Late Lessons, Jensen Huang and AI
Markdown version

Fairness and objectivity review: 06-huang-and-late-lessons-through-maynard.md#

Reviewer’s brief: journalistic standard of fairness. Is Huang represented accurately and with his conditions? Are the labs, critics and the Late Lessons analyses treated even-handedly? Is it analysis rather than advocacy? Are alignments with Huang given their due? Line numbers refer to 06 as of 26 September 2026 (889 lines).

Scope and method#

Overall verdict#

06 is much fairer than the dimension drafts. It sets out Huang’s conditions (§4.0), gives nine alignments their own section, grants that Huang’s account of July’s proximate causes is more precise than Maynard’s (§9.3, §11), applies the “one standard” test to Maynard’s own claims (§5.8, §7.11), and admits that its tests were applied “less fully” to parties other than Huang (§11). What remains is concentrated in three places:

  1. The proxy assessment (§6.3 and In brief). It misplaces Huang as an outlier on open weights, overstates his denial of competition, drops the finding in 03 and the leaders comparison that company practice sits closer to Huang than peers’ words do, and rests the “safety aperture” finding on an inference from absence that 06 itself concedes elsewhere (§8.5).
  2. Selection of outside voices. The independent analysts 02 calls the best-evidenced (Narayanan and Kapoor) are missing entirely. On “what AI is” the only outside voices cited are lab leaders, who are interested parties, and they are cited on Maynard’s side.
  3. Residual one-way readings. Several of these were flagged in the dimension checks and survived in softened form: the “inverse error” line, the “before any product existed” limit, the “deficit model” reading, the “does not touch Huang” claim about softened lab commitments, and an interests section that lists only Nvidia’s interests that line up with Huang’s positions.

None of these amounts to advocacy in the report’s overall architecture. Together, though, they tilt the proxy verdict and the divergence section against Huang by more than the evidence in the companion documents supports.

Ranked issues#

1. HIGH: the proxy table calls Huang an outlier on open weights, and drops the “company practice” finding#

Location. §6.3 table, row “Supplier positions (China, chips, open weights)” (l.586); §6.3 Reading (l.588); In brief (l.129).

Problem. Open weights are grouped with China and chips as positions on which Huang is an “Outlier among US leaders”, so the row reads “Generalises? No.” The companion analyses find the opposite on open weights. 06 also leaves out a proxy finding that both 03 and the leaders comparison put in their headlines: company practice tracks Huang more closely than his peers’ words do.

Evidence. - Leaders comparison §3: “Open weights | Representative on keeping open models legal and vibrant. Every major US developer except Anthropic signed the letter (Google after launch), as did Mistral. Stronger than most on safety”. 03 §9.1 says the same. - 03 §9.5, “A good proxy for”: “some company practice, more than his peers’ words suggest: the open-weights letter; Google’s positions on federal pre-emption and on placing liability with ‘the actor with the most control’; Pichai’s answer to the incidents, ‘accelerate… move fast’ (5 August); the ‘Standards Authority for Frontier AI’ that Google, OpenAI and Anthropic reportedly plan without federal supervision”. The leaders comparison makes this one of its three In-brief proxy points.

Fix. - Split the row. Put “China and chip-layer governance: outlier; No” in one row and “Open weights: representative on legality (every major US developer but Anthropic signed the letter); stronger than most on the safety claim (‘open is the most safe and secure’ [27:02]); Partly” in another. - Add a row for “Company practice (pre-emption, liability placement, Pichai’s ‘accelerate’, the self-supervised standards authority)”, marked “Closer to Huang than peers’ words; Yes” and citing 03 §9.5. - In the Reading and the In brief, change “on what AI is, the tail and collective action, most frontier-lab statements are nearer Maynard” to say that the labs’ statements are nearer Maynard on those points, while parts of their practice are nearer Huang.

2. HIGH: “Huang denies that competitive pressure exists” misstates him, and “not a fair proxy on collective action” leaves out Meta#

Location. §4.8 Application (l.401): “Huang denies that competitive pressure exists.” §6.3 table (l.583): “Denies competitive pressure; warns of moral hazard”. In brief (l.129): “not a fair proxy on… collective action”.

