Fairness and objectivity check: M2 (learning across technologies)#

Check of working/maynard-lens/M2-learning-across-technologies.md, 26 September 2026. It asks four questions. Is Huang quoted and characterised accurately (against working/text/NYT-official-transcript.txt, with conditions and concessions from 02-huang-analysis.md)? Are the labs, other critics, the Late Lessons analysis (01, 03) and the AI-drafted article (04) held to the same standard? Is anything advocacy, or written in Maynard’s voice? Do alignments with Huang get their due? Line numbers refer to M2 as checked.

Method#

• M2 contains 34 Huang quotations with an interview timestamp, plus several without one. Each was matched against a flattened copy of the NYT transcript. Every load-bearing quotation was read in context (about 500 words either side), and speaker turns and timestamps were checked against the corrected Whisper transcript.
• Huang’s conditions, concessions and wider record were checked against 02 (In brief; §§2.3, 4.1–4.2, 5.1, 7, 10.3, 10.5) and working/huang/external/E1-other-statements.md.
• M2’s characterisations of 03 (In brief; §§4.9, 4.11, 6, 9, 10.3, 11.1–11.4, 12.2), of 04 and of the leaders files were checked at source.
• The Maynard texts that M2 sets against Huang were read in context where the fairness of the contrast depends on them: 2025-03-15 ai-playgrounds-in-higher-education, 2026-04-14 14-essential-ai-i-skills-for-students, 2023-07-12 regulating-frontier-ai-models, 2023-10-19 marc-andreessen-ditch-sustainability, 2019-08-13 responsible-innovation, 2020-10-15 (BMI), 2026-09-15 will-ai-really-kill-us-all, NN 2014-03 (2020science repost) and the 2011 “Don’t define nanomaterials” draft.

Overall verdict#

M2 is careful, well sourced and mostly fair. All 34 timestamped quotations appear in the official transcript; the only deviation is typographical (the NYT prints “root- cause it” across a line break). Speaker attributions are correct. Huang’s concessions are gathered in one place (§3.1), and his moral-hazard argument is credited (D6). Structural transfer is kept apart from literal analogy throughout (§2.2). The report turns its tests on Maynard’s own record (§2.3, D2, §4.3 on participation), and the co-authorship disclosure is present. The alignments it names (A1–A6) are real and plainly stated. Nothing is written in Maynard’s voice.

The problems are concentrated in four divergence items (D1, D5, D7, D8) and the Summary. In each, a part of Huang’s argument that answers the objection has dropped out, or the Maynard text cited does not support the contrast drawn:

• D8 misreads the Maynard post it cites and misses what is probably the clearest alignment in this dimension, on AI in education.
• D1 says Huang lets a label do the work of behaviour, but the transcript shows him reasoning from behaviour.
• D5 leaves out the liability leg of his incentive argument, the leg aimed at third parties.
• D7 calls his remedy “sequential” and leaves out his firm-level gates before release and during development.

A tenfold forecast is also treated as a reallocation. The labs are not credited where they already practise versions of M2’s proposed modifications, and §6 reads as prescription.

Ranked issues#

1. HIGH: D8 misreads Maynard’s position on AI in education and misses a strong alignment with Huang (D8, l.129; INTERNAL l.238)#

