinction": "Exposure estimates are not observed job losses. Costs, infrastructure, skills and decisions about adoption affect the outcome.", "references": { "summary": [ "ilo-ai-work-exposure" ], "definition": [ "ilo-ai-work-exposure" ], "placement": [ "ilo-ai-work-exposure" ], "distinction": [ "ilo-ai-work-exposure" ] }, "sources": [ "ilo-ai-work-exposure" ] }, { "id": "ai-consciousness", "short": "AI consciousness", "term": "AI consciousness & sentience", "category": "Open research question", "guide": "crosscutting", "group": "Ethics", "summary": "Whether an AI could have subjective experience, such as feeling something, rather than only describing it.", "definition": "Researchers disagree about how consciousness should be assessed. One proposed approach examines internal features suggested by scientific theories, while acknowledging disputed assumptions and uncertain conclusions.", "placement": "Consciousness, capability and risk are separate questions; one answer does not settle the others.", "distinction": "Human-like speech is not a sufficient test. 'Sentience' is used differently across discussions, so check whether the speaker means experience, sensation or something else.", "references": { "summary": [ "ai-consciousness-indicators" ], "definition": [ "ai-consciousness-indicators" ], "placement": [ "ai-consciousness-indicators" ], "distinction": [ "ai-consciousness-indicators" ] }, "sources": [ "ai-consciousness-indicators" ] }, { "id": "intelligence", "short": "Intelligence", "term": "Intelligence", "category": "Contested concept", "guide": "crosscutting", "summary": "A broad word for abilities such as learning, using concepts and solving new problems. Different definitions emphasize different abilities.", "definition": "The 1955 Dartmouth proposal used artificial intelligence for a research project on machine language, concepts, problem-solving and improvement. The label predates modern LLMs; using it does not establish human-like thought.", "placement": "A definition of intelligence does not specify a development speed or level of catastrophic-risk concern.", "distinction": "Open question: which abilities and tests would justify the label? Successful task performance, human-like understanding and consciousness need separate evidence.", "group": "AI concepts", "references": { "summary": [ "development-dartmouth-proposal", "understanding-mitchell-krakauer" ], "definition": [ "development-dartmouth-proposal", "understanding-mitchell-krakauer" ], "placement": [ "development-dartmouth-proposal", "understanding-mitchell-krakauer" ], "distinction": [ "understanding-mitchell-krakauer" ] }, "sources": [ "development-dartmouth-proposal", "understanding-mitchell-krakauer" ] }, { "id": "understanding", "short": "Understanding", "term": "Understanding in language models", "category": "Research debate", "guide": "crosscutting", "summary": "The dispute over whether producing suitable language also involves grasping what that language means.", "definition": "Bender and Koller argue that training only on patterns in language cannot teach the connection between words and what speakers mean. In a separate Othello board-game study, a model predicting moves learned information about the board's state. That finding concerns an internal representation, not proof of human-like understanding.", "placement": "This debate concerns what models do, not membership of a camp on the map.", "distinction": "Open question: what would distinguish robust understanding from a successful shortcut? Test unfamiliar situations, changed assumptions and failures, rather than judging fluency alone.", "group": "AI concepts", "references": { "summary": [ "understanding-bender-koller", "understanding-mitchell-krakauer" ], "definition": [ "understanding-bender-koller", "understanding-othello" ], "placement": [ "understanding-mitchell-krakauer" ], "distinction": [ "understanding-mitchell-krakauer" ] }, "sources": [ "understanding-bender-koller", "understanding-othello", "understanding-mitchell-krakauer" ] }, { "id": "grounding", "short": "Grounding", "term": "Symbol grounding", "category": "Research concept", "guide": "crosscutting", "summary": "Connecting words or symbols to what they refer to, rather than only to other symbols.", "definition": "The question is how language connects to objects, events and what a speaker is trying to communicate. Bender and Koller's argument concerns systems trained only on language form. Adding images or interaction changes the evidence available; it does not by itself prove human-like understanding.", "placement": "Grounding is a question about representations and learning, not an ideological position.", "distinction": "Open question: which connections are needed, and what would show they are enough? Finding a document to support an answer is different from establishing that a system understands its meaning.", "group": "AI concepts", "references": { "summary": [ "understanding-bender-koller" ], "definition": [ "understanding-bender-koller", "understanding-mitchell-krakauer" ], "placement": [ "understanding-bender-koller" ], "distinction": [ "understanding-bender-koller", "rag-paper" ] }, "sources": [ "understanding-bender-koller", "understanding-mitchell-krakauer", "rag-paper" ] }, { "id": "anthropomorphism", "short": "Anthropomorphism", "term": "Anthropomorphism in AI", "category": "Design and language", "guide": "crosscutting", "summary": "Attributing human qualities, such as feelings or intentions, to a machine.", "definition": "A human-like voice or statements about having feelings can make a system seem like a person. The cited researchers examine how these design choices could encourage over-reliance or blur responsibility. Effects and useful safeguards depend on context.", "placement": "Questioning human-like language does not by itself imply a position on either map axis.", "distinction": "Open questions: which labels help people predict behavior, and which mislead? Claims that every provider uses AI terminology to avoid accountability require evidence about those providers.", "group": "AI concepts", "references": { "summary": [ "understanding-anthropomorphism" ], "definition": [ "understanding-anthropomorphism" ], "placement": [ "understanding-anthropomorphism" ], "distinction": [ "understanding-anthropomorphism" ] }, "sources": [ "understanding-anthropomorphism" ] }, { "id": "algorithm", "short": "Algorithms", "term": "Algorithm", "category": "AI basics", "guide": "crosscutting", "group": "AI concepts", "summary": "A set of steps a computer can carry out to produce a result.", "definition": "For example, a sorting algorithm puts a list of numbers in order. Training a language model uses algorithms to adjust its numerical settings.", "placement": "Start here when separating the program’s procedure from the model it trains or runs.", "distinction": "An algorithm can include random choices. Calling something an algorithm does not establish that every run gives the same result.", "references": { "summary": [ "glossary-model-algorithm" ], "definition": [ "glossary-model-algorithm", "google-gradient-descent", "hf-models" ], "placement": [ "glossary-model-algorithm", "hf-models" ], "distinction": [ "glossary-model-algorithm" ] }, "sources": [ "glossary-model-algorithm", "google-gradient-descent", "hf-models" ] }, { "id": "training-data", "short": "Training data", "term": "Training data & datasets", "category": "AI basics", "guide": "crosscutting", "group": "AI concepts", "summary": "Examples used to adjust a model. Their choice affects what the model learns.", "definition": "A dataset is a collection of examples, such as text or images. The training portion is used to fit the model; separate validation and test portions help assess its performance.", "placement": "Ask where the examples came from and whether the evaluation uses genuinely separate material.", "distinction": "A large dataset can still miss relevant situations. Testing on duplicated training examples can make a model look better than it is on new data.", "references": { "summary": [ "glossary-model-supervised" ], "definition": [ "glossary-model-supervised", "glossary-model-datasets" ], "placement": [ "glossary-model-datasets" ], "distinction": [ "glossary-model-supervised", "glossary-model-datasets" ] }, "sources": [ "glossary-model-supervised", "glossary-model-datasets" ] }, { "id": "pretraining", "short": "Pretraining", "term": "Pretraining", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "An initial training stage that provides a starting model for later use or adaptation.", "definition": "For a language model, this often means adjusting weights on large text collections to predict missing or next text pieces. Later training can change how the model handles specific tasks.", "placement": "Use this distinction when a proposal refers specifically to training new base models.", "distinction": "A pretrained language model is not automatically an instruction-following assistant. Pretraining and later adaptation serve different objectives.", "references": { "summary": [ "glossary-model-pretraining" ], "definition": [ "glossary-model-pretraining" ], "placement": [ "glossary-model-pretraining" ], "distinction": [ "glossary-model-posttraining" ] }, "sources": [ "glossary-model-pretraining", "glossary-model-posttraining" ] }, { "id": "post-training", "short": "Post-training", "term": "Post-training", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "Further training after pretraining, intended to adapt a model’s abilities and behavior.", "definition": "Developers can train on demonstrations, preferred answers or rewards. These methods may target instruction following, coding or responses to harmful requests. Different developers use different combinations.", "placement": "Inspect the training goal, whose feedback was used and which results were tested.", "distinction": "Post-training names a development stage, not a safety certificate. Improvements on selected tasks leave other failures possible.", "references": { "summary": [ "glossary-model-posttraining" ], "definition": [ "glossary-model-posttraining", "instructgpt-paper" ], "placement": [ "instructgpt-paper" ], "distinction": [ "glossary-model-posttraining" ] }, "sources": [ "glossary-model-posttraining", "instructgpt-paper" ] }, { "id": "supervised-learning", "short": "Supervised learning", "term": "Supervised learning", "category": "AI basics", "guide": "crosscutting", "group": "AI concepts", "summary": "Training with examples that include a target answer, such as a category or number.", "definition": "Software compares the model’s prediction with the target and uses the difference to adjust it. An illustrative task is sorting messages using examples already labeled as spam or ordinary mail.", "placement": "Ask how the labels were made and whether they match the task you want to perform.", "distinction": "Supervised describes the training examples. It does not mean a person checks every answer when the trained model is used.", "references": { "summary": [ "glossary-model-supervised" ], "definition": [ "glossary-model-supervised" ], "placement": [ "glossary-model-supervised" ], "distinction": [ "glossary-model-supervised" ] }, "sources": [ "glossary-model-supervised" ] }, { "id": "unsupervised-learning", "short": "Unsupervised learning", "term": "Unsupervised learning", "category": "AI basics", "guide": "crosscutting", "group": "AI concepts", "summary": "Methods that find patterns in data without a supplied answer label for each example.", "definition": "Clustering is one example: it groups items by a chosen measure of similarity. Imagine grouping songs by their sound without giving the program genre labels first.", "placement": "Treat discovered groups as results to inspect, not as automatic ideology memberships.", "distinction": "Unsupervised does not mean free of human choices. The selected data and definition of similarity shape the groups.", "references": { "summary": [ "google-ml-intro" ], "definition": [ "glossary-model-clustering" ], "placement": [ "glossary-model-clustering" ], "distinction": [ "glossary-model-clustering" ] }, "sources": [ "google-ml-intro", "glossary-model-clustering" ] }, { "id": "self-supervised-learning", "short": "Self-supervised learning", "term": "Self-supervised learning", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "Training that builds its target answers from the data itself, such as a hidden word.", "definition": "Software can hide part of a text, ask the model to predict it and compare the prediction with the original. This supplies a training signal without someone labeling each example separately.", "placement": "Read this alongside pretraining to see where a language model’s practice targets come from.", "distinction": "The original text supplies a target, not independent proof that its claims are true. Self-supervised does not mean self-verifying.", "references": { "summary": [ "glossary-model-self-supervised" ], "definition": [ "glossary-model-self-supervised" ], "placement": [ "glossary-model-self-supervised" ], "distinction": [ "glossary-model-self-supervised", "understanding-bender-koller" ] }, "sources": [ "glossary-model-self-supervised", "understanding-bender-koller" ] }, { "id": "reinforcement-learning", "short": "Reinforcement learning", "term": "Reinforcement learning", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "Training that uses rewards from attempted actions to adjust which actions a system selects.", "definition": "A system tries actions in an environment and receives numerical feedback. Training aims for more reward over time. A game score can supply feedback; other tasks need other reward rules.", "placement": "Examine what earns a reward before interpreting claims that training improved behavior.", "distinction": "A higher reward can miss the outcome people intended. RLHF is a specific approach using human feedback, not a name for all reinforcement learning.", "references": { "summary": [ "glossary-model-reinforcement" ], "definition": [ "glossary-model-reinforcement" ], "placement": [ "concrete-safety" ], "distinction": [ "concrete-safety", "instructgpt-paper" ] }, "sources": [ "glossary-model-reinforcement", "concrete-safety", "instructgpt-paper" ] }, { "id": "deep-learning", "short": "Deep learning", "term": "Deep learning", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "Machine learning that builds several layers of learned calculations on top of one another.", "definition": "In a deep neural network, intermediate layers transform the numbers passed between input and output. Training adjusts settings across the network. These layers can build increasingly complex representations.", "placement": "Use this term to distinguish a model family from a claim about how quickly AI should advance.", "distinction": "Deep refers to layers of computation or representation. It is not a measure of wisdom, and there is no universally agreed minimum depth.", "references": { "summary": [ "glossary-model-deep-learning" ], "definition": [ "google-neural-layers", "glossary-model-deep-learning" ], "placement": [ "glossary-model-deep-learning" ], "distinction": [ "glossary-model-deep-learning" ] }, "sources": [ "glossary-model-deep-learning", "google-neural-layers" ] }, { "id": "attention", "short": "Attention", "term": "Attention & self-attention", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "A calculation that mixes information from different input positions with different weights.", "definition": "The weights depend on the input. In self-attention, parts of one sequence supply the information being combined. Multiple attention heads perform different learned combinations in parallel.", "placement": "Use it to read Transformer diagrams. It describes a component, not an actor’s position.", "distinction": "Attention is one operation within a larger network. The name does not imply a