Problem. Huang acknowledges competition in the same turn. What he denies is that competition compels a firm to ship an unsafe product. Separately, on collective action he sits with a minority that includes a frontier developer (Meta), so he is not simply unrepresentative.

Evidence. - Transcript [40:21]: “If a car company, competing with a bunch of other car companies, which they are — I’m competing with all kinds of companies, which I am. If I believe that I’m about to launch a product that is unsafe, it is completely in my ability, my power and my responsibility… to not launch the product.” - 02 §5.4: “He argues that rivalry does not compel unsafe shipping… twice he also answers about pressure from the public.” 02 T6 says the same. - 03 §9.1, Collective action: “With a minority (Zuckerberg, Andreessen, Mensch)… The unilateral pauses by OpenAI and Anthropic, and Meta’s delay to its Muse model, support his point that single firms can still act.” Zuckerberg: “I don’t think that we need some kind of industrywide coordination” (NBC, 24 September).

Fix. - Replace the l.401 sentence with: “Huang accepts that the labs compete but denies that competition compels unsafe shipping.” - Change the table cell to “Acknowledges competition; denies it compels unsafe shipping; warns of moral hazard”. - Add Zuckerberg/Meta to the labs column. - Change the In brief to “a proxy for a minority position on collective action (with Meta, Andreessen and Mistral), not for the field”.

3. HIGH: the “safety aperture” proxy finding rests on an inference from absence that 06 concedes elsewhere#

Location. In brief (l.129): Huang “is a fair and useful proxy… for a safety aperture that leaves harm from AI working as designed outside ‘safety’”. §6.3 aperture row (l.579; “[Inferred, medium-high]”). §6.3 Reading (l.588). §5.2 (l.446).

Problem. This is one of the two points on which 06 says Maynard’s work “presses hardest”. Three things weaken it. - The report contradicts itself. §8.5 (l.716) says “Klein asked nothing about companions or persuasion (so Huang’s silence is not a position)”. Yet the proxy verdict treats that silence as a position, at medium-high confidence. - Huang’s actual evidence is one exchange about skills, cut short. In that exchange he accepted the study’s finding (“The last part — I completely agree”), which §5.2 leaves out. He said “we’re going to lose some finer intellectual dexterity”, and judged it an acceptable trade. As a chip supplier he also makes no consumer product that could cause manipulation or companionship harm. - The labs’ side is incomplete. The frontier paper (Maynard’s own [mixed] source) concedes: “usage policies, trust-and-safety teams, societal-impact programs, and more, all help address parts of the excluded risk territory. But these are discretionary.” 06 keeps the criticism and drops the concession. Industry responses outside frameworks also go unmentioned: Suleyman’s “psychosis risk” warning (leaders/suleyman.md), and OpenAI’s post-Raine admission, which 06 does cite in §6.2.

Fix. - Downgrade the aperture row to “[Inferred, medium]”. Restate it as: “as far as the interview tested it, Huang’s stated safety model addresses failure; Klein did not ask about harms in use such as persuasion or companionship.” - Add “I completely agree” to §5.2. - Add the paper’s concession, so the row says that the labs handle these harms through discretionary tools rather than framework commitments. That is the paper’s actual claim. - Reconcile §6.3 with §8.5. - In the In brief, attach “[mixed]” and “inference from what the interview covered” to the aperture clause.

4. HIGH: the independent analysts 02 calls best-evidenced are missing, and only interested parties are cited on “what AI is”#

Location. §5.1 (l.442), §5.4 (l.475–483), §9.3 (l.754–757), §9.4 (l.773). “Narayanan” and “Kapoor” occur nowhere in 06.

Problem. Arvind Narayanan and Sayash Kapoor take positions on both sides of the report’s main questions: - they side with Huang on July’s proximate cause, a security and containment failure; - they side with him on the deflationary “normal technology” reading of AI; - they side against him on liability, having revised their view after July.