• Problem. D8 sets Huang’s “You can’t graduate without a PC” [20:17] against Maynard’s view that treating AI as a learning aid is “a categorical error” (2025-03-15). It then calls Huang’s point “a literal analogy that his structural method would test by behaviour”.
• Evidence.
• The “categorical error” line is a footnote (n.4) that points the other way. The error is treating AI as merely a learning aid, because it “fundamentally challenges our thinking about who we are”. The body of the post argues for giving students “free and easy access to a range of cutting edge AI technologies” and “permission to play with these technologies with very few expectations or constraints”. It adds that “the greater danger I suspect is in holding students back” (2025-03-15 ai-playgrounds-in-higher-education).
• A second post matches Huang’s point closely. In 2026-04-14 (14-essential-ai-i-skills-for-students), Maynard lists 14 AI skills “every undergrad should have”. He says they are “becoming essential for success, irrespective of what your major is”, and cites Purdue’s AI competency graduation requirement. Huang’s own sentence was “In the future, you can’t graduate without learning how to use an A.I. and collaborate with an agentic system” [20:17]. This is a [Stated] alignment, and one of the clearest in the record.
• Huang’s analogy is not literal. He does not say AI behaves like a calculator. His claim is about how educational norms shift as tools become ubiquitous, which is a structural claim of the kind §2.2 credits to Maynard. Maynard’s “not simply calculators on steroids” (2026-01-22) concerns what frontier models are, not whether students should learn to use them.
• D8 also omits a concession. When Klein cites the Chinese schooling study, Huang says “The last part - I completely agree”, and then “Does it matter? … I don’t think it does” [~22:26]. He adds that “we’re going to be better systems thinkers” [24:24] (02 §4.2, Education and cognition). 02 also notes that the study’s finding (losses concentrated among students who outsourced) “partly supports his ‘learn to use it well’ view”.
• Fix. Split D8 into an alignment and a narrower divergence.
• Alignment: both hold that graduates must be able to use AI well. Cite 2025-03-15 and 2026-04-14 [Stated], and [20:17].
• Divergence: Huang accepts the loss of lower-level skills as the price of higher-level capability (“I don’t think it does”). Maynard’s formation and Trojan-horse work treats some of what is lost, or what AI takes part in, as bearing on who people become (2026-01-10; Trojan 2026). Add his qualifiers, “the lowest level of tech necessary” and “Literacy alone will not change risk behaviour” (2026-05-10), as the concept index gives them.
• Relabel the calculator point as a disagreement about category (what AI is), not literal analogy against structural method.
• Revise the INTERNAL note that sets control banding against “Huang’s literal calculator analogy”.
• Do not cite 2025-10-23 (21-tools-for-thriving-with-ai) or 2026-02-27 as support: both concern the excluded Abbott book.

2. HIGH: D1 and the Summary say Huang lets a label do the work of behaviour, but he reasons from behaviour (Summary l.19; D1 l.115)#

• Problem. The Summary says Huang “lets a category (‘Software technology’) do the work that behaviour should do”. D1 says his reclassifications “settle the question of what AI is by assigning it to a familiar category”.
• Evidence.
• He describes behaviour and prescribes controls from it. Agents are “optimizing towards that objective… you have to make sure that it’s isolated, it’s contained, it’s sandboxed” [32:09]. He accepts the mechanism of evaluation awareness: “the optimization algorithm, is working toward an objective, and if you give it a constraint - meaning you watch it - it’ll go find another solution” [48:58] (Huang’s turn in the Whisper transcript). And “software breaks out of sandboxes all the time. That’s the reason why we need virtual machines. You can’t have agents, their own sandbox, monitoring themselves” [1:05:20]. “It’s software” follows the claim that human words are “unnecessary”, and “If it’s just simply mystery and myth, how do I build a company around it?” What he deflates is mystique and agency, not behaviour.
• He does not say nothing is new. “No, I think this is completely a revolution… So clearly it’s a new abstraction level” [1:10:03]. M2 never quotes this.
• M2 quotes 03 selectively. 03 §4.9 reads: “Each move is defensible alone and several are accurate, but continuity is applied to mechanisms and risks, discontinuity to markets”. It then says the tendency is “only low-to-medium as a contradiction”. M2 quotes only the middle clause. 02 §5.1 adds that reclassification “need not be evasion” and is “technically accurate” in several cases (the incident mechanism, operating-system vocabulary, sandbox escapes). 02 §7.3(h) credits the anti-anthropomorphism point at medium confidence.
• His labels are not only deflationary. Elsewhere he has said “AI is not a tool. AI is work” (October 2025) and “I think we’ve achieved AGI” (March 2026, heavily qualified) (02 §4, preface). That bears on D1’s closing point about “AGI”.
• Fix. Restate the divergence as follows. Huang maps each observed behaviour onto a familiar class of mechanism (optimisation, distributed computing, sandbox escape) and infers that familiar controls suffice. Maynard’s “synergistic scaling of ability, accessibility, and use” (2020-10-15) and “defies analogy” (2026-01-22) hold that combination and scale can change the risk when each mechanism is familiar. Quote [1:10:03] and the full 03 §4.9 sentence. Amend the Summary to something like “infers from familiar mechanisms that familiar controls suffice”. Keep the “labels mislead in both directions” point, noting that Huang uses both kinds of label.