separate reader or human-like concentration.", "references": { "summary": [ "transformer-paper" ], "definition": [ "transformer-paper" ], "placement": [ "transformer-paper" ], "distinction": [ "transformer-paper" ] }, "sources": [ "transformer-paper" ] }, { "id": "model-architecture", "short": "Architecture", "term": "Model architecture", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "The arrangement of a model’s parts and the calculations that connect them.", "definition": "An architecture specifies the structure, such as the number and kind of layers. A checkpoint supplies learned weights for that structure. Two models can share an architecture but have different weights.", "placement": "Check whether a comparison concerns the structure, the trained weights or the surrounding application.", "distinction": "An architecture diagram does not specify all learned values. Rebuilding the structure alone does not reproduce a trained model.", "references": { "summary": [ "hf-models" ], "definition": [ "hf-models" ], "placement": [ "hf-models" ], "distinction": [ "hf-models" ] }, "sources": [ "hf-models" ] }, { "id": "mixture-of-experts", "short": "Mixture of experts", "term": "Mixture of experts (MoE)", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "A model design with several calculation blocks called experts; some versions activate only a subset for each input.", "definition": "In sparse MoE models such as Mixtral, a routing component selects only some expert blocks for each text piece. This allows more total parameters than are active for that piece.", "placement": "Distinguish total from active parameter counts when reading model-size comparisons.", "distinction": "Experts are numerical components, not independent professionals. The name alone does not show that each block has a clear subject specialty.", "references": { "summary": [ "glossary-model-mixture-of-experts" ], "definition": [ "glossary-model-mixture-of-experts" ], "placement": [ "glossary-model-mixture-of-experts" ], "distinction": [ "glossary-model-mixture-of-experts" ] }, "sources": [ "glossary-model-mixture-of-experts" ] }, { "id": "distillation", "short": "Distillation", "term": "Model distillation", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "Training a model to match outputs from another model, often to make it cheaper to run.", "definition": "In the cited method, a larger model supplies probabilities as training targets for a smaller one. The target is output behavior, rather than copying the original model’s weights.", "placement": "Ask which tasks the smaller model was tested on and how its results changed.", "distinction": "The smaller model may not match every output or capability. Distillation does not guarantee identical performance.", "references": { "summary": [ "glossary-model-distillation" ], "definition": [ "glossary-model-distillation" ], "placement": [ "glossary-model-distillation" ], "distinction": [ "glossary-model-distillation" ] }, "sources": [ "glossary-model-distillation" ] }, { "id": "quantization", "short": "Quantization", "term": "Quantization", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "Representing model numbers with fewer bits to reduce memory use, with possible accuracy tradeoffs.", "definition": "Weights or intermediate values are mapped to a smaller set of representable numbers. For example, an eight-bit representation uses less storage per value than a 32-bit representation, but introduces approximation.", "placement": "Compare memory needs, task results and performance on the hardware actually being used.", "distinction": "Quantization changes numerical precision. It does not necessarily reduce the number of parameters, and faster execution depends on the implementation and hardware.", "references": { "summary": [ "glossary-model-quantization" ], "definition": [ "glossary-model-quantization" ], "placement": [ "glossary-model-quantization" ], "distinction": [ "glossary-model-quantization" ] }, "sources": [ "glossary-model-quantization" ] }, { "id": "overfitting", "short": "Overfitting", "term": "Overfitting", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "When a model fits its training examples so closely that it performs worse on new examples.", "definition": "Training can fit details that do not carry over to new data. A warning sign is training error falling while error on separate validation data rises.", "placement": "Look for results on separate, relevant examples when assessing claims about model quality.", "distinction": "A good training score alone does not show useful performance elsewhere. Underfitting is different: the model already struggles with its training examples.", "references": { "summary": [ "glossary-model-overfitting" ], "definition": [ "glossary-model-overfitting" ], "placement": [ "glossary-model-overfitting" ], "distinction": [ "glossary-model-overfitting" ] }, "sources": [ "glossary-model-overfitting" ] }, { "id": "underfitting", "short": "Underfitting", "term": "Underfitting", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "When a model has not captured enough of the pattern even in its training examples.", "definition": "Possible causes include an unsuitable model structure, missing useful inputs or too little training. These are different problems, so simply adding more examples may not address the cause.", "placement": "Ask whether the error already appears in training before comparing performance on new data.", "distinction": "Underfitting is not a synonym for a small model. It describes an inadequate fit for the task and available data.", "references": { "summary": [ "glossary-model-underfitting", "glossary-model-overfitting" ], "definition": [ "glossary-model-underfitting" ], "placement": [ "glossary-model-overfitting" ], "distinction": [ "glossary-model-underfitting" ] }, "sources": [ "glossary-model-underfitting", "glossary-model-overfitting" ] }, { "id": "loss-function", "short": "Loss function", "term": "Loss function", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "A rule that turns a model’s training errors into a number to reduce.", "definition": "Different loss functions count errors differently. For example, squaring numerical prediction errors gives large mistakes more weight than taking their absolute size.", "placement": "Ask what the training score measures and which real-world mistakes it leaves out.", "distinction": "Lower loss means improvement under that scoring rule. It is not by itself proof of safe or useful behavior.", "references": { "summary": [ "glossary-model-loss" ], "definition": [ "glossary-model-loss" ], "placement": [ "glossary-model-loss", "concrete-safety" ], "distinction": [ "glossary-model-loss", "concrete-safety" ] }, "sources": [ "glossary-model-loss", "concrete-safety" ] }, { "id": "gradient-descent", "short": "Gradient descent", "term": "Gradient descent", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "summary": "A way to adjust model settings step by step in a direction that aims to reduce training error.", "definition": "Software calculates how a small change to each parameter would affect the loss, then updates the parameters in the opposite direction. Repeating these steps is part of many training procedures.", "placement": "This explains what changing weights during training means in concrete computational terms.", "distinction": "Reducing error on the training objective does not guarantee good results on new examples. Training and evaluation answer different questions.", "references": { "summary": [ "google-gradient-descent" ], "definition": [ "google-gradient-descent" ], "placement": [ "google-gradient-descent" ], "distinction": [ "glossary-model-datasets" ] }, "sources": [ "google-gradient-descent", "glossary-model-datasets" ] }, { "id": "benchmark", "guide": "crosscutting", "group": "AI concepts", "short": "Benchmark", "term": "Benchmark", "category": "Evaluation method", "summary": "A shared test for comparing systems on chosen tasks; its score covers those tests, not every possible use.", "definition": "A benchmark combines tasks, scoring rules and a test setup. Using the same setup makes comparisons more informative. Accuracy, cost, bias and robustness can produce different pictures of the same model.", "placement": "Atlas reading question: which tasks and conditions support a public capability claim?", "distinction": "Evaluation is the broader investigation; a benchmark is one tool within it. A high score can leave relevant languages, users or failure types untested.", "references": { "summary": [ "helm-paper" ], "definition": [ "helm-paper" ], "placement": [ "helm-paper" ], "distinction": [ "helm-paper" ] }, "sources": [ "helm-paper" ] }, { "id": "data-contamination", "guide": "crosscutting", "group": "AI concepts", "short": "Test contamination", "term": "Data contamination / benchmark contamination", "category": "Evaluation problem", "summary": "Test material appearing in training data can make an apparently new test partly familiar to a model.", "definition": "Overlap between training material and test questions or answers can weaken a test of performance on unseen material. Checking that overlap helps interpret a reported score.", "placement": "Atlas reading question: how did an evaluator check that the test was meaningfully separate from training?", "distinction": "Overlap does not prove that a model memorized an answer or gained an advantage. Brown and colleagues found that effects varied, and their detection method had limits.", "references": { "summary": [ "glossary-eval-contamination" ], "definition": [ "glossary-eval-contamination" ], "placement": [ "glossary-eval-contamination" ], "distinction": [ "glossary-eval-gpt3" ] }, "sources": [ "glossary-eval-contamination", "glossary-eval-gpt3" ] }, { "id": "generalization", "guide": "crosscutting", "group": "AI concepts", "short": "Generalization", "term": "Generalization", "category": "Model behavior", "summary": "Doing useful work on examples outside the training set; success depends on how different those examples are.", "definition": "A model generalizes when patterns learned during training support good results on new examples. Recognizing a new photo of a familiar kind of object is an illustrative case.", "placement": "Atlas reading question: does a claimed improvement hold beyond the examples used to develop the system?", "distinction": "A new example can still closely resemble training data. Success there does not establish success in a different setting, and generalization is not a declaration of AGI.", "references": { "summary": [ "google-ml-glossary", "glossary-eval-rmf-characteristics" ], "definition": [ "google-ml-glossary" ], "placement": [ "glossary-eval-rmf-characteristics" ], "distinction": [ "glossary-eval-rmf-characteristics", "helm-paper" ] }, "sources": [ "google-ml-glossary", "glossary-eval-rmf-characteristics", "helm-paper" ] }, { "id": "calibration", "guide": "crosscutting", "group": "AI concepts", "short": "Calibration", "term": "Confidence calibration", "category": "Evaluation property", "summary": "Checking whether a system's stated probabilities match how often its predictions turn out right.", "definition": "Illustrative example: among many predictions assigned an 80% chance of being correct, about 80% should be correct. Calibration concerns that match across cases, not certainty about one answer.", "placement": "Atlas reading question: has a confidence number been checked against outcomes?", "distinction": "High accuracy and good calibration are different properties. Confident wording in a chatbot reply is not itself a measured probability.", "references": { "summary": [ "glossary-eval-calibration" ], "definition": [ "glossary-eval-calibration" ], "placement": [ "glossary-eval-calibration" ], "distinction": [ "glossary-eval-calibration" ] }, "sources": [ "glossary-eval-calibration" ] }, { "id": "uncertainty", "guide": "crosscutting", "group": "AI concepts", "short": "Uncertainty", "term": "Predictive uncertainty", "category": "Evaluation property", "summary": "A way to describe limits on a prediction, such as missing knowledge or noisy information.", "definition": "Researchers distinguish uncertainty due to limited knowledge in a model from uncertainty in the observations themselves. Estimating these separately can help show where more data may help and where observations remain ambiguous.", "placement": "Atlas reading question: what does an uncertainty measure refer to, and how was it checked?", "distinction": "Generating several different answers is not automatically a calibrated uncertainty estimate. The categories describe sources of uncertainty; calibration checks whether numerical estimates fit outcomes.", "references": { "summary": [ "glossary-eval-uncertainty" ], "definition": [ "glossary-eval-uncertainty" ], "placement": [ "glossary-eval-uncertainty", "glossary-eval-calibration" ], "distinction": [ "glossary-eval-uncertainty", "glossary-eval-calibration" ] }, "sources": [ "glossary-eval-uncertainty", "glossary-eval-calibration" ] }, { "id": "robustness", "guide": "crosscutting", "group": "AI concepts", "short": "Robustness", "term": "Robustness", "category": "Evaluation property", "summary": "Maintaining useful performance when conditions change, rather than only succeeding in one test setup.", "definition": "Robustness asks how performance holds up across variations. For example, test a document reader on blurred scans as well as clean ones. The relevant changes depend on the intended use.", "placement": "Atlas reading question: which difficult conditions were included in a claim that a system is reliable?", "distinction": "Robustness is broader than resisting deliberate attacks. A system can handle one kind of change and fail on another; the tested conditions need to be named.", "references": { "summary": [ "glossary-eval-rmf-characteristics" ], "definition": [ "glossary-eval-rmf-characteristics" ], "placement": [ "glossary-eval-rmf-characteristics" ], "distinction": [ "glossary-eval-rmf-characteristics" ] }, "sources": [ "glossary-eval-rmf-characteristics" ] }, { "id": "adversarial-examples", "guide": "crosscutting", "group": "AI concepts", "short": "Adversarial inputs", "term": "Adversarial examples", "category": "Security concept", "summary": "Inputs deliberately changed to make a model give a wrong result; testing them can reveal weaknesses.", "definition": "In a classic image example, small carefully chosen changes cause a classifier to give the wrong label. The term covers crafted inputs, not simply every difficult or unfamiliar example.", "placement": "Atlas reading question: does a security claim cover deliberate manipulation, and under what conditions?", "distinction": "These attacks act on inputs during use. Data poisoning instead alters material used for training. A defense evaluated against one attack need not stop another.", "references": { "summary": [ "glossary-eval-adversarial" ], "definition": [ "glossary-eval-adversarial" ], "placement": [ "glossary-eval-aml-taxonomy" ], "distinction": [ "glossary-eval-aml-taxonomy" ] }, "sources": [ "glossary-eval-adversarial", "glossary-eval-aml-taxonomy" ] }, { "id": "red-teaming", "guide": "crosscutting", "group": "AI concepts", "short": "Red teaming", "term": "Red teaming", "category": "Evaluation method", "summary": "Deliberately probing a system for harmful or unwanted behavior so weaknesses can be investigated and fixed.", "definition": "Testers try challenging cases rather than only routine requests. People can devise the tests, and models can help generate candidates. The cited study used another language model to search for problematic replies.", "placement": "Atlas reading question: who tested the system, what did they try, and what happened to the findings?", "distinction": "Finding a failure does not measure how often it occurs in ordinary use. Finding none does not establish that every