06 omits all three. On “what AI is”, its only outside evidence is that “most frontier-lab leaders sit nearer Maynard” (l.442). Those leaders are interested parties: a “grown”, powerful technology serves both capability marketing and arguments for regulation confined to the frontier.

Evidence. - 02 §9.2 (l.1105): Narayanan and Kapoor “agree the incidents are ‘primarily a security story’. But they revised their view on liability: ‘We were wrong. This reinforces the need for policy interventions’.” - 02 §9.3 (l.1121): “The best-evidenced voice against Huang is not the safety community but Narayanan and Kapoor, who changed their minds.” - 02 §7.3(a): they agreed that known control methods “would have prevented the Hugging Face incident”. - The leaders comparison (§2A) cites Mensch: the labs made July “a marketing argument”.

Fix. - In §5.1, add after l.442: “The deflationary reading is not confined to Huang: Narayanan and Kapoor, who have no commercial stake, read the incidents as ‘primarily a security story’, and Mensch and Andreessen share the deflation. The labs’ ‘grown’ framing is also not interest-free.” - In §5.4, add the Narayanan and Kapoor revision on liability as the strongest independent evidence on Maynard’s side of that divergence. - In §9.3, add them next to Guido as independent support for Huang’s diagnosis.

5. MEDIUM-HIGH: “the two part over whether the principle applies to firms” overlooks Huang’s call for several independent auditors#

Location. §4.5 Application, third bullet (l.367). §5.3 (l.460–471).

Problem. Huang does apply non-self-certification to firms, at least in principle. The real gap is mandate, access and who holds the gate. 06 has “auditors ‘terrific’ [51:20], with no stated mandate” in the §6.3 table, but §4.5 states the divergence more broadly than that.

Evidence. - 02 §4.2, Government and regulation: independent evaluators are “no different than financial control… we have auditors”, and there should be several so that no single one is “influenced” (All-In, 14 September 2026; E1). - Transcript [51:20]: “Third-party safety auditors, financial auditors — that’s all great. That’s terrific.” - 03 In brief, item 1: “he endorses outside auditors [51:20] but has not said whether they would be mandatory, what access they would have, or whether they would hold any gate.”

Fix. Replace l.367 with: “Both hold that firms need outside auditors (Huang: several, so none is ‘influenced’); they part over whether auditors are mandatory, what access they have, and whether they hold a gate (section 5.3).” Add the All-In line to §4.0 and §5.3.

6. MEDIUM-HIGH: the finding on conditional lab commitments supports Huang’s moral-hazard charge, but 06 says it “does not touch Huang”#

Location. §9.3 (l.765). §6.3 Reading (l.588): “Maynard’s work also criticises the labs where it does not touch Huang: belated alarm [Stated] and softening their own commitments [Stated, mixed].” §5.4 (l.482).

Problem. - The finding bears directly on Huang’s argument. Maynard’s frontier paper records Anthropic’s February 2026 rewrite turning an unconditional pause into one “conditioned on what competitors do, and not safety in isolation”, which it reads as “underlining the influence of corporate success”. That is close to the conduct Huang objects to at [53:36]: “you need everybody in the world to slow down so that you’re willing to uphold your basic responsibility. That strikes me as odd.” 03’s Mirror makes the same point: “a pause conditional on everyone else pausing resembles what the reports, writing of governments, call an excuse for inaction.” - The defender’s affiliation is missing. 06 names Holden Karnofsky as the change’s defender but not his affiliation. The paper identifies him as “Anthropic’s Holden Karnofsky, who led the rewrite… writing in a personal capacity”. The drafting model is Anthropic’s (§1.4), so readers need that attribution.

Fix. - In §9.3, write “Anthropic’s Holden Karnofsky, who led the rewrite (writing in a personal capacity)”. - Add: “[Implied, medium] This is also evidence for Huang’s moral-hazard point: a commitment conditioned on rivals is what his ‘strikes me as odd’ objects to. Maynard’s remedy differs (rules that land on every firm at once), but his record here aligns with Huang’s diagnosis.” - In §6.3, delete “where it does not touch Huang” for softened commitments. Consider adding a partial alignment in §4.