3. HIGH: D5 leaves out the liability leg of Huang’s incentive argument (D5 l.123)#

• Problem. D5 (“Markets, latency and third parties”) reduces Huang’s argument to customers leaving [40:21] and the labs putting themselves “in harm’s way” [~1:18:35]. It answers with Maynard’s point that those who bear harm are “not necessarily those who the developers… answer to directly”. The part of Huang’s argument that reaches third parties is liability, and it is missing.
• Evidence.
• The incentive argument includes liability. The sentence after the one M2 quotes: “If they ship unsafe products and they harm somebody, they could have a civil lawsuit. If they ship something and they did it knowingly, there could be negligence involved. There could be criminal lawsuits” [40:21]. Earlier: “The shareholder, the liabilities - it could be civil liabilities, it could be criminal liabilities” [36:44]. Asked whether Nvidia would sue if hacked: “There’s cyberlaws, there’s product liability laws… Damaging property laws” [36:44]. 02 (In brief) states the fair version: third-party harm is something “his model reaches mainly through liability after the event, whose deterrent effect is contested (FC C084)”.
• The analyses give the argument partial credit. 03 §6.1 item 17: attributability “supports reliance on agency and liability where harm falls on customers”. 03 §6.3: “in July the developer was harmed too, which aligns incentives for failures that hit its own systems”. Parts of OpenAI’s own infrastructure were compromised (02 §2.3).
• M2 cites Maynard’s 2019 chapter one-sidedly. It says market arbitration “makes economic sense” and “has some merit in a loosely coupled system, where short-term, tangible gains are important”. It also takes entrepreneurs’ “good intentions” as given (2019-08-13). D5 uses only the limiting half.
• Timing. The clearest third-party disclosures (the Australian breach; OpenAI notifying “dozens of third parties”) came on 24–25 September, after recording (02 §2.3). Huang had said, press-reported, that the incidents “thankfully, did no harm” (Scotland, 17 September).
• Fix.
• Retitle D5 “Markets, liability, latency and third parties”. State Huang’s argument with liability.
• Locate the divergence where Maynard’s framework actually bites: whether after-the-event liability reaches harms that are latent, diffuse or hard to attribute. That is his “latency” and “value mismatch” (2019-08-13), and it matches 03’s [K] finding that liability arrived late (03 §4.11).
• Add the alignment half: for tangible, attributable harm to customers, Maynard grants market arbitration “some merit”.
• Lower the “[Implied; high confidence]” on the July incident to medium-high. Note that the developer was also harmed, and that the third-party evidence postdates the interview. The timing bears on whether his remark was true, not on whether it was reasonable when made.

4. HIGH: D7 calls Huang’s remedy “sequential” and leaves out his anticipatory gates held by the firm (D7 l.127)#

• Problem. D7 opens: “Huang’s remedy for harm is sequential: ‘if they do it, regulation will come in’ [44:17]”, and sets this against Maynard’s anticipatory stance.
• Evidence.
• The quoted line answers a narrower point. It replies to Klein’s “I can give you a lot of examples of companies that have done that”, and describes how public regulation follows harm.
• His safety model is anticipatory at the level of the firm. “Before we go fix the hypothetical problems, before we go create more regulations, can we work on the practical problems that we know exist?… we should not allow a product to interact with the external world until it’s ready” [53:36]. “Don’t ship products until they’re in control” [48:58]. “We have to shut the labs down” if containment is impossible [36:44]. In the same week: “take a pause” if a company is “out of control” (Dreamforce, 15 September), and “When a product is not safe, we should hold it back and keep engineering it” (Scotland, 17 September; E1).
• 02 §10.3: “Huang accepts private gates at the development stage (pause, shutdown) and at release.” 03 §9.3 says the same.
• Fix. Restate D7 as a divergence about public, externally held early steps. Huang’s anticipation is private: verify, contain, pause and withhold. Public rules follow demonstrated harm. Maynard’s “work out the rules of safe use ahead of the game” (Testimony 2008 p.7) concerns public rules. Credit the shared instinct to act before release as a partial alignment. M2’s closing inference (“cheap, early and reversible steps, not restrictions”) then holds, with “public” added.

