Fidelity check: M9, The AI moment and the industry#

Working file. Checks working/maynard-lens/M9-the-ai-moment-and-the-industry.md (version of 26 September 2026) for fidelity to Maynard’s own texts. Line numbers (L) refer to that file. Points about Huang are included only where they change how Maynard is read; the fairness check covers Huang.

Method#

• Every quotation attributed to Maynard in the public part of M9 (about 120) was matched by script against the corpus posts (working/maynard/corpus/), the text of every PDF in Resources/maynard-papers/ (extracted afresh, with page markers; AI and the Art of Being Human files excluded), the web and paper .md files, the Films from the Future page text (working/maynard/book/FFTF-*.txt), Maynard stuff/substack article 1 draft 3.md (the Series introduction) and S1–S7. All but one were found verbatim. The context was then read for every quotation that carries an argument, and page numbers were checked for the papers.
• Labels were checked against the sources, the map (05: §1 provenance, C1–C18, §8 tensions and gaps) and the earlier M8 fidelity check, which found several of the same problems.
• Also checked: excluded and AI-written text, [mixed] handling, co-authorship weighting, dates, the two September 2026 clarifications, and proportion.

Verdict#

M9 is accurate at the level of wording. Its quotations are exact, its dates are right, and it treats provenance with care. It does not use AI and the Art of Being Human in any form (see issue 14 for three near-misses that it handles correctly). It uses no AI-generated text as evidence, and it weights Maynard & Garbee (2019-08-13) as fully his. Both September 2026 clarifications are reflected faithfully (§2.4, D3, D7, §6 item 3). The INTERNAL section is at the end.

The problems are in how evidence is weighted and labelled, and most of them lean the same way: they bring Maynard closer to Huang, or to a tidier position, than his texts allow. - Two quotations are cut so that they tilt him towards Huang (issue 1). - [Stated] is attached to alignments and applications he has never stated (issue 2). - Five sources are used outside their context (issue 3). - The “AI working as designed” / “narrow aperture” thesis is drawn more narrowly than his risk landscape, and it leaves out his own concessions to the frameworks (issue 4). - Several directly relevant statements of his are missing. The most important is his own description of Huang’s depiction of AI in the Series introduction (issue 5). - D4 settles a tension in his record in one direction only (issue 6).

Most fixes are a relabel, a restored clause or an added sentence. Issues 4 and 5 need a paragraph each.

Ranked issues#

1. HIGH: Two truncated quotations tilt him towards Huang (§5 L188; A4 L109)#

(a) The costs of alarm (§5, last bullet, L188). “Forgone innovation and false alarms have victims (Testimony 2006; FFTF p.163; NN 2014-03 p.160), as Huang’s radiology example shows [Stated].” - Source. NN 2014-03 p.160 praises the RS–RAE report as “a voice of scientific reason and social concern at a time when speculation could have scuppered the nanotechnology enterprise or, worse, led to materials and products that showed a blatant disregard for health and environmental risks.” The map (§2 “It counts both sides”; C7; tension 16) records that he ranks disregard of risk as the worse outcome. M8’s fidelity check (issue 1) raised the same point. - Fix. “…have victims (Testimony 2006 p.52; FFTF p.163) [Stated]. He does not weigh the two equally, though: in 2014 he judged products showing ‘a blatant disregard for health and environmental risks’ a worse outcome than an industry scuppered by speculation (NN 2014-03 p.160) [Stated]. That Huang’s radiology case illustrates the point is [Implied].”

(b) Alignment research (A4, L109). “…and wrote in 2026 that alignment research ‘deserves the attention it’s getting’ (FWB 2026) [Stated].” - Source. FWB 2026 continues: “The technical challenges are real and important. The alignment problem deserves the attention it’s getting. But underneath all of it is a challenge that I suspect matters more than any of the technical ones: figuring out what kind of future we actually want…” - Fix. Quote the “But…” sentence. Also add that 2024-06-20 (“always a social and political endeavor as well as an engineering challenge”, already in §2.4) qualifies the alignment. This matters because A4 is used in the Summary (L20) as agreement “on treating safety spending as serious engineering”.