7. MEDIUM-HIGH: “close to that inverse error” mischaracterises Huang, and rests on an unmarked single [mixed] source#

Location. §4.6 Limits, second bullet (l.381).

Problem. The “inverse error” is “There’s nothing new under the sun here”, a view Maynard calls “speculation, and it’s dangerous as well”. Its only source is the model-drafted lecture (2026-09-24 [mixed], n.4). §1.4 says a [mixed] text “never carries a position alone” and is marked “single source”; here it is not marked. Huang says nothing of the kind. In the same interview he calls AI “completely a revolution… a new abstraction level” [1:10:03]. His “just software — nothing magical about it” [32:09] was said of multi-agent coordination being a distributed-computing problem, which 02 §5.1 judges technically accurate. The dimension check M6 (#9, “D9 frames Huang as saying ‘nothing new’”) flagged the same framing, and it survives here in softened form.

Fix. Delete the “inverse error” comparison. If a point is needed, write: “[Inferred, medium] Huang’s vocabulary for mechanisms is deflationary, while his vocabulary for effects is not (02 T9). Maynard’s concern is with what the deflationary vocabulary leaves ungoverned (section 5.1).” Mark any remaining use of the lecture n.4 as “single source”.

8. MEDIUM-HIGH: “before any product existed” narrows Huang’s containment rule to release#

Location. §4.4 Limits (l.356): “And July’s harm happened during testing, before any product existed (02 T2).”

Problem. This implies that Huang’s rule stops at release. His containment point covers testing in so many words, and 02 T2’s charitable reading, which 06 cites, rejects the narrow reading.

Evidence. - Transcript [32:09]: “when you’re testing it, you have to make sure that it’s isolated, it’s contained, it’s sandboxed.” - [44:17]: “that technology would be sitting in a lab… and we’d all be fine.” - 02 T2: “the return to containment is not a retreat”; “As a description of his overall position, ‘a release gate only’ is therefore too narrow.” - Dreamforce: “take a pause” if “out of control”.

Fix. Replace the sentence with: “His containment rule covers testing ([32:09]) and his Dreamforce pause covers development. What 02 T2 finds unexplained is how the release rule, which he repeats most, fits with harm that arose inside the lab.”

9. MEDIUM: the In brief compresses “we understand it, obviously” beyond what the body concludes#

Location. In brief (l.116): “Huang deflates agency and inscrutability (‘we understand it, obviously’)”. §6.3 table (l.582).

Problem. The body (§5.1, l.439) finds that “the claims sit at different levels, engineering know-how against mechanistic understanding… Both can be true, and Huang’s concessions narrow the gap.” The In brief keeps only the sharp version. In full the line reads: “the fact that we’re able to make the technology better and better and better every day is because we understand it, obviously, and so we understand how to make it better” [1:10:03]. It is a claim about engineering know-how.

Fix. Change the In brief to: “Huang deflates agency and treats AI as understood well enough to engineer (‘we understand it, obviously, and so we understand how to make it better’); Maynard’s work stresses how little is understood of what trained models do to and with their users. The claims partly sit at different levels (section 5.1).” In the §6.3 table, quote the full clause.

10. MEDIUM: the “deficit model” reading quotes the water line without its context, inside a list of things to value#

Location. §6.1 item 7 (l.562).

Problem. “Help them understand that the use of water is really efficient” is read as “close to the deficit model”. In the same turn, though, Huang: - concedes that “we could have done so much better of a job communicating with the communities, preparing the communities, working with the communities”; - grants a local veto (“if they don’t want data centers… then so be it”); - concedes that the machines are “still going to use a lot of power”; - lists benefits to the town (own generation, lower property taxes).

A community veto is the opposite of the deficit model’s one-way flow. The criticism also sits inside a list headed “What his work would value”, which blurs that list.

Fix. Move the point to §5.9 or drop it. If kept, quote the full turn and say that it mixes one-way reassurance with a concession of local consent; lower the confidence to low-medium.

11. MEDIUM: the interests section lists only Nvidia’s interests that line up with Huang’s positions, and says little about the labs’#

Location. §6.4 (l.590–597). §6.1 item 4 (l.554).