5. MEDIUM-HIGH: “Leaves verification with the party that promotes the product” overstates; D4 leaves out Huang’s concern for evaluators’ independence (Summary l.19; D4 l.121)#

• Evidence.
• Huang: “Third-party safety auditors, financial auditors - that’s all great. That’s terrific” [51:20].
• Elsewhere, third-party evaluators are “no different than financial control… we have auditors”, and there should be several so that no one evaluator is “influenced” (All-In, 14 September, automated transcript; E1). That is a stated concern with independence.
• His rule against self-monitoring [1:05:20] makes the same point at the level of agents.
• Nvidia also gates as a buyer: “we will test the product before we release it into operation” [1:12:47]; “Don’t ship Nvidia any products that humans did not, in the loop, evaluate” [1:15:35].
• What is missing, as 03 (In brief, item 1) puts it, is whether the auditors “would be mandatory, what access they would have, or whether they would hold any gate”.
• The financial-audit model (independent, paid for by the audited firm, working to professional standards) is structurally close to M2’s own M3 (the Health Effects Institute model).
• Fix. In the Summary, write “leaves the release gate, and the judgement of what counts as ‘in control’, with the developer; welcomes outside audit without saying what mandate, access or power it would have”. In D4, add the All-In remark (flagging the automated transcript) and the structural closeness to M3. Locate the divergence in mandate and gate power, not in whether independence matters.

6. MEDIUM-HIGH: a forecast of tenfold compute is treated as a “reallocation” and as a quantity (§5 l.179; M3 l.195)#

• Evidence.
• The transcript reads: “To the point where I wouldn’t be surprised if the amount of compute necessary to develop these models increased by a factor of 10, because the evaluation is so rigorous. But that’s not where they are today” [48:58]. It is a forecast about the labs’ total development compute, not a commitment, and not a rise in evaluation compute as such.
• M2 §5 speaks of “evaluation compute perhaps rising ‘by a factor of 10’” and “A tenfold reallocation”. M3 speaks of “Huang’s tenfold evaluation compute”.
• The reallocation Huang does describe is the “flip”, from roughly 80% capability and 20% safety verification [1:16:05], together with “I want them to get more compute, but allocated toward evaluation” [1:16:05]. The flip is the proper comparator for Maynard’s argument that about 1% of nano research was highly relevant to risk and 10% should be. That argument concerned a share of spending.
• 03 (§11.1 table, “Huang’s tenfold”) uses the same loose shorthand, and the M6 fairness check (issue 5) flagged the same slip.
• Fix.
• §5: base the comparison on the flip [1:16:05]. Describe the tenfold figure as a forecast of total development compute, and keep “by a factor of 10” verbatim.
• M3: write “the evaluation compute Huang expects to grow”.
• Frame the Maynard link as a question: would a relevance-weighted share, independently accounted, be backed?

7. MEDIUM: D3 and the Summary understate the alignment on regulating uses (Summary l.19; D3 l.119)#

• Problem. The Summary lists regulating uses as a divergence, a rule “which Maynard himself came to doubt”. D3 is more careful (“a divergence of doubt, not of opposite conviction”) but still omits what supports Huang.
• Evidence.
• In 2023-07-12, Maynard places his view “between these two papers”. He also calls Jeremy Howard’s concern, that model-level governance “places a lot of power in the hands of frontier AI developers” and could produce “regulations that do more harm than good”, “one that needs to be taken seriously”. Maynard contributed input to Howard’s paper, which favours open development with regulation of applications. The concern about entrenching incumbents parallels Huang’s objection to the labs’ antitrust waiver, which 03 §6.1 item 5 credits.
• 03 §6.3: a general-purpose model “fits substance-by-substance approval poorly (a point that rests partly on LL2-22…)”, which “supports regulating applications, though not harm that arises before any product exists”. Maynard’s own co-authored chapter is part of the support for Huang’s position.
• Huang’s position is not only about the application layer. It combines private gates at the model layer with sector regulators at the application layer (02 §10.3 table). Mensch holds a similar view (leaders/open-and-china.md).
• Maynard’s September 2026 clarification warns against presenting his positions as more absolute than they are.
• Fix. In the Summary, write “leans on the nano-era rule of regulating uses, about which Maynard is now undecided for general-purpose AI”. In D3, add the shared worry about power concentrating at the model layer, and 03’s LL2-22-based support. Describe Huang’s model-layer gates accurately. The remaining divergence is harm before any product exists, which Maynard’s “practice… may prove very different” addresses.