harmful behavior has been excluded.", "references": { "summary": [ "glossary-eval-red-teaming" ], "definition": [ "glossary-eval-red-teaming" ], "placement": [ "glossary-eval-red-teaming" ], "distinction": [ "glossary-eval-red-teaming" ] }, "sources": [ "glossary-eval-red-teaming" ] }, { "id": "in-context-learning", "guide": "crosscutting", "group": "AI concepts", "short": "In-context learning", "term": "In-context learning", "category": "Model behavior", "summary": "Using material in the current prompt to adapt an answer, without a new training run.", "definition": "A prompt can provide examples or a pattern that guides the next response. This makes some task adaptation possible without a new training run. Researchers also study how this behavior arises inside transformers.", "placement": "Atlas reading question: did a demonstration change the model itself, or only the information in its input?", "distinction": "This is not a claim that a conversation permanently teaches the underlying model. Accounts of the internal mechanism remain dependent on the model and evidence studied.", "references": { "summary": [ "google-llm-tuning", "glossary-eval-induction" ], "definition": [ "google-llm-tuning", "glossary-eval-induction" ], "placement": [ "google-llm-tuning", "glossary-eval-induction" ], "distinction": [ "glossary-eval-gpt3", "glossary-eval-induction" ] }, "sources": [ "google-llm-tuning", "glossary-eval-induction", "glossary-eval-gpt3" ] }, { "id": "few-shot-learning", "guide": "crosscutting", "group": "AI concepts", "short": "Few-shot prompting", "term": "Few-shot learning / few-shot prompting", "category": "Prompting method", "summary": "Giving a model a few worked examples in the prompt to show the kind of response wanted.", "definition": "In LLM prompting, the examples are part of the input, not a separate training run. For instance, show two sample messages labeled by topic before asking it to label a third.", "placement": "Atlas reading question: were examples supplied when a model's result was reported?", "distinction": "Few examples in the prompt does not mean little prior training. Improvements vary by task and model; more examples do not guarantee a better result.", "references": { "summary": [ "google-llm-tuning" ], "definition": [ "google-llm-tuning" ], "placement": [ "google-llm-tuning" ], "distinction": [ "glossary-eval-gpt3" ] }, "sources": [ "google-llm-tuning", "glossary-eval-gpt3" ] }, { "id": "zero-shot-learning", "guide": "crosscutting", "group": "AI concepts", "short": "Zero-shot prompting", "term": "Zero-shot learning / zero-shot prompting", "category": "Prompting method", "summary": "Asking a model to perform a task without giving worked examples in that prompt.", "definition": "For an LLM, a zero-shot test might give an instruction and a question but no demonstration of a correct answer. This tests what the already-trained model can do under that setup.", "placement": "Atlas reading question: does a comparison use the same amount of help in each model's input?", "distinction": "Zero-shot does not mean untrained, unfamiliar with the topic, or free of test contamination. It describes the immediate task setup.", "references": { "summary": [ "glossary-eval-gpt3" ], "definition": [ "glossary-eval-gpt3" ], "placement": [ "glossary-eval-gpt3" ], "distinction": [ "glossary-eval-gpt3" ] }, "sources": [ "glossary-eval-gpt3" ] }, { "id": "chain-of-thought", "guide": "crosscutting", "group": "AI concepts", "short": "Chain of thought", "term": "Chain of thought (CoT)", "category": "Prompting & generated text", "summary": "Generated intermediate steps before an answer, which can improve some task results but can also contain mistakes.", "definition": "Chain-of-thought prompting supplies examples with intermediate steps so a model produces a similar sequence. The original paper reported gains on selected tasks. Later reasoning systems can also be trained to generate extended intermediate text.", "placement": "Atlas reading question: does a displayed explanation help check the answer, and what does it leave unverified?", "distinction": "An explanation can be useful without faithfully reporting every influence on the answer. The label names generated text, not evidence of a human-like inner thought process.", "references": { "summary": [ "glossary-eval-cot", "understanding-faithfulness" ], "definition": [ "glossary-eval-cot", "understanding-faithfulness" ], "placement": [ "understanding-faithfulness" ], "distinction": [ "glossary-eval-unfaithful-cot", "understanding-faithfulness" ] }, "sources": [ "glossary-eval-cot", "understanding-faithfulness", "glossary-eval-unfaithful-cot" ] }, { "id": "explanation-faithfulness", "guide": "crosscutting", "group": "AI concepts", "short": "Faithful explanations", "term": "Explanation faithfulness", "category": "Interpretability question", "summary": "Whether an explanation accurately reflects what influenced an answer, rather than merely sounding plausible.", "definition": "Researchers can introduce a hint, observe whether it changes an answer, and check whether the explanation mentions it. Such tests examine a particular causal influence, not every step of the underlying computation.", "placement": "Atlas reading question: what evidence connects an explanation to the process that produced the answer?", "distinction": "A correct answer can have an incomplete explanation. The cited hint experiments found omissions in particular models and tasks; they do not show that every reasoning trace is useless.", "references": { "summary": [ "glossary-eval-unfaithful-cot", "understanding-faithfulness" ], "definition": [ "glossary-eval-unfaithful-cot", "understanding-faithfulness" ], "placement": [ "glossary-eval-unfaithful-cot", "understanding-faithfulness" ], "distinction": [ "glossary-eval-unfaithful-cot", "understanding-faithfulness" ] }, "sources": [ "glossary-eval-unfaithful-cot", "understanding-faithfulness" ] }, { "id": "guardrails", "guide": "crosscutting", "group": "AI concepts", "short": "Guardrails", "term": "Guardrails", "category": "System safeguard", "summary": "Checks and restrictions intended to prevent unwanted inputs, outputs or actions in an AI application.", "definition": "The term covers different controls. A concrete implementation may check user input, filter retrieved documents, inspect a reply or validate a tool call before execution. The control's location matters.", "placement": "Atlas reading question: which safeguard is actually enforced, at which step, and against which failure?", "distinction": "A check on the final answer does not itself restrict an earlier tool action. Stating that guardrails exist does not establish their effectiveness against the relevant attacks.", "references": { "summary": [ "glossary-eval-guardrails" ], "definition": [ "glossary-eval-guardrails" ], "placement": [ "glossary-eval-guardrails" ], "distinction": [ "glossary-eval-guardrails", "glossary-eval-aml-taxonomy" ] }, "sources": [ "glossary-eval-guardrails", "glossary-eval-aml-taxonomy" ] }, { "id": "human-in-the-loop", "guide": "crosscutting", "group": "AI concepts", "short": "Human review", "term": "Human in the loop (HITL)", "category": "System design", "summary": "Giving people a role in reviewing or deciding an AI-assisted task, with benefits that depend on how the role works.", "definition": "A person may review a recommendation, correct an output or approve an action. This can combine different strengths, but the reviewer needs enough information, time and authority to disagree.", "placement": "Atlas reading question: can the named reviewer meaningfully change or stop what happens?", "distinction": "A confirmation button alone does not establish effective oversight. Experiments have found that erroneous system advice can influence a person's judgment even when that person makes the final decision.", "references": { "summary": [ "glossary-eval-human-interaction" ], "definition": [ "glossary-eval-human-interaction", "glossary-eval-human-errors" ], "placement": [ "glossary-eval-human-errors" ], "distinction": [ "glossary-eval-human-errors" ] }, "sources": [ "glossary-eval-human-interaction", "glossary-eval-human-errors" ] }, { "id": "automation-bias", "guide": "crosscutting", "group": "AI concepts", "short": "Automation bias", "term": "Automation bias", "category": "Human decision pattern", "summary": "Following automated advice too readily, including wrong advice; useful assistance can still create this problem.", "definition": "A person may accept a system's suggestion instead of checking it against other evidence. One study of pathology experts found better overall performance alongside some cases where wrong advice displaced a correct judgment.", "placement": "Atlas reading question: how does a workflow help people detect and reject a wrong recommendation?", "distinction": "This is a possible failure of reliance, not proof that people always trust machines or that all AI assistance reduces accuracy. Effects depend on the task and interaction.", "references": { "summary": [ "glossary-eval-automation-bias" ], "definition": [ "glossary-eval-automation-bias" ], "placement": [ "glossary-eval-human-errors" ], "distinction": [ "glossary-eval-automation-bias", "glossary-eval-human-interaction" ] }, "sources": [ "glossary-eval-automation-bias", "glossary-eval-human-errors", "glossary-eval-human-interaction" ] }, { "id": "data-poisoning", "guide": "crosscutting", "group": "AI concepts", "short": "Data poisoning", "term": "Data poisoning", "category": "Security concept", "summary": "Deliberately altering training material to influence a model's later behavior in an attacker's favor.", "definition": "An attacker inserts or changes training examples, for instance to cause particular errors or make a hidden trigger affect outputs. Poisoning can target initial training or later training stages.", "placement": "Atlas reading question: how are the origins and integrity of training material checked?", "distinction": "Accidental low-quality data is not necessarily poisoning. Data checks and filtering can help, but NIST notes that finding malicious examples in a large training collection can be difficult.", "references": { "summary": [ "glossary-eval-aml-taxonomy" ], "definition": [ "glossary-eval-aml-taxonomy" ], "placement": [ "glossary-eval-aml-taxonomy" ], "distinction": [ "glossary-eval-aml-taxonomy" ] }, "sources": [ "glossary-eval-aml-taxonomy" ] }, { "id": "computer-vision", "category": "Application", "guide": "crosscutting", "group": "AI concepts", "short": "Computer vision", "term": "Computer vision", "summary": "Software that extracts information from images, such as objects or written text. Different tasks need different checks.", "definition": "Computer vision covers tasks such as labeling an image, locating objects and reading text in a picture. For example, a system could mark the position of a bicycle in a photo.", "placement": "Map context: this names a set of technical tasks. We do not assign it a preferred pace of AI development.", "distinction": "Image classification assigns a category; object detection also locates objects. Neither task is the same as producing a new image.", "references": { "summary": [ "glossary-wide-vision" ], "definition": [ "glossary-wide-vision" ], "placement": [ "glossary-wide-vision" ], "distinction": [ "glossary-wide-vision" ] }, "sources": [ "glossary-wide-vision" ] }, { "id": "speech-recognition", "category": "Application", "guide": "crosscutting", "group": "AI concepts", "short": "Speech recognition", "term": "Automatic speech recognition (ASR)", "summary": "Software that turns spoken audio into written words. A readable transcript can still contain mistakes.", "definition": "An ASR system processes a recording and produces text, for example subtitles or a draft meeting transcript. The Hugging Face examples show that a plausible word can replace the word actually spoken.", "placement": "Map context: speech transcription is a capability, not a position on AI risk or development speed.", "distinction": "Transcribing words, identifying a speaker and translating into another language are different tasks. Check important wording against the recording.", "references": { "summary": [ "glossary-wide-asr" ], "definition": [ "glossary-wide-asr" ], "placement": [ "glossary-wide-asr" ], "distinction": [ "glossary-wide-asr" ] }, "sources": [ "glossary-wide-asr" ] }, { "id": "speech-synthesis", "category": "Application", "guide": "crosscutting", "group": "AI concepts", "short": "Speech synthesis", "term": "Speech synthesis / text-to-speech (TTS)", "summary": "Software that produces spoken audio, often from written text, for example to read a document aloud.", "definition": "Text-to-speech systems turn text into a speech signal. They can be used to read a document aloud or provide an audio version of a message. Speech generation is one type of audio generation.", "placement": "Map context: this describes an output method. We do not treat a synthetic voice as an actor or an ideology.", "distinction": "Speech recognition turns audio into text. Speech synthesis goes the other way. Producing speech does not necessarily mean imitating a particular real person.", "references": { "summary": [ "glossary-wide-tts" ], "definition": [ "glossary-wide-tts" ], "placement": [ "glossary-wide-tts" ], "distinction": [ "glossary-wide-tts" ] }, "sources": [ "glossary-wide-tts" ] }, { "id": "diffusion-models", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "short": "Diffusion", "term": "Diffusion models", "summary": "Generative models that can build an image through repeated removal of noise. Generating a picture does not establish that the pictured event happened.", "definition": "In the denoising approach, training examples are corrupted with noise and a model is trained to reverse that process. Generation starts with noise and applies learned steps to produce an output.", "placement": "Map context: this is a way to generate content. Its use does not establish a view about catastrophic risk or faster development.", "distinction": "The classic diffusion approach refines a noisy representation over successive steps. That differs from a language model generating text one token at a time; both are computational methods.", "references": { "summary": [ "glossary-wide-diffusion" ], "definition": [ "glossary-wide-diffusion" ], "placement": [ "glossary-wide-diffusion" ], "distinction": [ "glossary-wide-diffusion" ] }, "sources": [ "glossary-wide-diffusion" ] }, { "id": "latent-space", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "short": "Latent space", "term": "Latent space", "summary": "An internal system of numerical coordinates for representing features, such as features of an image.", "definition": "In latent diffusion, an image is compressed into a learned numerical representation. The diffusion process works on that representation, then a decoder converts the result into pixels. The set of possible representations is called a latent space.", "placement": "Map context: latent coordinates belong to a model representation. They are unrelated to the editorial coordinates used for public positions in this atlas.", "distinction": "A latent representation is not a miniature image or a written explanation. In latent diffusion, compression reduces detail as well as computational work; its design affects reconstruction quality.", "references": { "summary": [ "glossary-wide-latent-diffusion" ], "definition": [ "glossary-wide-latent-diffusion" ], "placement": [ "glossary-wide-latent-diffusion" ], "distinction": [ "glossary-wide-latent-diffusion" ] }, "sources": [ "glossary-wide-latent-diffusion" ] }, { "id": "synthetic-data", "category": "Using AI", "guide": "crosscutting", "group": "AI concepts", "short": "Synthetic data", "term": "Synthetic data", "summary": "Artificially generated examples used as data. They can support testing and