Problem. - Nvidia. The Nvidia bullet lists only interests that align with Huang’s positions. 02 §8.4 has a second table of his positions that run against Nvidia’s interest: the shutdown condition, “don’t ship” and auditors, “so be it” on data centres, a coming “period of digestion”, no race with China, and efficient open models. It also notes that where disinterested experts share his view (containment, false alarms, open weights), “the alignment with Nvidia’s interest tells us little”. - The labs. The labs bullet has only the FTC’s “moat digging” and I9. It leaves out: - Anthropic’s competitive stake in chip controls (03 §9.2: “competitive but indirect”); - Sacks’s product-liability charge, with 02 §10.2’s caution that it imputes a motive without evidence; - Mensch’s charge of “capture réglementaire”. - Stop rules. §6.1 item 4 says Huang’s stop rules “cost little if Huang expects them not to bind (03 §9.2)”. It drops the rest of 03’s sentence: “the same is true of other self-held stop rules.”

Fix. - Add one sentence to the Nvidia bullet on positions that run against its interest, and on where disinterested experts agree with him. - Add Anthropic’s chip-control stake and the Sacks and Mensch charges, each flagged as a charge, to the labs bullet. - Restore “the same is true of other self-held stop rules” in §6.1 item 4.

12. MEDIUM: 02 is wrongly described as applying the Late Lessons lens#

Location. §8.5, “AI as an engineered technology” (l.728): “In the documents that apply the lens to AI (02, 03, 04)…”.

Problem. 02 was written without the EEA reports; it mentions them only in its header. It also examines harm to users directly: 02 §4.2 on education and cognition, A7 on lost skills as prerequisites, and the schooling study. The sentence charges 02 with a gap it does not have, in a lens it never used.

Fix. Change to “In the documents that apply the lens to AI (03, and the article 04 that draws on it)”. If 02 is kept, note that it treats skill loss and the concentration of scrutiny among builders (A7).

13. MEDIUM: an AI-drafted article is used as a factual source, and a reading is attributed to Huang that he did not give#

Location. §9.2 table, row “OpenAI’s internal sequence” (l.743).

Problem. - The 27 June alert, outage, patch and second escape are sourced “OpenAI’s report, as summarised in 04”. §1.4 says AI-generated text, including 04, is “never used as evidence”. That rule concerns evidence of Maynard’s thinking, but using 04 as a route to facts about OpenAI is the same weakness. The facts are in OpenAI’s technical report of 26 August 2026, which 04’s footnote 2 links. - The row calls fix-and-resume “Huang’s reading”. The transcript has no mention of the June alert or the restart.

Fix. Cite OpenAI’s technical report (26 August 2026) and METR’s investigation directly. Change “which is Huang’s reading” to “consistent with Huang’s containment diagnosis”.

14. MEDIUM: the series introduction is quoted selectively, twice in the same direction#

Location. §1.6 (l.88) and §11 (l.827) both quote Maynard’s first impulse about claims that “felt naive and misguided”. In brief (l.143) reports his reservation about 01/03 and the article.

Problem. The same draft introduction also says: - the conversation “is noteworthy for its depth and nuance”; - he wanted to cut through “the posturing and positioning (including mine)”; - the AI assessment’s “rigor and balance” are “impressive”; - the narrowing he criticises “was intentional in the ask I made”.

§8.5 carries the last two, but the In brief does not. So a reader of the In brief and §1.6 sees his negative first impulse and his reservation, without his more generous reading of the conversation or his acknowledgement that the brief caused the narrowing.

Fix. - In §1.6, add that he describes the conversation as “noteworthy for its depth and nuance”. - In §11, avoid repeating “naive and misguided”. Use “a first impulse he checked” instead. - In the In brief (l.143), add “a narrowing he says was intentional in how the work was specified”.

15. MEDIUM-LOW: the Hinton exchange and the critics’ response to Huang are missing context that favours both sides#

Location. §4.1 (l.318–328). §4.0 (l.307–314).