2. HIGH: [Stated] attached to alignments and applications he has not stated (systemic)#

§8 of the map and M9’s own §2.8 say, rightly, that he has written almost nothing on the July–September events and nothing on Huang beyond the Series introduction. Several labels contradict this: the component quotation is his, but the relation to Huang, the labs or the event is the analysis’s.

3. HIGH–MEDIUM: Five sources used outside their context#

(a) 2025-05-04 (§2.4 L57; D3 L123). Cited [Stated] under “tests have limits”, and paraphrased in D3 as “simulated environments leave real-world risk unclear”. - Source. “Simulated environments” is Kasirzadeh and Gabriel’s category of operating environment for agents (beside mediated and physical). His sentence is a doubt about their efficacy scale: “I’m not fully convinced that the proposed efficacy scale has it right yet, as it’s unclear how the risks of an AI working within a simulated environment (the ‘safest’ type of environment) might be realized”. It says nothing about testing, and the D3 paraphrase reverses his direction (he is unsure how risk arises in a simulated environment, not how simulation predicts the real world). M8’s check (issue 6) found the same misuse. - Fix. Delete it from §2.4 and D3. Testimony 2007 p.21, NN 2015-06 p.483 and the September 2026 clarification carry the point. If it is kept, gloss it accurately, “he doubted that a framework rating simulated environments ‘safest’ had captured how risk arises there”, and label it [Inferred, low] for the July incident (a “supposedly isolated sandbox”, 2026-09-24). That reading is at least on topic.

(b) 2023-11-18, “a chance to take a breath (a pause even)” (§3.1 L92; §5 L185). Offered as “Costly unilateral action is what he has asked for”, and in §5 as pauses “which his work asked for”. - Source. He speculates that the OpenAI board crisis “may be a good thing — a reset from the hype of the past year, and a chance to take a breath (a pause even) and rethink and recalibrate the relationship between AI and society”. It is a hope that disruption might give society a pause, not a request for firms to pause at their own cost. His record on pauses is mixed: - he declined the 2023 pause letter (2023-04-04); - he called for “pausing — or even rethinking” emotion-exploiting companion bots (2024-10-27); - he said in 2026, “We can’t pause it” (2026-09-24 [mixed]). - Better evidence. His paper treats Anthropic’s move from an “unconditional” pause to one “conditioned on what competitors do” as a softening (2026-07-16 [mixed], arXiv p.8). That implies he values unconditional pause commitments. The 2025-08-31 call for “working harder on safety checks and protocols before releases” can stay. - Fix. Replace “what he has asked for” with “consistent with his paper’s treatment of unconditional pause commitments as the stronger form (2026-07-16 [mixed]) and his call for more safety work before release (2025-08-31)”. Cite 2024-10-27 as his one explicit call to pause a class of products, note the 2023 decline and “We can’t pause it”, and label [Implied, medium].

(c) 2025-07-23, open weights (§2.7 L74; §3.1 L97; A7 L115). The text says he “listed … support for open-weight models among its better elements” and speaks of “his” “worry about ‘corporate control’ of closed models”. - Source. “On the other hand, the Action Plan promotes open-source and open-weight models — something that will be welcomed by many who worry about the corporate control…”. He then adds that researchers will applaud it, and that these are open models “with strings attached”. The worry is attributed to “many”; the framing (“On the other hand”) is mildly positive, but it is reportorial. - Fix. “He noted, in a positive aside, that its open-weight support ‘will be welcomed by many who worry about the corporate control’ of closed models, adding ‘strings attached’ (2025-07-23) [Stated]; whether he shares the worry is [Implied, medium].” In the §3.1 open-weights row, attribute “corporate control” to those he describes. For balance, §2.7 should add that he gave input to Howard’s pro-openness paper (“myself included”, 2023-07-12).

(d) 2025-08-31, “better-managed, but probably not eliminated entirely” (§2.5 L64; D2 L121; §3.1 L92). Presented generically (“Harmful behaviour that emerges from a model’s properties…”) and applied to agentic intrusion as [Stated]. - Source. The sentence is about “The alleged behavior of ChatGPT that led to Adam Raine’s death”, an emergent property of a conversational model, contrasted with apps “intentionally designed” to exploit cognitive biases. - Fix. Name the context in §2.5. Label the transfer to July’s agents [Implied, medium]; it is a reasonable structural transfer (emergent behaviour managed, not eliminated), but it is not his statement about agents. Stronger [Stated] support for D2 exists: see issue 5b and 5f.