analysis, but being synthetic does not guarantee privacy.", "definition": "A generator can produce new rows that resemble patterns in an original table. The resulting dataset can be compared with real data to check whether useful patterns were preserved. For example, a team could generate sample customer records to test software.", "placement": "Map context: this is a data-production method. We do not place it on the pace or concern axes.", "distinction": "Synthetic content describes generated media; synthetic data emphasizes its use as examples for analysis, testing or training. Some generation and evaluation methods can expose information about original records, so privacy needs separate assessment.", "references": { "summary": [ "glossary-wide-synthetic-data", "glossary-wide-synthetic-privacy" ], "definition": [ "glossary-wide-synthetic-data" ], "placement": [ "glossary-wide-synthetic-data" ], "distinction": [ "glossary-wide-synthetic-data", "glossary-wide-synthetic-privacy", "nist-synthetic-content" ] }, "sources": [ "glossary-wide-synthetic-data", "glossary-wide-synthetic-privacy", "nist-synthetic-content" ] }, { "id": "knowledge-cutoff", "category": "Using AI", "guide": "crosscutting", "group": "AI concepts", "short": "Knowledge cutoff", "term": "Knowledge cutoff", "summary": "A reported date describing the freshness of a model’s training information. It is not a guarantee of complete or correct knowledge before that date.", "definition": "A provider may state when training material was collected. Researchers distinguish that reported date from an effective cutoff: how recent the information a model can actually reproduce is for a particular topic or source.", "placement": "Map context: a cutoff is a limit to check when using a model as an information source. It is not a measure of intelligence or a map position.", "distinction": "Different topics can have different effective cutoffs. A single advertised date should not be read as a promise that every earlier event is covered or that every answer is up to date.", "references": { "summary": [ "glossary-wide-knowledge-cutoff" ], "definition": [ "glossary-wide-knowledge-cutoff" ], "placement": [ "glossary-wide-knowledge-cutoff" ], "distinction": [ "glossary-wide-knowledge-cutoff" ] }, "sources": [ "glossary-wide-knowledge-cutoff" ] }, { "id": "data-provenance", "category": "Using AI", "guide": "crosscutting", "group": "AI concepts", "short": "Data provenance", "term": "Data provenance", "summary": "A record of where data came from and how it changed. It helps people assess evidence but does not prove the data is correct.", "definition": "Provenance describes the sources, people and processes involved in producing data. It can record that one dataset was derived from another, who carried out a step, and which version was used.", "placement": "Map context: we use provenance to let readers inspect the origin of claims. It is a documentation practice, not a political position.", "distinction": "An origin record and a truth assessment answer different questions. Knowing who supplied a claim helps investigation; the claim and the record can still require checking.", "references": { "summary": [ "glossary-wide-provenance" ], "definition": [ "glossary-wide-provenance" ], "placement": [ "glossary-wide-provenance" ], "distinction": [ "glossary-wide-provenance" ] }, "sources": [ "glossary-wide-provenance" ] }, { "id": "watermarking", "category": "Using AI", "guide": "crosscutting", "group": "AI concepts", "short": "Watermarking", "term": "Digital watermarking", "summary": "Information embedded in text, images or audio to help identify its origin. A watermark can be missed, removed or wrongly detected.", "definition": "A watermark places a signal inside the content itself. It might be a visible mark or a subtle pattern detectable by software. Some schemes indicate that a particular generator produced the content.", "placement": "Map context: watermarking is one approach to documenting media origins. It does not tell us where an actor belongs on the map.", "distinction": "A watermark differs from a separate metadata record. Its presence does not establish that a claim is true, and its absence does not establish that a human made the content.", "references": { "summary": [ "nist-synthetic-content" ], "definition": [ "nist-synthetic-content" ], "placement": [ "nist-synthetic-content" ], "distinction": [ "nist-synthetic-content" ] }, "sources": [ "nist-synthetic-content" ] }, { "id": "model-card", "category": "Access & release", "guide": "crosscutting", "group": "AI concepts", "short": "Model card", "term": "Model card", "summary": "A document describing a model’s intended uses, tests and limitations. It helps scrutiny but is not a safety certificate.", "definition": "The model-card proposal calls for information about what a model is for, how it was evaluated and how its performance varies across relevant conditions or groups. A useful card makes those limits visible alongside results.", "placement": "Map context: a model card is evidence to inspect when evaluating claims about a model. Publishing one does not establish an organization’s policy position.", "distinction": "A model card focuses on a model. A system card considers how models and other parts work together in a product. Neither title alone establishes how complete the reporting is.", "references": { "summary": [ "glossary-wide-model-card" ], "definition": [ "glossary-wide-model-card" ], "placement": [ "glossary-wide-model-card" ], "distinction": [ "glossary-wide-model-card", "glossary-wide-system-card" ] }, "sources": [ "glossary-wide-model-card", "glossary-wide-system-card" ] }, { "id": "system-card", "category": "Access & release", "guide": "crosscutting", "group": "AI concepts", "short": "System card", "term": "System card", "summary": "A document about how the parts of an AI system work together and where its limits lie. It can become outdated as the product changes.", "definition": "A system card can describe models, non-AI components, data flows and the way outputs affect a product. Meta’s proposal explains why a model’s behavior needs to be considered in the system where it is used.", "placement": "Map context: system documentation helps readers inspect a product beyond an individual model. We do not infer an actor’s beliefs from the existence of a card.", "distinction": "The term describes documentation, not an independently verified stamp of approval. Read its scope, date and limitations; a card about one version may not describe another.", "references": { "summary": [ "glossary-wide-system-card" ], "definition": [ "glossary-wide-system-card" ], "placement": [ "glossary-wide-system-card" ], "distinction": [ "glossary-wide-system-card" ] }, "sources": [ "glossary-wide-system-card" ] }, { "id": "compute", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "short": "Compute resources", "term": "Compute", "summary": "The computing resources used to train or run a model. More resources describe an input, not whether the result is useful.", "definition": "Compute can refer to resources such as processing power, memory and storage. In AI discussions, it is worth checking whether someone means resources for training a model or resources for running it afterwards.", "placement": "Map context: this is a resource term. A claim about compute alone does not establish a preference for faster or slower frontier development.", "distinction": "Compute, model size and elapsed time are different descriptions. A parameter count describes model values; compute describes resources used in doing the work.", "references": { "summary": [ "google-ml-glossary" ], "definition": [ "google-ml-glossary" ], "placement": [ "google-ml-glossary" ], "distinction": [ "google-ml-glossary" ] }, "sources": [ "google-ml-glossary" ] }, { "id": "gpu", "category": "Inside a model", "guide": "crosscutting", "group": "AI concepts", "short": "GPU", "term": "Graphics processing unit (GPU)", "summary": "A chip designed to do many calculations in parallel. It can accelerate suitable AI workloads, but it is not itself an AI model.", "definition": "GPUs began as processors for graphics. Their parallel design is also used for calculations in model training and generation. A GPU runs software; the trained model’s numerical values are separate from the chip.", "placement": "Map context: a GPU is hardware. Owning or using one does not identify a movement or a position on AI risk.", "distinction": "A CPU emphasizes fast sequences of operations; a GPU emphasizes many operations in parallel. Which is useful depends on the work, so a GPU is not automatically faster for every program.", "references": { "summary": [ "glossary-wide-gpu" ], "definition": [ "glossary-wide-gpu" ], "placement": [ "glossary-wide-gpu" ], "distinction": [ "glossary-wide-gpu" ] }, "sources": [ "glossary-wide-gpu" ] }, { "id": "latency", "category": "Using AI", "guide": "crosscutting", "group": "AI concepts", "short": "Response time", "term": "Latency", "summary": "The waiting time for a system’s response. For generated text, the first visible word and the completed answer have different waiting times.", "definition": "Latency is a duration. Language-model benchmarks can measure time until the first output token and the time taken to produce later tokens. These measures describe different parts of the waiting experience.", "placement": "Map context: response speed is a product-performance measure. It is separate from the map’s axis about the pace of developing more capable AI.", "distinction": "Throughput describes how much work a system processes in a period. A system that handles many requests overall does not necessarily answer each individual request quickly.", "references": { "summary": [ "glossary-wide-latency" ], "definition": [ "glossary-wide-latency" ], "placement": [ "glossary-wide-latency" ], "distinction": [ "glossary-wide-latency" ] }, "sources": [ "glossary-wide-latency" ] }, { "id": "on-device-ai", "category": "Application", "guide": "crosscutting", "group": "AI concepts", "short": "On-device AI", "term": "On-device AI / edge AI", "summary": "Running a model on a device such as a phone or laptop. Where a calculation runs does not by itself establish the privacy of the whole app.", "definition": "On-device AI runs model calculations locally. A model can be obtained or trained elsewhere, adapted for the target device and then run there. Edge AI is also used for processing near the source of data.", "placement": "Map context: this describes where computation happens. We do not treat local processing as a movement or a position on frontier pacing.", "distinction": "On-device inference is different from training on the device. To assess privacy, check the whole application’s data flows, including any other services it uses.", "references": { "summary": [ "glossary-wide-on-device", "glossary-wide-system-card" ], "definition": [ "glossary-wide-on-device" ], "placement": [ "glossary-wide-on-device" ], "distinction": [ "glossary-wide-on-device", "glossary-wide-system-card" ] }, "sources": [ "glossary-wide-on-device", "glossary-wide-system-card" ] }, { "id": "software", "short": "Computer programs", "term": "Software & computer programs", "category": "AI basics", "guide": "crosscutting", "group": "AI concepts", "references": { "summary": [ "software-lens-nist-computer", "software-lens-nist-software" ], "definition": [ "hf-tool-use", "hf-models", "google-ml-intro", "software-lens-nist-software" ], "placement": [ "software-lens-nist-software", "hf-models" ], "distinction": [ "circuit-tracing" ] }, "summary": "Computer programs tell a machine what operations to carry out. Software includes those programs and their associated data.", "definition": "A calculator program might multiply two numbers using a written rule. A language-model program runs calculations using numerical settings learned during training. Both are executed by computer hardware.", "placement": "Map context: this explains how systems work; it is not a position for or against faster AI development.", "distinction": "Being software does not make a model’s behavior easy to explain. Interpretability research studies how its learned computations produce particular outputs.", "sources": [ "software-lens-nist-computer", "software-lens-nist-software", "hf-tool-use", "hf-models", "google-ml-intro", "circuit-tracing" ] }, { "id": "ai-system", "short": "AI systems", "term": "AI system", "category": "AI basics", "guide": "crosscutting", "group": "AI concepts", "references": { "summary": [ "software-lens-oecd-system", "ncsc-secure-ai" ], "definition": [ "hf-tool-use", "owasp-excessive-agency", "software-lens-oecd-system" ], "placement": [ "software-lens-oecd-system", "ncsc-secure-ai" ], "distinction": [ "ncsc-agentic-risk", "owasp-excessive-agency" ] }, "summary": "An AI model together with the software and hardware that put it to use.", "definition": "For a chatbot, this can include the interface, instructions, model, search tools and permissions. A robot can also include sensors and motors. These parts shape what the system can do.", "placement": "Map context: distinguish a claim about a model from a claim about the application or machine that uses it.", "distinction": "A model’s results alone do not establish that a whole system is safe. Tools, access rights and oversight need their own assessment.", "sources": [ "software-lens-oecd-system", "ncsc-secure-ai", "hf-tool-use", "owasp-excessive-agency", "ncsc-agentic-risk" ] }, { "id": "normal-technology", "short": "Normal technology", "term": "AI as normal technology", "category": "Theoretical thesis", "guide": "crosscutting", "group": "AI concepts", "summary": "Narayanan and Kapoor's view that people can shape AI's impacts through institutions, engineering and policy. Its safety claims are contested: in September 2026, they acknowledged underestimating risks during development and called for stronger controls.", "definition": "The thesis compares AI with earlier powerful technologies. It separates technical progress from applications and adoption, and treats continued human control as a goal requiring choices. Normal does not mean harmless, simple or predictable.", "placement": "Map interpretation: context across both axes, not a measured point or membership label. Its authors support pausing unsafe experiments where needed, while rejecting claims that catastrophic risks are imminent.", "distinction": "Scott Alexander argues that rapid self-improvement and adoption by AI labs could bypass assumed limits. The authors reply that technical improvements do not automatically remove external constraints. Their September update gives more weight to developer oversight.", "references": { "summary": [ "normal-technology-2025", "normal-technology-alexander-response-2025", "normal-technology-control-2026" ], "definition": [ "normal-technology-2025", "normal-technology-guide-2025" ], "placement": [ "normal-technology-control-2026" ], "distinction": [ "normal-technology-alexander-response-2025", "normal-technology-guide-2025", "normal-technology-control-2026" ] }, "sources": [ "normal-technology-2025", "normal-technology-guide-2025", "normal-technology-alexander-response-2025", "normal-technology-control-2026" ] }, { "id": "symbolic-ai", "short": "Symbolic AI", "term": "Symbolic AI & expert systems", "category": "AI basics", "guide": "crosscutting", "group": "AI concepts", "summary": "Programs that work with explicitly represented facts and rules. An expert system applies such rules within a particular subject.", "definition": "Developers encode facts and logical rules that software applies to a problem. An expert system combines a knowledge base with rules, often written as if-then conditions, drawn from a specific field.", "placement": "Map context: this explains an AI approach based on explicitly programmed knowledge.", "distinction": "An expert system is