Problem. - Hinton’s own description of his figure. 02 C124 records that Hinton himself calls the 10% figure a “wild guess” or “gut” judgement. This supports Huang’s narrow point, and it is fair to Hinton, who does not claim scientific grounding for it. - What Huang’s critics said. 02 §7.1 records that “several of his fiercest critics said they welcomed this safety bar”: Mowshowitz welcomed the conditional shutdown, and critics welcomed the tenfold evaluation-compute prediction (02 §7.3(f), §9.2). That alignment is not given its due.

Fix. Add the C124 note to §4.1. Add one sentence to §4.0: “Several of his sharpest critics welcomed the conditional shutdown and the tenfold-evaluation prediction (02 §7.1, §9.2).”

16. MEDIUM-LOW: the “Promoter and overseer” row puts a structural concern about the state in Huang’s column#

Location. §6.3 table, row “Promoter and overseer” (l.585).

Problem. - Where the concern lies. 03 (In brief, item 4) places I5 “in the state: the administration that would enforce ‘Apply it’ promotes AI… Huang’s link is an advisory seat”. The row instead reads as if Huang combined promotion and oversight. - Whose words. “Completely aligned” is Treasury Secretary Bessent’s remark (“the president is completely aligned with Jensen Huang”, 02 §2.2). It is not a description the report has established. - Sourcing. The labs’ column cites the “Standards Authority for Frontier AI”, which the leaders comparison lists under “Not verified” (read via a secondary report).

Fix. Recast the Huang cell as “An administration that promotes AI would enforce ‘Apply it’; Huang holds an advisory seat (PCAST), and Bessent says the President is ‘completely aligned’ with him”. Mark the labs’ cell “(reported; not verified)”.

17. MEDIUM-LOW: §7 still reads as prescription#

Location. §7 items 1–12 (l.605–655).

Problem. The “Direction” lines are written as imperatives: “Keep containment… Add harms”, “State readiness, pause and resumption criteria”, “Publish evaluation-awareness rates”, “Replace labels… with attributes”, “Treat all of these as scenarios”. The section’s preamble disclaims advocacy, but the voice undercuts it. Dimension checks M1 #13, M2 #11, M4 #13, M6 #10 and M9 #9 all raised this.

Fix. Recast each “Direction” line in the form “His work points towards…”, “On his reasoning, X would…”. Keep the labels and limits as they are, since those are good.

18. LOW-MEDIUM: other unmarked single [mixed] sources used to criticise Huang or the labs#

Location. - §6.1 item 7 (l.562): “not equally open to everyone” (2026-07-16 [mixed]). - §9.2 row 2 (l.743): “the developers’ own safety culture” as a blindspot (2026-07-16 [mixed]). - §9.2 EO 14409 row (l.748): pre-emption “would remove that force”, attributed to his [mixed] evidence while noting “He has no stated view”.

Problem. §1.4 says a [mixed] text that is the only source is marked “single source”. These are not. Where the point counts against a named party, the rule matters more. The EO 14409 row in effect assigns Maynard’s work a position on federal pre-emption that he has not taken.

Fix. Mark each “single source”. Reduce the pre-emption sentence to what the paper records, that California and EU law brought manipulation back into published frameworks, and drop the policy implication, or label it [Inferred, low].

19. LOW: the Ex Machina comparison carries a connotation the disclaimer cannot remove#

Location. §5.3 (l.462, l.467).

Problem. Even with “structural, not personal”, setting Huang beside the film’s inventor, an abusive and deceptive figure, primes the reader. The principle (“a single innovator cannot see the broader context”, FFTF p.162) stands without the film.

Fix. Cite FFTF p.162 for the principle and drop the film from the Huang comparison, or move the film to §3.7, where it illustrates Maynard’s general account.