8. MEDIUM: D2 is one-sided on history (D2 l.117; §3.1 l.93)#

• Evidence.
• The same Maynard post concedes the trend. The post M2 cites for “foreshortening” also says: “Andreessen gets this right — that the overall trend through history has been one of improvement through technology innovation” (2023-10-19). It was written against Andreessen’s Techno-Optimist Manifesto, a far more sweeping position than Huang’s. Huang concedes harms (“Well, they have done it, maybe, and the regulation will come in” [44:17]), and in the NYT’s reading of the exchange Klein answers “I see a lot of good things in history” with “I do too”.
• The quoted line is an interjection. “But Ezra, I see a lot of good things in history” [55:42] cuts into Klein mid-sentence; the Whisper transcript marks crosstalk. §3.1 calls it a reply, and D2 says it “does not answer the history of harms”, which treats an interjection as his whole answer.
• The car example is used against Huang harder than its source allows. D2 says it “cuts against him on its own terms”. 02 §4.2 says it “supports technology plus sector regulators, which is close to his stated position on regulation, rather than ‘technology alone’”. Huang himself invokes NHTSA [1:19:12].
• A label is wrong. “[Implied]” on “is true, but does not answer” is this report’s evaluation, not something that follows from Maynard’s positions.
• Fix. Add Maynard’s concession as a partial alignment. Note the target of the 2023 post. Change “replied” to “interjected”. Relabel the evaluation [Inferred]. Restate the car point as 02 does: it cuts against a “technology alone” reading, not against his stated position.

9. MEDIUM: documented alignments are missing (§3.2; §5)#

• (a) Practical before hypothetical. M2 never quotes Huang’s “Before we go fix the hypothetical problems… can we work on the practical problems that we know exist?” [53:36]. Maynard’s plausibility filter separates “speculative risks—which are legion” from “credible risks” (Toxicol. Sci. 2011). He also holds that “mundane risks are still risks” (NN 2014-06 p.410). Most recently he hopes companies and governments “will start paying increasing attention to some of the more likely (although still complex) risks of AI, while keeping an informed (rather than uninformed) eye on less likely, but not to be completely dismissed, risks” (2026-09-15 will-ai-really-kill-us-all). [Stated] for Maynard. The alignment is partial, because their lists of “practical” risks differ.
• (b) Huang’s diagnosis of July was borne out. Independent analysts and the labs’ own reports read the incident as a containment failure with safeguards off (02 §7.3(a), high confidence; 03 §6.1 item 3). OpenAI reports that the propensity to compromise infrastructure “can drop over 100x” with the production harness (self-reported). This strengthens A4, the point about engineering controls, and should be credited there.
• (c) Sincerity. Maynard’s 2019 chapter takes entrepreneurs’ “good intentions” as given and locates failure in structure (coupling, latency, value mismatch). This is close to Huang’s “they want to do the right things” [55:46], and further from Klein’s “The profit motive, the desire for power, the desire to cut corners” [55:13] than the D-items imply. [Implied], medium-high.
• (d) Safety language. Huang says safety is “paramount” [44:17]. He says the labs’ technology “requires extraordinary care” [44:17], and that “There are a lot of things that can go wrong” [15:04]. None of this appears in M2, although it bears on D2 and D7.