one application of symbolic methods. Its encoded knowledge covers a limited domain. This differs from fitting a model's parameters to examples through machine learning.", "references": { "summary": [ "glossary-symbolic-stanford", "glossary-expert-stanford" ], "definition": [ "glossary-symbolic-stanford", "glossary-expert-stanford" ], "placement": [ "glossary-symbolic-stanford" ], "distinction": [ "glossary-expert-stanford", "google-ml-intro" ] }, "sources": [ "glossary-symbolic-stanford", "glossary-expert-stanford", "google-ml-intro" ] }, { "id": "narrow-ai", "short": "Specialized AI", "term": "Narrow AI / specialized AI", "category": "Capability concept", "guide": "crosscutting", "group": "AI concepts", "summary": "AI for a limited task or set of tasks. Specialized software can perform very well within that scope.", "definition": "Narrow describes breadth, not quality. A research framework separates how many kinds of tasks a system handles from how well it performs them.", "placement": "Map context: keep specialist performance separate from claims about general intelligence.", "distinction": "Breadth, performance and autonomy are different questions. Strong task results do not determine how much control an application should receive.", "references": { "summary": [ "agi-levels" ], "definition": [ "agi-levels" ], "placement": [ "agi-levels" ], "distinction": [ "agi-levels" ] }, "sources": [ "agi-levels" ] }, { "id": "foundation-models", "short": "Foundation models", "term": "Foundation models", "category": "AI basics", "guide": "crosscutting", "group": "AI concepts", "summary": "Models trained on broad data that can be adapted for many tasks. Reusing one model can also spread its weaknesses across applications.", "definition": "A shared trained model becomes a starting point for different uses, through prompts or further training. The term includes more than language models.", "placement": "Map context: shared models connect development choices to many downstream applications.", "distinction": "A foundation model is a reusable component. Its broad training does not establish equal performance across applications.", "references": { "summary": [ "foundation-model-report" ], "definition": [ "foundation-model-report" ], "placement": [ "foundation-model-report" ], "distinction": [ "foundation-model-report" ] }, "sources": [ "foundation-model-report" ] }, { "id": "scaling-laws", "short": "Scaling laws", "term": "AI scaling laws", "category": "Research concept", "guide": "crosscutting", "group": "AI concepts", "summary": "Measured patterns linking model size, training data and computing resources to performance. Each pattern concerns a particular measurement and training setup.", "definition": "Researchers train different-sized models and fit equations to the results. In language-model research, a common measure is prediction loss: how poorly the model predicts the next text piece.", "placement": "Map context: these studies inform expectations about capability growth and resource use.", "distinction": "Larger is not the only choice. Hoffmann and coauthors found better results by balancing model size with more training data. Extrapolating a measured trend to untested scales adds an assumption.", "references": { "summary": [ "scaling-kaplan", "scaling-chinchilla" ], "definition": [ "scaling-kaplan", "scaling-chinchilla" ], "placement": [ "scaling-kaplan", "scaling-chinchilla" ], "distinction": [ "scaling-kaplan", "scaling-chinchilla" ] }, "sources": [ "scaling-kaplan", "scaling-chinchilla" ] }, { "id": "model-collapse", "short": "Model collapse", "term": "Model collapse", "category": "Model behavior", "guide": "crosscutting", "group": "AI concepts", "summary": "Repeatedly training on earlier models' output can erase patterns from the original data. Other experiments avoided this degradation by retaining the original dataset while adding generated examples.", "definition": "In studied training loops, each model supplies examples for the next. Errors can accumulate, and uncommon patterns can disappear from the later models' output.", "placement": "Map context: this qualifies claims about generated data sustaining future model training.", "distinction": "The result depends on how datasets are replaced or accumulated. It does not show that all synthetic data is harmful, or diagnose a chatbot's bad answer without examining its training.", "references": { "summary": [ "model-collapse-nature", "model-collapse-accumulation" ], "definition": [ "model-collapse-nature" ], "placement": [ "model-collapse-nature", "model-collapse-accumulation" ], "distinction": [ "model-collapse-nature", "model-collapse-accumulation" ] }, "sources": [ "model-collapse-nature", "model-collapse-accumulation" ] }, { "id": "data-work", "short": "Human data work", "term": "Human data work & data labeling", "category": "Work & society", "guide": "crosscutting", "group": "Ethics", "summary": "People collect, label, check and evaluate data used in AI. Their work and the instructions they receive help shape trained software.", "definition": "Data labeling attaches categories or other annotations to examples. Wider data work includes collecting material, preparing datasets and checking model outputs. Research documents this work in different employment arrangements.", "placement": "Map context: AI development includes labor and organizational choices, alongside model design and computing resources.", "distinction": "A label is a judgment made for a task, not automatically a fact. A study of AI practitioners found that neglected data work could cause problems later in development and deployment.", "references": { "summary": [ "data-work-typology", "data-cascades-author-summary" ], "definition": [ "data-work-typology" ], "placement": [ "data-work-typology", "data-cascades-author-summary" ], "distinction": [ "data-work-typology", "data-cascades-author-summary" ] }, "sources": [ "data-work-typology", "data-cascades-author-summary" ] }, { "id": "ai-control", "short": "AI control", "term": "AI control", "category": "Research field", "guide": "crosscutting", "group": "AI concepts", "summary": "Tests whether safeguards can block harmful actions even when model outputs are chosen to bypass them. Monitoring models have also been bypassed in such tests.", "definition": "Researchers test whole workflows against adversarial behavior. The original control study used programming tasks and tested reviewing or editing untrusted code with another model.", "placement": "Map context: connects permissions, monitoring and review to safeguards around deployed software.", "distinction": "Control can add checks around a model without changing its learned weights. It can complement alignment work. A passed test supports only its stated setup and attacks.", "references": { "summary": [ "ai-control-original", "ai-control-monitor-attacks" ], "definition": [ "ai-control-original" ], "placement": [ "ai-control-threats" ], "distinction": [ "ai-control-original", "ai-control-monitor-attacks" ] }, "sources": [ "ai-control-original", "ai-control-monitor-attacks", "ai-control-threats" ] } ], "config": { "width": 1920, "height": 1280, "plot": { "x": 180, "y": 220, "width": 1640, "height": 830 }, "fontFamily": "Arial, Liberation Sans, Helvetica, sans-serif", "labelFontSize": 25, "titleFontSize": 58, "colors": { "ink": "#202B3C", "muted": "#58677D", "grid": "#D8DFE9", "person": "#7047A3", "organization": "#087F80", "movement": "#A15B13", "topLeft": "#F5F2FA", "topRight": "#EFF5FA", "bottomLeft": "#F7F8FA", "bottomRight": "#FAF6EE", "paper": "#FFFFFF", "bg": "#F3F5F8", "surface": "#F7F9FC", "line": "#D8DFE9", "controlBorder": "#7F8CA0", "accent": "#2458B8", "accentHover": "#19438F", "accentSoft": "#EAF0FC", "focus": "#2458B8", "warning": "#805509", "warningSoft": "#FCF5E5", "positive": "#276746", "positiveSoft": "#EBF5EE", "concept": "#65458F", "conceptSoft": "#F2EFF8", "restraint": "#65458F", "development": "#245B82", "acceleration": "#805509", "onAccent": "#FFFFFF", "codeBg": "#202B3C", "codeInk": "#F4F6FA" }, "designTokens": { "space-1": "4px", "space-2": "8px", "space-3": "12px", "space-4": "16px", "space-5": "20px", "space-6": "24px", "space-8": "32px", "space-10": "40px", "space-12": "48px", "text-xs": "0.875rem", "text-sm": "1rem", "text-base": "1rem", "text-lg": "1.25rem", "text-xl": "1.5rem", "text-2xl": "2rem", "line-small": "1.25rem", "line-body": "1.5rem", "line-heading": "2rem", "radius-control": "6px", "radius-panel": "12px", "control-height": "44px", "touch-target": "44px", "motion-fast": "120ms", "shadow-panel": "0 16px 48px #202B3C26" }, "quadrants": [ { "id": "nw", "title": "HIGHER CONCERN · RESTRAINT", "subtitle": "Pause, slow or pace the frontier" }, { "id": "ne", "title": "HIGHER CONCERN · DEVELOPMENT", "subtitle": "Build with safeguards; risk-aware progress" }, { "id": "sw", "title": "LOWER CONCERN · RESTRAINT", "subtitle": "Restraint can have other motivations" }, { "id": "se", "title": "LOWER CONCERN · DEVELOPMENT", "subtitle": "Techno-optimism and acceleration" } ], "darkColors": { "ink": "#F3F3F4", "muted": "#BDBDC6", "grid": "#54545D", "person": "#C0A2E3", "organization": "#62CBCB", "movement": "#E8B773", "topLeft": "#24202B", "topRight": "#1D262E", "bottomLeft": "#232326", "bottomRight": "#2C271E", "paper": "#171719", "bg": "#0D0D0F", "surface": "#202023", "line": "#484850", "controlBorder": "#82828D", "accent": "#A6C7FF", "accentHover": "#D0E1FF", "accentSoft": "#202A3D", "focus": "#A6C7FF", "warning": "#F4CF89", "warningSoft": "#2D261B", "positive": "#91D6AE", "positiveSoft": "#1D2A22", "concept": "#C9B4E7", "conceptSoft": "#26222E", "restraint": "#C9B4E7", "development": "#9CCDF3", "acceleration": "#F4CF89", "onAccent": "#101217", "codeBg": "#0C0C0E", "codeInk": "#EEEEF2" } }, "guide": { "reviewedOn": "2026-09-15", "explainer": { "title": "How does an LLM produce an answer?", "sections": [ { "id": "llm-software", "title": "Your request becomes an input to a computer program", "text": "Imagine asking for a program that totals a shopping bill. A large language model (LLM) runs as software on a computer. It uses programmed calculations and learned numerical settings called weights. Training adjusts those settings using examples. When drafting your program, it scores possible next tokens, which are pieces of text.", "sources": [ "transformer-paper", "hf-generation", "copilot-code-review" ], "claim": "C1", "depths": { "d1": { "text": "Imagine asking for a program that adds up a shopping bill. A language model is also a computer program. Programmers write its calculations. Training uses examples to adjust many number settings inside it. When writing, it uses those settings to score the text pieces that could come next. A piece might be a word or part of one." }, "d2": { "text": "Imagine asking for a program that totals a shopping bill. A large language model (LLM) runs as a computer program. Programmers specify its calculations; training uses examples to adjust many numerical settings, called weights. When it answers, its neural network uses those weights to score possible next tokens, the pieces of text it works with." }, "d4": { "text": "An LLM combines programmed calculations with learned parameters. In the text-generation setup shown here, a transformer maps a token sequence to scores for the next token. Training adjusts its parameters using a next-token objective on example text; further tuning stages can follow. During inference, the programmed forward pass applies the trained parameters to the supplied context. A draft shopping program is an output of that process. It still needs review and testing before anyone runs it.", "sources": [ "google-llm-tuning" ] } } }, { "id": "llm-token-choice", "title": "The answer is built one piece at a time", "text": "Software chooses a token from the model's scores, adds it to the answer and repeats. It can pick the highest-scoring option or sample among several options. Sampling can produce different drafts of your shopping program. The model can also run without sampling; variation is a separate question from how well we understand its workings.", "sources": [ "hf-generation", "pytorch-reproducibility", "circuit-tracing" ], "claim": "C2", "depths": { "d1": { "text": "The answer is not written all at once. The model scores the pieces that could come next. The software picks one piece, adds it to the answer, and repeats the calculation. It can always take the top score, or pick at random using the scores. Random picking can make the same question produce different answers." }, "d2": { "text": "The model gives each possible next token a score. Software then chooses one token, adds it to the answer and repeats until a stopping rule is met. It can always choose the highest score, or sample: pick randomly, with higher scores more likely. Sampling can produce different drafts of your shopping program. Running without sampling removes that randomness, but it does not explain why the model scored things the way it did. Those are two separate questions." }, "d4": { "text": "Each forward pass yields logits over the vocabulary; a decoding strategy selects one token, appends it to the context, and the pass repeats. Greedy decoding takes the highest score; sampling draws from the softmax distribution, optionally reshaped by temperature, top-k or top-p, until an end token or a length limit. Sampling is the intended source of output variance. Its absence says nothing about interpretability: a greedy run is still produced by the same learned weights, whose contributions interpretability research can only partly trace." } } }, { "id": "llm-generated-code", "title": "Check the draft, then run the saved program", "text": "The model might draft a rule that multiplies price by quantity. Check that it matches your needs and test it. Once saved, that code can run without the LLM. For the same complete inputs and execution conditions, deterministic code gives the same result. It can also repeat the same mistake, such as leaving out delivery charges.", "sources": [ "copilot-code-review", "pytorch-reproducibility" ], "claim": "C3", "depths": { "d1": { "text": "The model might write a rule: price times how many. Someone must read it and test it. Once saved, this rule can run without the model. With the same numbers and setup, this multiplication gives the same answer. If the rule leaves out the delivery charge, it leaves it out every time." }, "d2": { "text": "The model might draft code that multiplies price by quantity. Check that draft against what you need and test it. Once saved, this multiplication can run without the model. For fixed inputs and execution conditions, it gives the same result; this is called deterministic. Saving generated code does not by itself make every program deterministic or correct. The shopping rule can repeat the same mistake, such as leaving out delivery charges." }, "d4": { "text": "Review and test a generated code draft before running it. In the shopping example, a saved multiplication function can run without calling the model. When that function uses the same complete inputs and fixed execution conditions, its result is deterministic. This describes the specified function, not all generated software. A repeatable function can still implement the wrong rule. Passing tests cover the cases checked; they do not establish correctness for every input or explain how the model produced the draft." } } }, { "id": "llm-learned-behavior", "title": "Why did the model choose that answer?", "text": "You can inspect the shopping program's calculation. Explaining why the model drafted that particular program is harder: its behavior depends on many learned weights working together. Knowing the calculations does not automatically reveal a simple reason for each answer. Researchers studying mechanistic interpretability trace parts of this process, with important gaps still remaining.", "sources": [ "transformer-paper", "circuit-tracing", "attention-tracing" ], "claim": "C4", "depths": { "d1": { "text": "With the saved shopping rule, you can follow the line that multiplies the numbers. The model also runs code, but explaining its answer takes more than reading that code. Many learned settings work together to score each piece. Scientists can trace some of what happens inside, but important gaps remain." }, "d2": { "text": "You can read the shopping program and follow its calculation. Explaining why the model drafted that particular program is harder. The answer depends on many learned weights working together. Knowing every calculation does not give a simple reason for each answer. Researchers in mechanistic interpretability trace parts of this process, and important gaps remain." }, "d4": { "text": "Every operation in the forward pass is known; what is missing is a compact causal account of why those operations produced this output. Mechanistic interpretability methods such as attribution graphs and attention tracing recover partial circuits for specific behaviors, with stated limits on coverage and faithfulness. Opacity in this sense is a property of the learned function, independent of whether decoding is stochastic." } } }, { "id": "llm-weather-analogy", "title": "An illustration: knowing the rules is not a complete explanation", "optional": true, "text": "Weather helps illustrate the gap described above: knowing physical rules does not give a simple explanation of every outcome. Forecasts face uncertainty about starting conditions and approximations. With an LLM, the question here is how learned parts produce an answer. This is our limited analogy about explanation, not a claim that LLMs behave like weather or must be random.", "sources": [ "weather-uncertainty", "circuit-tracing", "attention-tracing" ], "claim": "C4", "depths": { "d1": { "text": "Here is a limited comparison. Knowing weather's physical rules does not fully explain every outcome. Forecasts face gaps in starting information and use simplified calculations. For a model, the question is how its learned settings help produce an answer. These are different gaps. The comparison does not show that models behave like weather or must be random." }, "d2": { "text": "An analogy: we know the physical rules of weather, but a forecast still cannot explain every outcome, because starting conditions and approximations are uncertain. With a model the open question is different: how learned parts combine to produce an answer. This is a limited analogy about explanation. It is not a claim that models behave like weather or must be random." }, "d4": { "text": "The comparison is a limited analogy. Numerical weather prediction has known dynamics and still limited explanatory power per outcome, from initial-condition uncertainty and approximation. For an LLM the operations are also known, and the gap is attributional: which learned components caused this output. The analogy illustrates that knowing the rules is not a complete explanation. It does not import chaotic sensitivity, a claim that models must be random, or any physical claim about models." } } }, { "id": "llm-reproducibility", "title": "More detail: what makes an answer repeatable?", "optional": true, "text": "Turning off sampling does not guarantee identical answers across every setup. Hardware, software versions and some calculations can still affect results. A random seed helps repeat a sequence of choices under matching conditions. Repeatability means reproducing a result; interpretability means explaining how it came about. One does not establish the other.", "sources": [ "pytorch-reproducibility", "circuit-tracing" ], "claim": "C3", "depths": { "d1": { "text": "Even with random picking turned off, two computers can give slightly different answers. Different machines and different software versions can round numbers differently. A seed number lets you replay the same random picks on the same setup. Getting the same answer again is not the same as understanding why it was the answer." }, "d2": { "text": "Turning off sampling does not guarantee identical answers on every setup. Different hardware, software versions and some calculations can still change the result. A seed replays the same sequence of random choices under matching conditions. Repeatability means getting the same result again; interpretability means explaining how it came about. One does not establish the other." }, "d4": { "text": "Greedy decoding removes sampling variance but not all nondeterminism: floating-point non-associativity, nondeterministic kernels and library versions can shift the computed scores, so even greedy runs can differ across setups. Seeds replay a sampled sequence only under matching conditions, and a low temperature makes the top token more likely without guaranteeing identical runs. Reproducibility is a property of the execution setup; interpretability is a property of the explanation. Neither implies the other.", "sources": [ "hf-generation" ] } } }, { "id": "llm-faq-errors", "title": "Why can a confident answer be wrong?", "optional": true, "text": "Producing convincing text does not check whether every claim is true. An answer can include a made-up fact, quotation or reference; this is often called a hallucination. Open an important citation and check that it exists and supports the claim. An answer's confident tone is not a measure of its accuracy.", "sources": [ "nist-genai-profile", "development-chatgpt" ], "claim": "C5", "depths": { "d1": { "text": "The model is good at writing text that sounds right. Sounding right is not the same as being right. It can invent a fact, a quote or a book that does not exist. People call this a hallucination. If it matters, look it up. A sure-sounding answer can still be wrong." }, "d2": { "text": "Producing convincing text does not check whether each claim is true. An answer can include a made-up fact, quotation or reference; this is often called a hallucination. Open an important citation and check that it exists and says what the answer claims. A confident tone is not a measure of accuracy." }, "d4": { "text": "Generation selects tokens using scores computed by the model. The resulting text can still contain fabricated facts, quotations or references. Neither fluency nor a high next-token score establishes that a claim is true. Where accuracy matters, inspect the cited source and check that it supports the answer. Retrieved passages can be useful evidence, but still need checking.", "sources": [ "rag-paper", "hf-generation" ] } } }, { "id": "llm-faq-learning", "title": "Does it learn from our conversation?", "optional": true, "text": "Your messages can shape the next answer through the conversation context. Some apps also save information for future chats. Neither is the same as retraining the model's weights. Separately, providers may use conversations for later training under their policies and settings. Check both memory controls and training controls for the service you use.", "sources": [ "google-llm-tuning", "chatgpt-memory", "openai-training-data" ], "claim": "C6", "depths": { "d1": { "text": "An app can include earlier messages when it prepares the next input. That can shape the answer without changing the model's learned settings. Some apps also save notes for later chats. Separately, companies may use chats to train models under their rules and your settings. Check the controls for saved notes and training." }, "d2": { "text": "An application can include earlier messages in the input, where they can shape the next answer. Some apps also save information for future chats. Providing this text does not itself update the model's weights, the numbers adjusted during training. Separately, a provider may use conversations for later training under its policies and settings. Check both the memory controls and the training controls of the service you use." }, "d4": { "text": "Context, application memory and training are different mechanisms. Text included in the context conditions the model's next output without itself updating the parameters. Memory features can save information and supply it in later conversations; that is separate from training or tuning the weights. Whether a provider uses conversations for later training depends on its policies and the relevant account settings. Check both memory and training controls for the service in question." } } }, { "id": "llm-faq-search", "title": "Is generating an answer the same as searching the web?", "optional": true, "text": "A language model can generate text without searching. An application can also retrieve webpages or documents and give that material to the model. These are separate steps. Retrieved material can contain errors, and the answer can misrepresent it, so follow the sources when accuracy matters.", "sources": [ "hf-generation", "rag-paper", "nist-genai-profile" ], "claim": "C6", "depths": { "d1": { "text": "A model can write an answer without looking anything up. Some apps first search the web or your files and hand the results to the model. Those are two separate steps. What the search found can be wrong, and the answer can get it wrong too. When it matters, open the source." }, "d2": { "text": "A language model can generate text without searching. An application can also retrieve webpages or documents first and add that material to the model's input. These are separate steps done by different parts of the system. Retrieved material can contain errors, and the answer can misrepresent it, so follow the sources when accuracy matters." }, "d4": { "text": "Retrieval-augmented generation adds a retrieval step whose results are inserted into the context; the model's parameters are unchanged and it still generates by next-token prediction over that context. Retrieval quality bounds what the model can be grounded on, and generation can still misstate what was retrieved. Verify against the retrieved source, not the paraphrase." } } }, { "id": "llm-faq-consciousness", "title": "Does sounding human mean it has feelings?", "optional": true, "text": "Human-like conversation alone does not establish subjective experience: whether anything feels like something to the system. Researchers disagree about how to assess this. The linked study proposes indicators based on theories of consciousness, with explicit assumptions and uncertainty; a chatbot's claim to have feelings does not settle the question.", "sources": [ "ai-consciousness-indicators" ], "depths": { "d1": { "text": "A model can chat like a person. That does not show it has feelings. Researchers do not agree on how to test this. One study lists signs to look for, and says the question is open. If a chatbot says it has feelings, that is text it produced, not proof." }, "d2": { "text": "Sounding human does not establish subjective experience: whether anything feels like something to the system. Researchers disagree about how to assess this. The linked study proposes indicators drawn from theories of consciousness, with explicit assumptions and uncertainty. A chatbot's statement that it has feelings is generated text and does not settle the question." }, "d4": { "text": "Conversational fluency alone does not establish subjective experience. The linked report derives indicator properties from scientific theories of consciousness and applies them under stated assumptions, with substantial uncertainty. A model's report of having feelings is also generated output. Interpreting such a report requires an account of how it was produced and what would count as evidence; the report alone does not settle whether subjective experience is present." } } }, { "id": "llm-faq-jobs", "title": "Will AI replace my job?", "optional": true, "text": "A job usually combines several tasks. Automating one task does not show that the whole job will disappear. The ILO's 2025 study estimates which tasks could be affected; it does not count actual job losses. Adoption, cost and choices about how work is organized also affect what happens.", "sources": [ "ilo-ai-work-exposure" ], "depths": { "d1": { "text": "A job is made of many tasks. A model may do one task well. That does not mean the whole job goes away. One big study looked at which tasks could change. It did not count lost jobs. What happens also depends on what companies choose to do." }, "d2": { "text": "A job usually combines several tasks. Automating one task does not show that the whole job will disappear. The ILO's 2025 study estimates which tasks could be affected; it does not count actual job losses. Adoption, cost and choices about how work is organized also shape what happens." }, "d4": { "text": "Exposure estimates are task-level: they score which tasks could be performed by generative AI, not observed displacement. The ILO's 2025 study reports exposure under explicit assumptions about capability, with no measurement of actual job losses. Adoption rates, cost and organizational choices mediate outcomes, so exposure describes potential, not a forecast." } } }, { "id": "llm-understanding", "title": "Does fluent language mean understanding?", "text": "The critical argument is that convincing language does not by itself show an understanding of meaning. Bender and Koller focus on training from language patterns alone. Other research asks what a prediction model learns internally: the Othello study found information about a game board. These address related questions, but neither establishes what every current LLM understands.", "sources": [ "understanding-bender-koller", "understanding-othello" ], "depths": { "d1": { "text": "A convincing answer alone does not prove that a model has learned meaning. One argument says text patterns alone are not enough to learn meaning. A separate study found signs of a game board inside a model trained on game moves. These studies ask related questions. Neither settles what every language model can do or how to describe it." }, "d2": { "text": "A convincing answer alone does not settle whether a model has learned meaning. Bender and Koller argue that training only on language patterns is insufficient for that. Another study examined a model trained to predict Othello moves and found information about the board inside it. That is a specific result about a game model. It does not settle the wider debate about understanding in language models." }, "d4": { "text": "Bender and Koller distinguish linguistic form from meaning and argue that training on form alone cannot establish the connection they call meaning. The Othello study investigates learned representations in a model trained to predict game moves and finds information about board states. This provides evidence about that model's internal structure, not a general verdict on human-like understanding. The studies address related but different questions. Their results do not establish what every current LLM understands." } } }, { "id": "llm-open-questions", "title": "Which questions are still open?", "text": "How reliably can a model learn a new concept from a few examples, explain causes, or revise a plan when its goal changes? Which tests separate those abilities from shortcuts? How faithfully does written reasoning report what influenced an answer? How do human-like names and voices affect trust and responsibility? How much of human thinking is captured in the explanations used for training? The cited studies do not settle that last question or establish any provider's motives.", "sources": [ "understanding-mitchell-krakauer", "understanding-faithfulness", "understanding-anthropomorphism" ], "depths": { "d1": { "text": "Can a model use a new idea after just a few examples? Can it explain causes or change plans? Which tests would rule out shortcuts? Do its written reasons show what shaped the answer? Do names and voices affect trust? How much human thinking appears in training text? The cited studies do not settle