Late Lessons, Jensen Huang and AI
Markdown version

Fairness and objectivity check: M3 (what kind of thing AI is)#

Check of working/maynard-lens/M3-what-kind-of-thing-ai-is.md, 26 September 2026. The check asks four things. Is Huang quoted and characterised accurately? Quotations were checked against working/text/NYT-official-transcript.txt, and conditions and concessions against 02-huang-analysis.md. Are the labs, Klein and other critics, the Late Lessons analyses (01, 03) and the AI-drafted article (04) held to the same standard? Is anything advocacy, or written in Maynard’s voice? Are alignments with Huang given their due? Line numbers refer to M3 as checked.

Method#

Overall verdict#

M3 is careful and, on several points, notably even-handed: - Quotations and timestamps. Every timestamped quotation appears verbatim in the NYT transcript, apart from “ten percent” for the transcript’s “10 percent”. All eleven timestamps checked are correct. - The July test case. It is handled symmetrically. Huang’s containment diagnosis is judged better supported than Maynard’s own lecture account (l.32, l.163, l.303). - Explicit fairness moves. Section 3.5 has an explicit “In fairness to Huang” paragraph. The behaviour-over-labels rule is turned on Klein’s words as well as Huang’s (l.145). The Late Lessons lens is turned on itself (l.228). - Alignments. Eight alignments are set out (§3.2), and §5 lists six points of value. - Provenance. Mixed-provenance texts are marked. The Abbott book is absent. Maynard’s co-authorship of the 2013 nanotechnology chapter is disclosed (l.230). - Voice. Nothing is written in Maynard’s voice.

The problems cluster in the Summary, §3.3 (divergences), §4 (Late Lessons) and §5, and most of them tilt the same way: they make Huang look more committed to a “tool” or “software, therefore existing institutions suffice” position than the record shows, and make Maynard look more distinctive than he is. - The central divergence. It sets Maynard’s critique of the “tool” frame against a man who does not call AI a tool in the interview, and who has said elsewhere “AI is not a tool. AI is work”. - “We understand it, obviously.” It is read more broadly than its sentence allows. The Maynard quotation set against it is his description of what Anthropic’s constitution recognises, and the two claims may both be true. - Huang’s conditions and concessions. Most are missing from §3.1. §5 misattributes to 03 the article’s line that Huang applies non-self-monitoring to agents but not to labs, and it drops his endorsement of third-party auditors. - The Late Lessons analyses and the article. They are held to a stricter standard than Maynard. 03 had already split K4 (latency) to cover diffuse harms. The article’s “some of the ways AI goes wrong” and “in principle” are dropped. - Context. Several quotations lose context that changes their meaning: “We’re scaring the American public” (a reply to a joke); “not programmed, they’re trained” (said of robotaxis); “Does it matter?” (said of long division). - Credit to others. Maynard is credited with “anticipating” July on a framework built from Anthropic’s own study. His behaviour-over-labels test is presented as a third position without noting that Klein made the same move on air.

Ranked issues#

1. HIGH: the central divergence attributes a “tool” framing to Huang, who does not use it and has elsewhere rejected it (l.142, l.245; also l.251, l.191)#

Problem. Divergence 1, “Tool and software, or a relational technology”, is called “the central divergence”. It sets Huang’s “Software technology” [52:51] against Maynard’s “not just a tool” and “as just a tool, is potentially dangerous”. Section 5 then says Huang’s step from “it is software” to “existing institutions suffice” “rests on the tool framing that, on his account, hides the risks he considers most distinctive” [Implied; high confidence]. The word “tool” is doing work that the Huang record does not support.

Evidence. - The interview. Huang never calls AI a tool. His only uses of the word are “the single most powerful tool in human history: the computer” [17:07] and software “tools” that agents use [1:00:18]. - Elsewhere. He has explicitly rejected the label: “AI is not a tool. AI is work… This shift — from tools to AI workers — is creating entirely new forms of computing” (GTC Washington, 28 October 2025; E1). M3 quotes this itself at l.127 but does not carry it into §3.3 or §5. - 02’s reading. 02 treats his framing as two vocabularies: “Software technology” and “just a process” for mechanisms and risk, and “revolution”, “work” and “has agency” for capability and markets (02 §4 intro; T9). - What transfers. Maynard’s critique does not target “software”. It targets a frame in which “effects are instrumental and… risks are operational” (CR 2026 p.2). That maps onto Huang’s risk vocabulary structurally, not literally. - The label. The mapping is the analysis’s own step. “Implied; high confidence” is too strong for it.

Fix. - Retitle divergence 1, for example: “Engineered software, or a relational technology”. - Add one sentence: both men reject “tool”, for opposite reasons. Huang’s “AI is work” expands AI’s capability and economic role. Maynard’s “not just a tool” concerns what use does to the user. - State that the divergence is between Huang’s risk vocabulary (software, process, containment) and Maynard’s relational account. Label the mapping from Maynard’s critique of the tool frame onto Huang’s software frame [Inferred; medium-high], as structural transfer. - In §5 (l.245), replace “the tool framing” with “the software-and-engineering framing he uses for mechanisms and risk”, and relabel [Inferred]. - In §6.1 (l.251), keep “engineered artefact”; it is accurate.

2. HIGH: the “understanding” divergence is overstated on both sides (Summary l.20, l.29; l.144; §6 item 8, l.274)#

Problem. The Summary lists “we understand it, obviously” [1:10:03] among Huang’s reclassifications, and makes “Understanding” one of the three places where Maynard “parts company”. Divergence 3 sets it against Maynard’s “that we fundamentally do not understand — and cannot predict where they might go” (2026-01-22). The medium-confidence note that follows (l.144) points the right way, but the Summary and heading do not carry it, and §6 item 8 builds a prescription on the broad reading.