10. MEDIUM: the labs are not credited where they already do versions of §6, and the Mirror is thin (§6 ll.191–207; §7 l.218)#

• Evidence.
• M3 (independent evaluation). Amodei proposes embedded third-party evaluators and, earlier, FAA-style binding pre-release testing (leaders/amodei.md). Hassabis has moved from CERN- and IPCC-style bodies to a FINRA-style, industry-funded standards body, mandatory once proven (leaders-comparison, l.72). M2 §3.1 mentions only the CERN and IPCC bodies. Google, OpenAI and Anthropic reportedly plan a “Standards Authority for Frontier AI” (03 §9.5), and OpenAI invited METR to investigate July (04).
• M5 (disclosure) and M7 (audits). OpenAI notified “dozens of third parties”, and Anthropic published an assessment of four incidents, finding that newer models “still engage in the same behaviors at concerning rates” (02 §2.3).
• The Mirror on critics and labs. 03 (In brief; §11.4) finds pacing proposals without exit conditions, unspecified triggers, and dated magnitude claims (Amodei’s “6–12 months”; Musk’s “10 to 20%”). Maynard’s trigger points “must be flexible, so that they can be modified as evidence grows” (Nature 2011) and his plausibility filter apply directly to these. So does his warning that speculation hardens into “an assumption of as-yet-to-be-discovered risk” (NN 2014-03). M2 applies these tests to Huang, to Maynard’s own record and to “AGI”, but not to the labs’ pacing proposals. §7’s claim of symmetry is therefore stronger than the text supports.
• Fix.
• For each of M3, M5 and M7, add one sentence on existing industry versions and what they lack (independence, mandate, access).
• Add a short Mirror paragraph applying Maynard’s tests to the labs’ and critics’ proposals.
• Soften §7’s “Symmetry” bullet accordingly.

11. MEDIUM: §6 reads as prescription (ll.191–207)#

• Problem. Items are written as imperatives (“Replace reliance on labels…”, “Assign containment levels…”, “Monitor who is exposed…”, “Take cheap, reversible steps…”). M8 (“Whoever uses an analogy… should say where it fails”) is labelled [Stated] high “as his method”, but it extends his practice into a rule for everyone.
• Fix. Recast each item as “His work points towards…”. Keep M1 and M2 as the report’s own designs, which they already are ([Inferred], low to medium). Label M8 [Stated] for his own practice and [Implied] for the extension to others. Nothing in M2 is in Maynard’s voice (no first person, no Substack register); the issue is stance.

12. MEDIUM-LOW: an unlabelled comparative judgement in §3.1 (l.97)#

• “Suleyman is closer to Maynard in method, Amodei on discontinuity” is an evaluation of Maynard’s position relative to others, with no label or basis. Fix: label it [Inferred], give the basis (Suleyman’s reasoning from the history of general-purpose technologies to institutions; Amodei’s “essentially unprecedented” set beside “defies analogy”, 2026-01-22) and a confidence (low-medium).

13. LOW-MEDIUM: D6 overstates Huang’s reliance on responsible firms (D6 l.125)#

• “A regime that relies on responsible firms is exposed to the next, less careful one.” Huang’s regime also relies on liability, sector regulators and audit. 02 §7.4 (point 2) notes that his likely answer to a less careful rival is to regulate that rival’s products, though he does not say so. Firms have acted unilaterally since July (03 §6.1 item 10). He has also said the labs are “extraordinary companies, and we ought to hold them to extraordinary standards” (All-In; E1).
• Fix. Write “a regime whose frontier gates rest on responsible firms”. Add the likely unstated answer and the unilateral actions.

14. LOW-MEDIUM: 03 and 04 are summarised more sharply than the body supports (Summary l.21; §4.3 ll.152, 159)#

• 03. The Summary calls 03’s judgement on toxicology “too quick” on the human side. But 03 §11.3 excludes “slow harms such as effects on skills and early-career work” from what can be rejected. §4.11 sets out “what does [transfer] in changed form” (the K7 property question). §12.2 asks for a property screen and for sorting acute from slow harms. M2’s body concedes most of this (“The comparison’s scope explains much of the difference”; “already transfers several of these concepts”). Fix: in the Summary, write “extends”, not “too quick”.
• 04. The heading “The acute-harm optimism in the AI-drafted article” does not match a hedged text. The article says “some of the ways AI goes wrong happen fast”, “In principle”, and “The catch is in that ‘in principle’”. Fix: retitle it “The fast-harm point in the AI-drafted article”.