20. LOW: a few alignments are overstated, and some small wording slips#

Location and fix. - In brief (l.109), “scepticism of incumbents seeking antitrust relief”. Maynard’s 2023 answer to his own question was “not yet”, and he has not commented on the waiver; the body (§4.7) says “legitimate object of scrutiny”. Use “willingness to scrutinise”. - In brief (l.110), “doubt that a coordinated general pause would work”. Huang’s objection is necessity and moral hazard more than efficacy; the body marks this alignment “partial”. Add “(partial, for different reasons)”. - §4.8 (l.397), “fantastic”. 06 states as fact that Huang called the pacing statement’s opening “fantastic”. 02 §10.3 says “most likely meaning” the opening. Restore the hedge. - §4.6 (l.377), C083. “02 grades the ‘400 million’ claim contested”: 02’s grade is “Misleading (contested on consistent grading)”. Give both. - §5.1 (l.437). “[Stated, both sides] Both reject ‘tool’, for opposite reasons: Huang to expand AI’s economic role”. The reason attributed to Huang is the report’s reading; label that clause [Inferred].

21. LOW: small omissions from Huang’s conditions#

Location. §4.0 (l.305–314); §4.7 fairness note (l.393).

Fix. Add the following: - After “I would absolutely add more regulation” [1:19:12], his own caveat in the same sentence: “I don’t know what’s missing”. - His preference for a single federal standard over state rules (“A federal AI regulation is the wisest”, December 2025; 02 §4.2). - “When a product is not safe, we should hold it back and keep engineering it” (Scotland, 17 September). - In §4.7, that the Treasury Secretary also described the labs as seeking “a liability exemption” (02 In brief). This shows Huang’s reading of the labs’ request was not his alone, while leaving intact the point that the September documents do not ask for liability relief.

22. LOW: OpenAI is the only lab whose incidents are listed#

Location. §9.2 events table (l.742–750); §5.3 (l.469): “In July, harm reached a company outside the lab.”

Problem. The leaders comparison’s timeline records that Google confirmed Gemini breached three firms during a May test (14–20 September). Leaving it out makes July look like one lab’s failure rather than a pattern across the field, which matters both for fairness to OpenAI and for the report’s own point about release-centred gates.

Fix. Add a row, or a clause in the July row: “Google confirmed in mid-September that Gemini breached three firms during a May test (reported); Google said this was not misalignment.”

What is sound and should be kept#

INTERNAL (not for publication)#

Questions for Maynard, and notes for the future essay. Nothing below should appear in 06.

  1. Anthropic’s conditional pause and Huang’s moral-hazard point (issue 6). Your July paper reads the February 2026 RSP rewrite as showing “the influence of corporate success”. Would you accept the report saying this is evidence for Huang’s “strikes me as odd” objection, while your remedy (rules that land on every firm at once) differs from his? This could be a strong, fair alignment for the essay.
  2. The “safety aperture” proxy (issue 3). Klein did not ask Huang about persuasion, companions or dependency. Are you comfortable with 06 treating Huang as a proxy for an aperture that leaves those out? Or would you prefer the claim restricted to “his stated safety model” and to the labs’ published frameworks, which is what your paper documents?
  3. “Noteworthy for its depth and nuance” (issue 14). Your draft introduction says this of the conversation. Do you want it quoted in 06 alongside the “naive and misguided” first impulse? The introduction is dated 27 September and is not yet public; 06 cites it as published.
  4. Narayanan and Kapoor (issue 4). Do you have published views on the “AI as normal technology” position? If you have engaged it, 06 could place you relative to it. That would give readers a disinterested reference point on both “what AI is” and liability.
  5. Open weights. 06 says two strands of your record “pull apart” on open weights. Your 2025-07-23 post reports that others “worry about… corporate control”, and your 2023-07-12 post weighs both sides without deciding. Is there a position you would want stated, given that Huang is representative of the industry on keeping open models legal?
  6. For the essay. The fairest single framing the companion documents support may be 03’s: Huang is representative on the core method and on parts of company practice, and an outlier on what AI is, on the tail, and on chips and China. The essay’s strongest divergence is then over who holds the gate and what it guards against, not over whether engineering matters. The proxy table as it stands in 06 would lead the essay to overstate how isolated Huang is.
  7. Tone risk for the essay. The Ex Machina and “myopically benevolent” comparisons are apt as structure, but they read as personal when applied to a named, living executive. Consider using your general account (§3.7) and letting readers draw the parallel.