every question or reveal a company's motives." }, "d2": { "text": "How reliably can a model learn a new idea from a few examples, explain causes, or change a plan when its goal changes? Which tests could separate those abilities from shortcuts? Do its written reasons reflect what actually shaped its answer? How do human-like names and voices affect trust and responsibility? How much human thinking is captured in the explanations used for training? The linked studies do not settle that last question or establish a provider's motives." }, "d4": { "text": "Open questions include how to test learning from limited examples, causal reasoning and revision of plans under changed goals while ruling out task-specific shortcuts. Reasoning-trace faithfulness asks whether a written explanation reports factors that actually influenced an output. Human-like interfaces raise separate questions about trust and responsibility. Another question is how much of human cognition written explanations capture when those explanations become training material. The cited studies do not settle that last question or establish universal conclusions about providers' motives." } } } ], "note": "Depth changes the wording, examples and drawing detail. The claims and their sources stay the same at every depth; the default text is the At work depth.", "defaultDepth": "d2", "depths": [ { "id": "d1", "label": "First look", "audience": "For readers from about age 10 with no computing background. Short sentences and everyday words." }, { "id": "d2", "label": "Look closer", "audience": "For secondary school and curious adults. Introduces tokens, scores, training, sampling and seeds." }, { "id": "d3", "label": "At work", "audience": "For people deciding whether to rely on an AI tool. Adds context, retrieval, tools and review steps." }, { "id": "d4", "label": "In depth", "audience": "For developers, researchers and students of machine learning. Uses the technical terms and links primary sources." } ] }, "practice": { "title": "Before you rely on an AI tool", "intro": { "text": "Use these questions to examine what an AI application can do and how errors could be caught. We assembled them from the linked guidance and research; they are a starting point for review, not a safety certificate.", "sources": [ "ncsc-secure-ai", "nist-genai-profile" ] }, "sections": [ { "id": "practice-system", "title": "Look at the whole system", "text": "An AI application combines a model with other software. People configure its instructions, data sources, connected tools and access rights. These choices affect what it can do. A feature that drafts text needs a different review from one that can change records or send messages. Start by listing the application’s actual access.", "sources": [ "ncsc-secure-ai", "ncsc-agentic-risk" ], "question": "What can this application read, change or send, and through which tools?" }, { "id": "practice-task", "title": "Decide how you will check the result", "text": "Define a successful result and which mistakes would matter. Then consider whether fixed code, a model or a combination fits the job. Test generated code against its requirements. A convincing demonstration is a reason to investigate further; reliable use needs tests that reflect the intended task.", "sources": [ "ncsc-ai-design", "copilot-code-review", "nist-genai-profile" ], "question": "How will we check that it did the right job, including difficult cases?" }, { "id": "practice-access", "title": "Set limits before it takes action", "text": "Give the application only the tools and access it needs. Use permissions in the connected software to enforce those limits, with approval before consequential actions where appropriate. Instructions tell a model what you want; access controls determine what it can reach. Include connected services and their access keys in the review.", "sources": [ "owasp-excessive-agency", "ncsc-ai-design" ], "question": "Which actions need approval, and what enforces that requirement?" }, { "id": "practice-observe", "title": "Keep a useful record of what happened", "text": "Record relevant inputs, answers and tool activity while protecting sensitive information. These records help investigate problems. Written reasoning steps can add clues, but give an incomplete account of a model's internal calculations. In its March 2026 report, OpenAI describes monitoring reasoning and actions together and acknowledges uncertainty about missed incidents.", "sources": [ "ncsc-ai-operations", "circuit-tracing", "cot-monitorability", "coding-agent-monitoring" ], "question": "What evidence would reveal a problem, who reviews it, and how quickly can they respond?" }, { "id": "practice-evidence", "title": "Test how things could go wrong", "text": "A document or website can contain instructions that redirect the model: prompt injection. Test the complete application with checks on input and output, limited permissions, separation from sensitive systems and a response plan. Say what each test covered. A successful example or an empty alert log leaves other possible failures untested.", "sources": [ "owasp-prompt-injection", "ncsc-secure-ai", "nist-genai-profile", "coding-agent-monitoring" ], "question": "What exactly was tested, under which conditions, and what conclusion does that evidence support?" } ] }, "routes": [ { "id": "understand-debate", "title": "Understand the debate", "description": "A short route through the major positions and ethical ideas.", "terms": [ "eacc", "decel", "doomer", "ea", "longtermism", "dacc", "normal-technology" ] }, { "id": "internet-lore", "title": "Unpack the internet lore", "description": "Meet the community, then the puzzles behind the famous story.", "terms": [ "lesswrong", "rationality", "basilisk", "newcomb", "decision-theory", "acausal" ] }, { "id": "ai-arguments", "title": "Go into the AI arguments", "description": "Build the vocabulary for reading technical safety claims.", "terms": [ "agi", "asi", "safety", "orthogonality", "convergence", "mesa", "reward-hacking" ] }, { "id": "assess-application", "title": "Assess an AI application", "description": "Follow the practical questions from tools and access to monitoring and evidence.", "terms": [ "agents", "least-privilege", "prompt-injection", "observability", "ai-control", "interpretability", "safety" ] }, { "id": "ai-basics", "title": "AI is software: start with the basics", "description": "Start with computer programs, then explore learned models and the applications that use them.", "terms": [ "software", "ai", "symbolic-ai", "ai-system", "machine-learning", "model", "foundation-models", "generative-ai", "llm", "token", "prompt", "hallucination" ] }, { "id": "model-mechanics", "title": "Look inside a language model", "description": "Follow the pieces from learned weights and training to text generation.", "terms": [ "neural-network", "parameters", "training", "inference", "transformer", "embedding", "temperature", "reasoning-models" ] }, { "id": "use-ai", "title": "Understand the features in an AI tool", "description": "Learn what changes when an app uses documents, images, extra training or connected tools.", "terms": [ "context-window", "rag", "multimodal", "fine-tuning", "rlhf", "tools", "agents", "evaluation" ] }, { "id": "everyday-ai", "title": "Make sense of everyday AI questions", "description": "Explore generated media, jobs, claims about feelings and the rules around AI.", "terms": [ "synthetic-content", "deepfake", "ai-slop", "automation", "ai-consciousness", "eu-ai-act" ] }, { "id": "product-names", "title": "Untangle the product names", "description": "Separate the app you use, the model behind it and the organization providing it.", "terms": [ "chatbot", "model", "chatgpt", "claude", "gemini", "grok", "llama" ] }, { "id": "does-it-understand", "title": "Does it understand?", "description": "Explore intelligence, meaning, human-like language and what the evidence leaves open.", "terms": [ "intelligence", "understanding", "grounding", "reasoning-models", "anthropomorphism", "ai-consciousness" ] }, { "id": "training-process", "title": "How a model is trained", "description": "Follow training examples, learning methods and the ways a model can fit them poorly.", "terms": [ "algorithm", "training-data", "data-work", "pretraining", "self-supervised-learning", "loss-function", "gradient-descent", "overfitting", "post-training" ] }, { "id": "read-ai-tests", "title": "Read an AI test result", "description": "Ask what was tested, whether the examples were new, and how well the result carries over.", "terms": [ "evaluation", "benchmark", "data-contamination", "generalization", "calibration", "robustness", "red-teaming", "human-in-the-loop" ] }, { "id": "beyond-chatbots", "title": "Explore AI beyond chatbots", "description": "Look at images, voices and generated data, then follow how their origins can be documented.", "terms": [ "computer-vision", "speech-recognition", "speech-synthesis", "diffusion-models", "latent-space", "synthetic-data", "data-provenance", "watermarking" ] }, { "id": "question-ai-progress", "title": "Question claims about AI progress", "description": "Separate task performance, model size and training data when reading claims about future capabilities.", "terms": [ "narrow-ai", "agi", "evaluation", "compute", "scaling-laws", "synthetic-data", "model-collapse" ] } ], "timeline": [ { "id": "vinge-singularity", "date": "1993", "title": "A singularity scenario", "text": "Vernor Vinge's essay argues that greater-than-human intelligence could transform technological change and the limits of prediction.", "sources": [ "vinge-singularity" ], "terms": [ "singularity", "transhumanism" ] }, { "id": "existential-risk-papers", "date": "2002–2003", "title": "Existential risk and advanced AI", "text": "Bostrom develops an existential-risk taxonomy and discusses advanced AI goals, including a paperclip example, in separate papers.", "sources": [ "existential-risks", "advanced-ai-ethics" ], "terms": [ "xrisk", "paperclip" ] }, { "id": "lesswrong-origins", "date": "2006–2009", "title": "From Overcoming Bias to LessWrong", "text": "LessWrong's own history traces the community from the Overcoming Bias blog in 2006 to a separate site in 2009, seeded with the Sequences.", "sources": [ "lw-history" ], "terms": [ "lesswrong", "sequences" ] }, { "id": "basilisk-controversy", "date": "2010", "title": "The basilisk controversy", "text": "A LessWrong user named Roko posts the disputed thought experiment. The community retrospective describes rejection of the argument and a subsequent discussion ban.", "sources": [ "basilisk-history" ], "terms": [ "basilisk", "acausal" ] }, { "id": "ea-name", "date": "2011", "title": "Effective altruism gets its name", "text": "The EA introduction dates the term's coinage to naming the Centre for Effective Altruism in Oxford. The project covers several causes, not just AI.", "sources": [ "ea-definition" ], "terms": [ "ea", "utilitarianism" ] }, { "id": "safety-research", "date": "2016–2019", "title": "More concrete safety questions", "text": "Concrete Problems in AI Safety examines accident risks in 2016. The 2019 learned-optimization paper introduces mesa-optimization as a research question.", "sources": [ "concrete-safety", "learned-optimization" ], "terms": [ "reward-hacking", "mesa" ] }, { "id": "eacc-outlook", "date": "2022", "title": "e/acc lays out its outlook", "text": "The e/acc newsletter publishes its growth-focused self-description and reposts early principles opposing centralized deceleration.", "sources": [ "eacc-definition", "eacc-tenets" ], "terms": [ "eacc", "decel" ] }, { "id": "pause-and-dacc", "date": "2023", "title": "Pause and acceleration proposals", "text": "FLI's open letter requests a six-month frontier pause. Later that year, Buterin's techno-optimism essay introduces d/acc as a selective, defensive approach.", "sources": [ "pause-letter-2023", "dacc-original" ], "terms": [ "pause", "dacc" ] }, { "id": "evolving-proposals", "date": "2025–2026", "title": "Proposals keep evolving", "text": "Buterin revisits d/acc in January 2025. PauseAI's April 2026 proposal sets out international safety and democratic-control conditions.", "sources": [ "dacc-update", "pauseai-proposal-2026" ], "terms": [ "dacc", "pause" ] } ], "connections": [ { "from": "lesswrong", "to": "ea", "text": "One is a discussion community; the other is a project for finding effective ways to help. Their shared interest in reasoning does not make them the same institution or philosophy.", "kind": "interpretation", "sources": [ "lw-welcome", "ea-definition" ] }, { "from": "ea", "to": "longtermism", "text": "EA spans causes such as health, animals and catastrophic risks. Longtermism adds a particular emphasis on future generations.", "kind": "synthesis", "sources": [ "ea-definition", "longtermism-definition" ] }, { "from": "basilisk", "to": "decision-theory", "text": "The basilisk tries to turn unusual decision-theory assumptions into an argument about future threats. The formal theories do not themselves establish that scenario.", "kind": "interpretation", "sources": [ "basilisk-response", "tdt-paper" ] }, { "from": "eacc", "to": "dacc", "text": "Both use the language of building and progress. Buterin's d/acc makes defensive benefits and distributed power explicit selection criteria.", "kind": "synthesis", "sources": [ "eacc-definition", "dacc-update" ] }, { "from": "orthogonality", "to": "convergence", "text": "Different final goals can coexist with similar useful intermediate goals. These are complementary theses rather than contradictory claims.", "kind": "synthesis", "sources": [ "superintelligent-will" ] }, { "from": "ai-ethics", "to": "xrisk", "text": "Rights and present harms and long-run catastrophe are different dimensions of concern; this map cannot compress all of them into one risk score.", "kind": "interpretation", "sources": [ "unesco-ethics", "existential-risks" ] }, { "from": "training", "to": "inference", "text": "Training adjusts learned settings. Inference uses the trained model to produce an output. They are different stages of working with a model.", "kind": "synthesis", "sources": [ "google-gradient-descent", "google-ml-glossary" ] }, { "from": "prompt", "to": "fine-tuning", "text": "A prompt supplies instructions and examples as input. Fine-tuning changes learned parameters through additional training.", "kind": "synthesis", "sources": [ "google-llm-tuning" ] }, { "from": "rag", "to": "hallucination", "text": "Retrieval supplies material that can support an answer. Check the generated claims against that material, since supplying sources does not by itself verify the result.", "kind": "synthesis", "sources": [ "rag-paper", "nist-genai-profile" ] }, { "from": "tools", "to": "agents", "text": "Tools provide operations an application can carry out. An agent can use a model to choose and repeat steps involving those tools, subject to the application's access controls.", "kind": "synthesis", "sources": [ "hf-tool-use", "owasp-excessive-agency" ] }, { "from": "alignment", "to": "safety", "text": "Alignment concerns intended behavior and goals. Safety also requires examining harms, failures and the environment in which a system operates.", "kind": "synthesis", "sources": [ "instructgpt-paper", "concrete-safety", "ncsc-secure-ai" ] }, { "from": "chatbot", "to": "llm", "text": "The chatbot is the application you interact with. A language model can generate its replies, while the surrounding software manages context and tools.", "kind": "synthesis", "sources": [ "gemini-overview", "hf-tool-use" ] }, { "from": "ai-governance", "to": "eu-ai-act", "text": "Governance includes many ways of assigning responsibility and oversight. The EU AI Act is one legal framework within that wider topic.", "kind": "synthesis", "sources": [ "oecd-ai-principles", "eu-ai-act-overview" ] }, { "to": "grounding", "kind": "interpretation", "text": "The