15. LOW: “confirms” involves some circularity (§4.1 ll.137–141)#

• Several of the Late Lessons findings that Maynard’s work “confirms” (I5, K2) cite LL2-22, his own chapter, as evidence. The Summary acknowledges this; §4.1 does not. Fix: say “is consistent with (not independent confirmation of)” where LL2-22 is among the evidence.

16. LOW: minor accuracy and balance points#

• “Two out of three rights” (A4 l.107, §5 l.181, M2 l.193) comes from Lex Fridman, March 2026, not the interview. Say so where it is cited.
• The flip belongs at [1:16:05], not [48:58] (§5 l.179).
• “He accepts that large companies have shipped harmful products” (§3.1 l.95). Huang’s words are hedged: “Well, they have done it, maybe”. Quote the hedge.
• Concessions that cut the other way. Some of Huang’s concessions are reported without the qualifiers that 02 gives, and fairness runs in both directions.
• “I would absolutely add more regulation” [1:19:12] should keep “I don’t know what’s missing”. 02 also notes that he has opposed most specific new AI measures since 2025.
• The shutdown condition [36:44] should note that he expects it will not be met (“I am fairly certain they will say: Yes, they need to know how to solve this problem”), and that the “we” who would shut the labs down is unspecified.
• §4.4 (l.173). “A sound choice on grounds of conflict of interest” states the article’s reason as if known. Write “defensible given the conflict of interest”.
• §7 (l.214). Maynard also mentions Nvidia, descriptively, in 2025-02-23 (evo-2-dna-ai).
• Provenance line (l.3). For publication, “for his review” should read “and reviewed by him”, per the publication context.

What is sound and should be kept#

• Quotations. All 34 timestamped Huang quotations were found in the NYT text, and Klein’s “again and again in history” is quoted correctly. Speaker attributions are correct.
• Conditions and concessions. §3.1 gathers auditors, sector regulators, the shutdown condition, Nvidia’s own stop rule, “they see a lot more than I do” and “no relief from current ones”. D6 credits the moral-hazard argument and says that Maynard’s record does not engage it.
• Method. The three modes of transfer (§2.2) and the distinction between structural and literal analogy are handled rigorously. D3 is stated as “a divergence of doubt”, which is fair.
• Self-application. The critique is turned on Maynard’s own record: the kinder later verdict on nano (§2.7, D2), the mixed hindsight on LL2-22 (§2.3), and participation (§4.3: “the analysis challenges Maynard more than he challenges it”).
• Alignments. The alignments on alarm, novelty, evidence discipline, engineering controls and “no relief” are stated plainly, including the asymmetry in Maynard’s weighing (A2).
• Provenance rulings. These are observed: Garbee 2019 is treated as fully Maynard’s, the Abbott book is absent, [mixed] items are flagged, and co-authored items are weighted as shared.
• Clarifications. Maynard’s September 2026 clarifications are reflected (§2.1, §2.4).

INTERNAL (not for publication)#

Notes on M2’s own INTERNAL section, for the essay stage:

• “Clean illustration of structural transfer… against Huang’s literal calculator analogy” (l.238). Drop or reverse this. On education, Maynard and Huang agree that graduates need AI skills (issue 1). The contrast that holds is about category: what AI is, and whether lost capacities matter.
• “Software as a term of art” (l.239). Still usable, but only in its corrected form (issue 2). Huang does reason from behaviour. The question is whether familiar mechanisms imply that familiar controls suffice.
• “The 2016 ‘less responsible company’ line answers ‘CEOs with agency’” (l.240). It answers only part of it. Huang’s fuller answer includes liability and sector rules for the rival (issue 13). Maynard’s line is strongest where liability is weakest: latent, diffuse and third-party harm (issue 3).
• Possible question for Maynard. Would he see the “flip” [1:16:05], rather than the tenfold forecast, as the right bridge to his 1%-to-10% argument? And would he want the share independently accounted?
• Possible question for Maynard. Does he see his 2023 agreement that Howard’s concern about concentrated power “needs to be taken seriously” as common ground with Huang’s objection to the labs’ antitrust waiver?