(e) 2024-05-15, “our anthropomorphizing cognitive biases” (§5 L180). Used to show that Huang’s objection to human words for software processes “fits” Maynard. - Source. The phrase defines “hyper-anthropomorphism — a concerted effort to create AI’s that are intentionally designed to engage our anthropomorphizing cognitive biases”, i.e. products “designed to make us fall a little in love with them”. It concerns designed intimacy, not how critics describe software. Huang’s target (“Spawn, create, kill…”, “A collection of people want to make the software more than it is”) is discourse. Maynard also keeps goal-directed language himself (issue 7). M8’s check (issue 11) found the same problem. - Fix. Narrow it to “both resist mystification”, note that his concern is designed anthropomorphism and its effects on users, which Huang’s framing does not address, and relabel [Inferred, low–medium].

4. HIGH–MEDIUM: “Harm from AI working as designed” and “narrow aperture” drawn more narrowly than his record, and his concessions to the frameworks omitted (D5 L127; §3.4 L140; §4.3 L171; Summary L20, L22)#

Text. D5: “Maynard’s most plausible AI harms come from AI working as designed, through fluency and relationship (Trojan 2026 p.1; Harness 2026 p.8) … [Stated]”. §3.4 row 2 and §4.3 build the main proxy finding on this.

Sources. - Trojan 2026 p.1 is a scope statement, not a ranking: “the analysis focuses on AI systems designed to be genuinely useful; the distinct challenges posed by intentional use of AI for manipulation … while important, fall outside the present scope.” His [Stated] plausibility ranking (FFTF p.159) sets manipulation above superintelligence. It does not set as-designed harm above failure. - His landscape is plural and includes failure-type risks. The week before the interview he reaffirmed as “amongst the top longer term (and more insidious) risks associated with frontier models” value misalignment, “Machines that alter their own instructions”, “smart-dumb decisions” and lethal autonomous weapons. He named cybersecurity and “governance of frontier AI models and systems” among the risks that have risen (2026-09-15). His own reading of the July incident is a loss-of-control reading (2026-09-24 [mixed]). The Series introduction’s reservation is about a missing “broader landscape of emergent AI characteristics, capabilities, threats, risks, and benefits”, which is broader than the cognitive thread. - His paper concedes the frameworks’ logic. “To be fair to the frameworks’ designers, there is a case for focusing on a narrow but deep risk layer … Concentrating limited safety resources on risks perceived to have the highest levels of severity makes sense; a framework that tried to manage sixteen hundred risks would end up managing none of them. And the frameworks’ architects have, sensibly, never claimed completeness.” Capability thresholds are “reasonable engineering judgment”, and “Severity floors make sense as triage” (2026-07-16 [mixed], arXiv pp.5, 7–8). None of this appears in D5, §3.4 or §5.

Fix. - D5: “The AI risk he has developed most fully, a threat to how people think and trust, operates, on his 2026 account, through AI working as intended (Trojan 2026 p.1; Harness 2026 p.8) and accumulates below catastrophe thresholds (2026-07-16 [mixed]) [Stated]. A containment-and-release model does not see it [Implied, high]. His landscape also includes failure and agency risks (2026-09-15), on which Huang’s containment model does engage [Stated].” - §3.4 row 2 and §4.3: add his concession (narrow-but-deep as defensible triage) as [Stated, mixed]. Say that his objection is to the missing second layer, not to the narrowness of the first (“a single safety layer is currently being asked to effectively stand in for two”, arXiv p.18). - §5: add a bullet, “Capability thresholds as defensible triage”, [Stated, mixed].

5. MEDIUM–HIGH: Directly relevant statements of his are missing#

Adding these would turn several [Implied] points into [Stated] ones, or balance them.