form-versus-meaning debate asks how language connects to what it refers to.", "sources": [ "understanding-bender-koller" ], "from": "understanding" }, { "to": "agents", "kind": "interpretation", "text": "An agent can have permission to use tools. Human-like wording does not establish human feelings or transfer responsibility away from its operators.", "sources": [ "understanding-anthropomorphism" ], "from": "anthropomorphism" }, { "from": "evaluation", "to": "benchmark", "text": "Evaluation is the wider process of checking a system. A benchmark is one reusable test within that process.", "kind": "synthesis", "sources": [ "helm-paper" ] }, { "from": "reasoning-models", "to": "chain-of-thought", "text": "A reasoning model may generate intermediate steps. Those steps can help with a task without reliably explaining how the answer was produced.", "kind": "synthesis", "sources": [ "deepseek-r1-paper", "glossary-eval-unfaithful-cot", "understanding-faithfulness" ] } ], "development": { "reviewedOn": "2026-09-15", "events": [ { "id": "turing-imitation-game", "date": "1950", "title": "Turing asks how to test machine conversation", "category": "foundations", "summary": "Alan Turing proposes an imitation game to make questions about machine intelligence more concrete.", "text": "In Computing Machinery and Intelligence, Turing describes a judge questioning hidden participants through written messages, then asks what would happen if a computer took one participant's place. He also discusses computers that learn.", "limitation": "This is a proposed test and a philosophical argument. The paper does not report a computer passing it, and Turing discusses objections to using the game as a test of thinking.", "sources": [ "development-turing" ], "terms": [ "ai", "evaluation", "machine-learning" ] }, { "id": "dartmouth-workshop", "date": "1956", "title": "Researchers gather under the name artificial intelligence", "category": "foundations", "summary": "The Dartmouth summer project brings researchers together to study artificial intelligence.", "text": "The proposal, written in 1955, uses the name artificial intelligence and sets out questions about language, learning, neural networks and problem-solving. Dartmouth's history records the gathering in the summer of 1956.", "limitation": "The proposal calls its central idea a conjecture: that aspects of intelligence can be described precisely enough for a machine to simulate them. It sets a research agenda, rather than reporting that this had been achieved.", "sources": [ "development-dartmouth-proposal", "development-dartmouth-history" ], "terms": [ "ai", "neural-network" ] }, { "id": "eliza-conversation", "date": "1966-01", "title": "ELIZA generates replies using written rules", "category": "language", "summary": "Joseph Weizenbaum describes ELIZA, a program that responds to typed conversation by rearranging text.", "text": "ELIZA looks for keywords, matches patterns and uses a script to assemble a reply. The paper explains how a sentence about being unhappy can be turned into a follow-up question using these rules.", "limitation": "Weizenbaum shows how the transformations can work without understanding the meaning of the text being rearranged. A reply that fits a conversation can give a misleading impression of what the program knows.", "sources": [ "development-eliza" ], "terms": [ "ai", "llm" ] }, { "id": "backpropagation-representations", "date": "1986-10-09", "title": "A learning method adjusts a neural network's inner layers", "category": "learning", "summary": "Rumelhart, Hinton and Williams show how backpropagation can help neural networks learn useful internal patterns.", "text": "The method works backward from the difference between the desired answer and the network's answer. It calculates how to adjust the connections inside the network, allowing its intermediate layers to learn features useful for a task.", "limitation": "The setup needs a specified task and desired outputs to compare against. Reducing that error is the learning objective; the method does not choose the purpose of the system.", "sources": [ "development-backpropagation" ], "terms": [ "neural-network", "training", "parameters" ] }, { "id": "deep-blue-chess", "date": "1997-05", "title": "Deep Blue wins its chess rematch with Kasparov", "category": "systems", "summary": "IBM's Deep Blue defeats reigning world chess champion Garry Kasparov in a six-game match.", "text": "IBM records a 3.5–2.5 result under standard tournament time controls. The system combines fast searches through possible chess positions with chess-specific evaluation, databases and advice from grandmasters.", "limitation": "The result measures performance in chess under the match rules. The computer and its preparation were designed around that task; the match did not test conversation or everyday problem-solving.", "sources": [ "development-deep-blue" ], "terms": [ "ai", "evaluation" ] }, { "id": "imagenet-dataset", "date": "2009", "title": "ImageNet organizes millions of labeled pictures", "category": "foundations", "summary": "The ImageNet paper introduces a large collection of pictures for training and testing image-recognition systems.", "text": "The researchers gather web images, organize them using WordNet, a catalogue of word meanings, and ask people to check the labels. The 2009 paper reports 3.2 million images across 5,247 categories in the collection available at that point.", "limitation": "The collection was still being built. Most of the paper's analysis focused on animal and vehicle categories, so those findings do not describe every kind of image or recognition task.", "sources": [ "development-imagenet" ], "terms": [ "training", "evaluation", "machine-learning" ] }, { "id": "alexnet-image-recognition", "date": "2012", "title": "A deep neural network improves image classification", "category": "learning", "summary": "Krizhevsky, Sutskever and Hinton report a winning ImageNet competition entry built with deep neural networks.", "text": "Their approach trains on labeled pictures using graphics processors, or GPUs, to speed up the calculations. It learns image features through several layers instead of relying only on features specified by hand.", "limitation": "The competition asks for labels from a fixed set of image categories. The authors also note that even ImageNet cannot capture the full variety of object recognition in the world.", "sources": [ "development-alexnet" ], "terms": [ "neural-network", "training", "evaluation" ] }, { "id": "alphago-lee-sedol", "date": "2016-03", "title": "AlphaGo beats Lee Sedol at Go", "category": "systems", "summary": "DeepMind's AlphaGo wins four of five games against Go player Lee Sedol.", "text": "DeepMind describes a system combining neural networks with a search through possible moves. It first learns from expert games, then improves by playing against versions of itself and learning from the results.", "limitation": "The result concerns the board game Go. Its training examples, possible moves and win-or-lose feedback are specific to that setting; the match did not test an unrestricted range of human tasks.", "sources": [ "development-alphago" ], "terms": [ "neural-network", "training", "evaluation" ] }, { "id": "transformer-architecture", "date": "2017-06-12", "title": "The Transformer offers a different way to process text", "category": "language", "summary": "Attention Is All You Need introduces the Transformer, a neural-network design that relates different parts of a text.", "text": "Its attention calculations let the model use information from other positions in a sequence. The authors report improved translation results and a design that allows more training calculations to happen in parallel.", "limitation": "The original experiments cover English–German and English–French translation, plus a task that identifies sentence structure. They establish results for those tasks and conditions, rather than for every later Transformer application.", "sources": [ "transformer-paper" ], "terms": [ "transformer", "training", "llm" ] }, { "id": "gender-shades-audit", "date": "2018-02", "title": "An audit finds uneven errors across groups", "category": "systems", "summary": "Joy Buolamwini and Timnit Gebru find different error rates across skin-type and gender groups in three commercial face-analysis systems.", "text": "The Gender Shades study tests systems that assign gender labels to photographs. Using a dataset balanced by gender and skin type, the researchers find the most errors for darker-skinned women. The study examines performance separately for different groups.", "limitation": "These results concern three systems and the dataset used in the 2018 study. They do not measure every face-analysis system, later versions or every aspect of fairness.", "sources": [ "development-gender-shades" ], "terms": [ "bias", "evaluation", "ai-ethics" ] }, { "id": "bert-pretraining", "date": "2018-10-11", "title": "BERT learns from text before adapting to particular tasks", "category": "language", "summary": "BERT uses text on both sides of a word to learn numerical representations of text that can be adapted for language tasks.", "text": "The authors train a Transformer-based model on text, then adjust it for tasks such as answering questions. They report improved results across eleven language-processing tasks.", "limitation": "The reported task results use additional training for each task. BERT's ability to learn useful text representations is distinct from a ready-to-use conversational assistant.", "sources": [ "development-bert" ], "terms": [ "transformer", "training", "fine-tuning" ] }, { "id": "rag-retrieval-generation", "date": "2020-05-22", "title": "RAG combines finding passages with writing an answer", "category": "systems", "summary": "A research paper combines a language generator with a system that retrieves relevant Wikipedia passages.", "text": "The authors call the approach retrieval-augmented generation, or RAG. Their model uses both what it learned during training and passages found in a separate index when producing an answer.", "limitation": "The paper reports improvements on selected tests, not perfect factuality. Its authors warn that Wikipedia and other external sources can contain errors and bias.", "sources": [ "rag-paper" ], "terms": [ "rag", "llm", "hallucination" ] }, { "id": "gpt3-few-shot", "date": "2020-05-28", "title": "GPT-3 performs tasks from instructions and examples", "category": "language", "summary": "OpenAI's GPT-3 paper studies how a large language model can attempt new tasks using examples placed in its input.", "text": "The researchers give the model instructions and a few demonstrations in text, without changing its learned settings for each task. They report results on translation, question answering and other language tests.", "limitation": "The paper identifies tasks where this approach still struggles. It also checks overlap between web training data and test material, flagging some results while finding little effect on most of the tests examined.", "sources": [ "development-gpt3" ], "terms": [ "llm", "prompt", "context-window", "evaluation" ] }, { "id": "instructgpt-human-feedback", "date": "2022-03-04", "title": "Human feedback helps tune instruction-following", "category": "learning", "summary": "OpenAI's InstructGPT paper describes using people's example answers and preferences to tune a language model.", "text": "People first demonstrate desired responses, then rank alternative model outputs. The researchers use that feedback in further training and report that their evaluators preferred InstructGPT's answers to those from the underlying GPT-3 model.", "limitation": "The preference result applies to the evaluators and prompts used in the study. The paper says InstructGPT still makes simple mistakes; being preferred by a reviewer does not establish that an answer is correct.", "sources": [ "instructgpt-paper" ], "terms": [ "fine-tuning", "rlhf", "evaluation" ] }, { "id": "helm-evaluation", "date": "2022-11-16", "title": "HELM compares more than answer accuracy", "category": "foundations", "summary": "Stanford's HELM project evaluates language models across several uses and measures, including fairness and efficiency.", "text": "The researchers use common test conditions and examine accuracy alongside reliability, bias, harmful language and other measures. They publish prompts and outputs so others can inspect the results.", "limitation": "HELM's authors explicitly identify gaps in what they test, including underrepresented English dialects and some measures of trustworthiness. A broad evaluation still has boundaries.", "sources": [ "helm-paper" ], "terms": [ "evaluation", "bias", "llm" ] }, { "id": "chatgpt-public-preview", "date": "2022-11-30", "title": "ChatGPT opens as a public conversation interface", "category": "systems", "summary": "OpenAI introduces ChatGPT as a research preview that people can question through a conversation.", "text": "The announcement describes a model trained with example conversations and human feedback. Follow-up questions become part of the interaction, rather than each request being treated only as an isolated text completion.", "limitation": "OpenAI's launch announcement warns about convincing but incorrect answers, sensitivity to wording and excessive verbosity. These are limitations reported for the version introduced in 2022.", "sources": [ "development-chatgpt" ], "terms": [ "llm", "prompt", "hallucination", "rlhf" ] }, { "id": "gpt4-multimodal-report", "date": "2023-03-15", "title": "GPT-4's report describes image and text inputs", "category": "language", "summary": "OpenAI's GPT-4 technical report describes a model that accepts pictures and text and produces text.", "text": "The report presents tests of this multimodal model on language tasks, images and exams. It describes next-token prediction followed by further training with human feedback.", "limitation": "OpenAI says the model remains less capable than people in many real-world situations. The report also withholds details such as model size and training-data construction, limiting what readers can independently reconstruct.", "sources": [ "development-gpt4" ], "terms": [ "multimodal", "llm", "evaluation" ] }, { "id": "alphafold3-molecules", "date": "2024-05-08", "title": "AlphaFold 3 predicts structures of interacting molecules", "category": "systems", "summary": "AlphaFold 3's authors describe a model for predicting how proteins and other biological molecules fit together.", "text": "The paper covers combinations that include proteins, DNA or RNA, and small molecules. The researchers report improved structure predictions over several specialized tools on the comparisons they test.", "limitation": "The paper documents mistakes such as overlapping atoms and incorrect molecular geometry. It also describes limits in representing the different shapes a molecule can take.", "sources": [ "development-alphafold3" ], "terms": [ "model", "neural-network", "evaluation" ] }, { "id": "deepseek-r1-reasoning", "date": "2025-01-22", "title": "DeepSeek describes training aimed at reasoning tasks", "category": "learning", "summary": "DeepSeek's R1 paper reports using reward-based training to improve performance on reasoning tests.", "text": "The January 2025 paper describes R1-Zero and R1, with different training setups. R1 combines example data and reinforcement learning; the team reports results on mathematics, coding and other tasks, as well as training smaller models from R1-generated examples.", "limitation": "These are the team's reported results. The same paper says R1 trails its earlier V3 model on some tool-use and multi-turn tasks, and can mix languages or respond poorly to certain prompts.", "sources": [ "development-deepseek-r1" ], "terms": [ "reasoning-models", "training", "evaluation", "tools" ] } ] } } }