• (a) The Series introduction on Huang’s depiction of AI (D1, D2, §2.8, §4.3). It is his only substantive statement about Huang’s view of AI: the analysis “does treat AI largely as it is depicted by Huang — a technology that has been designed and engineered like any other, and so is subject to the same management and control approaches and methods as any other”. He also says he is “not sure I fully agree” with the analysis “(and this includes Claude’s resulting article)”, and that it “struggled to apply conceptual rather than literal comparisons”. §2.8 quotes only his first reaction and self-check. Add this to §2.8, D1 and D2 as [Stated]; it is the strongest evidence for D1 and D2 in the record. It also bears on §4.3’s reading of LLH. (Citation caveat: issue 13.)
• (b) The lecture on whether the labs understood the incident (D2). “It’s got a lot of people worried, because they couldn’t work out how this happened”, and the model “escaped its supposedly isolated sandbox” (2026-09-24 [mixed], single source). This is the most direct counterpoint to Huang’s “I know they know what happened” [55:46]. It also fits K9 (§4.1). Note the single-source status and that OpenAI later published a technical report.
• (c) The lecture’s symmetric critique of speculation (D1, D7, A3). n.4: singularity, superintelligence and AGI speculation is “incredibly blinkered and naive”, and “Then there’s almost the inverse: the people who say, ‘There’s nothing new under the sun here; it’s all just going to go away.’ That’s not evidence-based either. It’s speculation, and it’s dangerous as well” (2026-09-24 [mixed]). This gives D7’s “applies to ‘0%’ as much as to 10%” a [Stated, mixed] basis. It also gives D1 his own objection to “nothing new” deflation. §6 item 8 quotes only the humility clause of this note.
• (d) System cards and the limits of their tests (§5, §3.1 Astra row, D3). 2024-09-01: OpenAI’s system-card approach “represents a sophisticated approach to assessing and addressing possible safety issues”, and n.1, “I very much appreciate OpenAI’s approach to publishing their system cards. It demonstrates the care they are taking internally”. But: “could this AI persona have a power of persuasion that far exceeds that of the simple test used in OpenAI’s system card?” This is [Stated] evidence both for the labs’ candour (§5, currently sourced to the [mixed] paper) and for the limits of tests (D3).
• (e) Nvidia’s other appearance (§2.7, §5). 2025-02-23 (evo-2-dna-ai): the Arc Institute’s Evo 2, built “in collaboration with NVIDIA”. He praises the team for “Rather smartly” leaving pathogen genomes out of the training data and red-teaming the model, then adds that “the domain of unexpected consequences … go way beyond harmful viruses”, and closes on “go fast and break things in the hope that someone else will clean up the mess”. It is his only other Nvidia-linked text and a [Stated] instance of the pattern M9 infers: builder-side engineering safeguards valued, but judged insufficient.
• (f) Emergent, uncontainable risk (D2, §3.4 row 1). 2025-07-23: “irresponsible (or simply unthinking) innovation is likely to lead to emergent risks that cannot easily be contained”, and checks and balances as “critical guardrails that help avoid triggering serious and irreversible failures” [Stated].
• (g) Commitments as assets (§5 L181, §3.1 Nvidia row). 2026-07-16 [mixed], arXiv p.14: commitments “are perhaps better understood as assets … and, like any assets, they can be spent”. This makes §5 bullet 2 [Stated, mixed] rather than [Implied].

6. MEDIUM: D4 and the Summary settle a tension in his record in one direction (L20, L125, L142)#

Text. D4 says his incentive field “agrees with the labs’ diagnosis rather than Huang’s [Stated]”. The Summary lists “whether competition drives drift” as a divergence from Huang.

Source. His only direct comment on the September pacing calls is 2026-09-15 n.3. It names the Coxon resignation and “calls from Dario Amodei and others in the industry for more caution”, and says: “it does flummox me a little as to why the people developing AI are the ones both saying they should go slower, and not doing so”. In substance that is close to Huang’s “That strikes me as odd” [53:36]. Against it stands his structural account: the incentive field ([mixed]; the map lists the “incentive field” analysis among concepts that may have originated with Fable) and its secure roots (2019-08-13; 2023-11-18; NN 2016-06 p.491; 2024-07-13).

Fix. - D4: present both halves as his and let neither absorb the other. His most recent direct reaction echoes Huang’s impatience [Stated]. His structural account explains the drift without bad faith and prescribes common rules rather than courage [Stated, mixed; roots Stated]. The divergence from Huang is over the remedy and the explanation, not over whether the labs’ position is odd [Implied]. - Summary (L20): “diverges … on whether competition, rather than courage, explains the drift and what would fix it”.

7. MEDIUM: A1 overstates his “deflation” and misplaces Huang’s remark (L103; Summary L20)#

• Huang’s [48:58] is about evaluation awareness, not the incident. It answers Klein’s account of Astra (“they know when they’re being tested”): “if you give it a constraint — meaning you watch it — it’ll go find another solution”. Pairing it with the incident in A1 (“The mechanism of the incident”) and in the Summary (“the mechanism behind the July incident”) misplaces it. The pairing with 2025-07-06 (reduced “degrees of freedom” lead to “bad behavior”) still works as a mechanism of constrained optimisation, but the heading should be “the mechanism of constrained optimisation”. The evaluation-awareness point belongs with D3.
• He keeps goal-directed language that Huang rejects. In 2025-07-06 n.1 he defined “motive” in answer to an AI reviewer’s charge of anthropomorphism, and defended the term: it “makes sense in this context to refer to the thing leading to the intentional action as a ‘motive’”. He also speaks of AIs developing “internal motives”. In 2026 agents use humans “as another cog in the machinery to achieve its ends” (2026-09-24 [mixed]). Huang: “I don’t think software’s relentless”; “it’s just on” [1:03:14]. So the shared ground is narrower than “Both reject the need for will”: both hold that no consciousness, sentience or AGI is needed [Stated for Maynard]. They differ on whether goal-directed language is apt [Stated]. §2.5’s gloss “to avoid anthropomorphism” should read “in reply to a charge of anthropomorphism, while defending the term”.

8. MEDIUM: The Summary misstates the [mixed] paper (L18)#

Text. “…with some set-aside risks reappearing only where law compels them [mixed]”.

Source. The paper documents a voluntary return: “Google DeepMind … moved the other way, adding an avowedly exploratory harmful-manipulation domain to its own framework in 2025”, which shows exclusion is “at least in some cases — a choice rather than a necessity”. It also documents Anthropic’s manipulation tiers, “made before any enforcement”. M9’s own §2.3 reports the first correctly.

Fix. “…reappearing mainly where law requires them, with Google DeepMind’s voluntary addition as the exception [mixed]”.

9. MEDIUM: His 2006–08 funding ask is equated with firm-internal evaluation compute (A4 L109; §5 L182)#

Text. A4: “asked Congress for a tenth of nanotechnology research spending for risk research (Testimony 2006–2008)”. §5: the tenfold prediction and the “flip” to evaluation “are the kind of investment he sought for risk research in 2006–08 [Implied, high]”.

Source. - The 10% ask is in Testimony 2007 and 2008. In 2006 he asked for a fixed sum, as M8’s check also found. - The ask was for “strategic risk-related research” directed by federal agencies with an oversight mandate, partly because industry “has an economic incentive to sell products” and its findings “might be considered suspect … if not supported by independent studies” (PEN 2006 pp.4–5, 32). He also proposed bodies funded jointly but operating “independently of the funders” (WEF 2008). - Firm-internal evaluation compute meets the scale condition, not the independence condition. - Huang’s tenfold figure is a forecast (“I wouldn’t be surprised if…”), not a demand.

Fix. Date the ask “Testimony 2007–08”. Relabel §5 [Inferred, medium], adding: “His 2006–08 ask was for independent, strategically directed research, so he would value the scale of the shift more than its location inside the firms.” A4 should say that it pairs a forecast with a demand.

10. MEDIUM: §6 items 4 and 5 present the analysis’s instruments as his direction (L200, L202)#

• Item 4. “Prefer common duties … (incident reporting, notification of affected third parties, liability that reaches internal evaluation)”, labelled [Implied; medium].
• None of the three instruments appears in his record. The map lists legal liability “beyond deepfakes” as little developed (§8, Gaps).
• His paper’s remedy is “consensus norms, rules, and costs that land on every organization at once”, plus disclosure of risk selection.
• Two of his own positions pull against a duties-first reading. In 2019 he held that top-down governance yields only “crude boundaries” in entrepreneurial cultures (2019-08-13). His Garbee lesson is “you do not hand it a compliance duty; you show it a threat to something it values”.
• Fix. Keep the direction [Implied, medium]. Mark the three instruments as the analysis’s examples [Inferred, low–medium], and note the tension with the Garbee lesson, which item 6 already develops.
• Item 5. “Keep state compliance duties until a federal equivalent exists”, labelled [Implied; medium], with 2025-07-23 as evidence.
• He has no stated view on pre-emption, and 2025-07-23 does not discuss state law.
• His paper cuts both ways: compliance brought manipulation back, but compliance coverage is “jurisdiction-bound and politically contingent”, and California’s statute “confines its mandated disclosures to catastrophic risk, narrowly defined”.
• Fix. Relabel [Inferred, medium]. Drop 2025-07-23 or replace it with its “checks and balances … critical guardrails” passage. INTERNAL Q5 already asks the right question.

11. MEDIUM: Proportion and provenance: the [mixed] paper carries the headline finding, and single-source items are not flagged (L7–8, L22, L140, L160, L164; §7 L221)#

• Weight of the paper. The frontier-AI paper [mixed] is cited 16 times by date and about 11 more as “his paper”. “Aperture” appears 14 times. The Summary’s proxy finding rests half on it (“a narrow aperture”). §7 acknowledges its prominence, but its secure antecedents are not cited where the concept is used:
• “The harder challenge is working out what we should be measuring” (NN 2015-06 p.483);
• community norms guiding “how risk is defined and evaluated” (NN 2015-09 pp.730–731);
• orphan risks as “those hard to quantify and easy to ignore risks that nevertheless have a habit of coming back to bite” (Nexus 2020);
• the 2018 orphan-risks article.

Fix. In §2.3 and §4.2, cite these first and the paper second. That anchors “aperture” in his own long-standing idea that what gets measured decides what gets managed. - Unflagged single-source items. The conventions (L8) promise that neither [mixed] text “carries a position alone unless marked ‘single source’”. Unflagged exceptions: - the four filters (§4.1 “[Inferred, high]”; possibly Fable-originated, per map §1); - the safety differential and aperture log (§4.2, §6 items 1–2); - “irrelevant to this conversation” (A1; lecture only); - the race-logic line (A6; §3.1 China row; lecture only); - “knowing that it’s speculation, not reality” (§6 item 8; corroborated by “These are explorations, not findings”, HNS 2026, which should be cited).

Flag each, or add the corroboration. - Balance of his thinking. Being human, the formative and cognitive thread, and justice appear only in passing (D5, D8, the Australia row). That is defensible for an industry-facing dimension, but §7’s list of what the reading “also rests on” should not imply more than the text delivers.

12. MEDIUM–LOW: A6 reads his race-logic criticism as support for Huang on China (L113, L98)#

Source. - The lecture line targets speed under ignorance, whoever argues for it: “the companies will tell you — and governments will tell you too — that we have to go as fast as possible with it, even though we don’t know what it is that’s happening, because if we don’t go fast, somebody else will” (2026-09-24 [mixed]). That applies to Huang’s own “We’re racing as fast as we can” (HA §8.1 T13) as much as to export controls. - The 2025-07-23 passage concerns a US-centric “dominance” vision, with collaboration preferred to “isolationism”. - Earlier he named a US–China AI “arms race” as a concern (2023-07-25), and treated arms-race logic as a limit on responsible innovation (2025-04-06, his framing only; the report is o1-pro’s).

Fix. Keep the alignment on collaboration versus zero-sum framing [Implied, medium]. Add that his race-logic criticism cuts against Huang’s speed rhetoric too [Implied, medium]. Do not use 2025-04-24 (us-and-china-vie-for-ai-k-12-leadership) as evidence: the concept index records it as “Developed by AI, checked by humans”.

13. LOW: Citation, wording and presentation slips#

14. LOW (no change needed): excluded text near cited passages#

M9 is clean, but three cited sources sit beside AI and the Art of Being Human. Future edits should keep it that way. - 2026-07-16, the Vaughan paragraph. The paragraph ends by equating Vaughan’s mechanism with “values drift” from the book (“the small yes that makes the next yes easier”, note 29). M9 uses only the Vaughan part (§2.2, §3.1, INTERNAL Q2). Do not add “values drift” or that quotation. - 2026-05-21. “the boat has already left the harbor” is in the sentence that follows a mention of the book. The sentence is his own prose; cite the post only, as M9 does. - 2026-05-10. Note 10 names the book. M9 uses only the main text (“safety message first”; “near-impossible”).

Checked and correct (selection)#

• Clarifications and rulings. Both September 2026 clarifications are reported accurately and dated (§2.4 L57, D3, D7, §6 item 3). Maynard & Garbee is weighted as fully his (L36). There is no use of AI and the Art of Being Human, Constituting Responsibility, the Fable-credited paper, AI-written notes 2–3 of 2026-01-31, the o1-pro report or Modem Futura. 2026-01-31 n.4 (“already beyond being contained”) is his own note and correctly used. [mixed] is marked on the paper and the lecture throughout.
• 2026-09-15. Every quotation, including “Machines that alter their own instructions”, “remarkably devoid of details on how, exactly”, “not that likely” and “without running around like headless chickens” (n.5), n.1, and “flummox” (n.3, which names Coxon and Amodei’s pacing post).
• 2026-09-24. “We can’t pause it”, “may be a flawed assumption”, “irrelevant to this conversation”, “hack another system”, “cog in the works”, “guardrails … we don’t even know how to do those effectively”, “Good (as in technically capable)”, n.3 “rein them in”, n.4 “knowing that it’s speculation, not reality”.
• 2026-07-16 [mixed]. Persuasion record (OpenAI 2023 → April 2025 → May 2026; Anthropic’s 2024 set-aside and 2026 tiers; DeepMind 2025); the four filters; “measurability in the accepted idiom”; “excellent exhibit, and a weak instrument”; Anthropic 2023 → February 2026; Meta “Stop development” → “Develop with Mitigations”, “uniquely enable” → “substantially contribute to”; “past 2028” and its direction; three blindsides; “does not abandon the idea of risk…”; “not as an alternative, but as an augmentation”; “catastrophic-capability apparatuses”; “not optimistic”; “jurisdiction-bound and politically contingent”; “surprisingly diligent”; “a public trail”; “de facto governance layer”; “your risk is my risk”; “named, mapped and watched”; incentive-field paragraph; posted 16 July, same day as the Hugging Face disclosure, and the text does not mention the incident.
• Other posts. 2019-04-15 (“smoke-and-mirrors”, at worst); 2019-08-13; 2023-05-15; 2023-07-12 (“between these two papers”; “once out of the bag”); 2023-11-18 (“outsized influence”, “far faster”); 2024-01-17; 2024-02-25 (Nvidia, “100% perfect”); 2024-06-20; 2024-07-13; 2025-01-07 (companies “still lack the breadth of vision”, corroborating the lecture); 2025-07-06 (“something akin to motive”, “degrees of freedom”, “weakest part of the link”, “write as well as read access”); 2026-01-22; 2026-01-31; 2026-05-10; 2026-08-30. Dates of all cited posts match their headers.
• Papers and book. PEN 2006 pp.13, 32; Testimony 2007 pp.16, 21; NN 2015-06 p.483; NN 2015-12 pp.1005–1006; NN 2016-06 p.491; Hansen et al. 2008 pp.445, 447; Nature 2011 (“modified as evidence grows”); Bulletin 2008; WEF 2008 (“operate independently of the funders”); Weighing09 (co-written, flagged); Trojan 2026 p.1; Harness 2026 p.8; FFTF pp.162, 163, 288; 30Y, NANO and FWB 2026.
• §2.8’s claim of silence. No other post from July to September 2026 mentions the incident, Astra, the pacing statement, the executive orders or Nvidia’s acquisition.

INTERNAL (not for publication)#

Questions for Maynard arising from this check (the M9 questions are not repeated): 1. Did 2023-11-18’s “a pause even” express a wish that firms would pause, or only a hope that the turmoil would give society room to recalibrate? 2. In 2025-07-23, was the open-weights aside (“welcomed by many who worry about the corporate control”) your own view? 3. May the Series introduction’s sentence on AI “as it is depicted by Huang” be quoted in the public documents before it is published?