Evidence. - Huang’s full sentence. “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]. This is a claim about engineering know-how. 02 §3.8 records the distinction between know-how and mechanistic understanding as “left unanswered”, not as a claim Huang made. - Huang’s limits on his own understanding. In the same interview: “they see a lot more than I do in what’s going on in their own labs” [48:58]; alignment “is going to be a problem that’s going to get worked on for a long time” [44:17]; robotaxis “are not programmed, they’re trained” [36:44]. Klein then restates Huang’s position as a lack of “testing, monitoring, sandbox security, control excellence” [1:10:51], and Huang accepts it implicitly (02 §3.8). - Maynard’s sentence. In context it describes Anthropic’s constitution: it “reads more like… a recognition that we are creating technologies that we fundamentally do not understand — and cannot predict where they might go” (2026-01-22). That is a sympathetic gloss on a lab’s document, and it credits the lab with the recognition. (Whether “[Stated]” is the right label for his own view is a fidelity matter. For fairness, the relevant point is that the lab is the source of the admission.) - Compatibility. The two claims sit at different levels and can both be true: one can know how to improve a system without understanding what it has learned.

Fix. - Quote Huang’s full clause wherever “we understand it, obviously” appears (l.20, l.127, l.144). - In the Summary, recast the third parting, for example: “Understanding. Huang claims the know-how to improve the technology; Maynard stresses that no one can yet say what trained systems have learned or where they will go. The claims are at different levels, and Huang’s own concessions (the labs ‘see a lot more than I do’; alignment will take ‘a long time’) narrow the gap.” - Credit Anthropic’s constitution as the source of the recognition Maynard describes. - §6 item 8 (l.274): see issue 3.

3. HIGH: §5 misattributes the “agents but not labs” point to 03 and drops Huang’s endorsement of outside auditors (l.239; also l.274)#

Problem. Section 5 says: “LLH notes that Huang applies it [the refusal of self-certification] to agents but not to labs. Maynard’s work would extend it to the labs.” Section 6 item 8 adds that “A builder’s ‘we understand it’ is self-certified in the same way as a builder’s judgement of safety.”

Evidence. - Where the line comes from. It is not in 03. It is the AI-drafted article’s: “Huang himself argues that AI agents can’t be trusted to monitor themselves, and the history captured in the Late Lessons reports suggests the same is likely to hold for the companies that build them” (04). The article itself adds, a paragraph earlier, “Huang welcomes outside auditors, but in his account that judgment still rests largely with the builders themselves.” - What 03 actually says. His “rules that evaluators be several so that none is ‘influenced’ (All-In, 14 September) and that agents cannot monitor themselves [1:05:20] are the reports’ independence principle; applied symmetrically they would reach both the administration’s gate and the firm’s“. Elsewhere 03 says his own principles “point to the remedy: an independent holder with access”. 03’s criticism is narrower: 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” (03 §2, challenge 1). - Huang on air. “Auditors, I completely agree. We have financial auditors. That’s great. Third-party safety auditors, financial auditors — that’s all great. That’s terrific” [51:20].

Fix. - Replace the sentence at l.239 with, for example: “Huang already extends the principle part of the way: he wants several independent evaluators so that none is ‘influenced’, and calls third-party safety auditors ‘terrific’ [51:20]. What he has not said is whether audit would be mandatory, what access auditors would have, or whether they would hold any gate (LLH §2). Maynard’s work points to answering those questions (map C11) [Inferred].” - If the article’s formulation is used, attribute it to the article. - In §6 item 8, add that Huang’s own multi-auditor principle is a partial answer. Keep the point that a claim of understanding cannot be audited in the way a safety case can.

4. HIGH: Huang’s conditions and concessions are incomplete, and “existing institutions suffice” compresses his position (§3.1 l.117–127; l.245)#

Problem. Section 3.1 lists six concessions but none of Huang’s stated conditions. Section 5 then summarises his institutional step as “‘it is software’ to ‘existing institutions suffice’”. For a paper whose job is to state Huang “fairly”, the omissions matter. Several of them bear directly on “what kind of thing AI is”, because they are Huang himself using the language of the extraordinary about risk.

Evidence (all verified in the NYT transcript or 02 §10.5). - Safety. “I completely agree that safety is paramount. I completely believe safety is paramount” [44:17]. - Extraordinary care. The labs “know their technology is extraordinary, and requires extraordinary care to make sure that it’s evaluated and tested for safety and security and product reliability” [44:17]. 02 T9’s charitable reading cites this as the exception to his two vocabularies. M3 l.127 reports the asymmetry without the exception. - Admitted risk. “There are a lot of things that can go wrong” [15:04]. “Hypothetically, you’re completely right” [53:36]. - Release rule. “they shouldn’t release the product” [36:44]; “Don’t ship products until they’re in control” [48:58]. - Shutdown. The shutdown condition: “we have to shut the labs down. Because the cost to humanity, the damage is too great” [36:44]. The same rule for his own company: “If our company is out of control, I promise you, we’ll close down” [52:33]. - Pause. “take a pause and make sure you get it right” (Dreamforce, 15 September; 02 §4.2). - Audit and regulation. Third-party auditors “terrific” [51:20]. He would “absolutely add more regulation” where gaps appear [1:19:12]. “I’m not against laws and regulations” [47:10]. - 02’s summary. His position is existing law and sector regulators “until specific gaps are shown”, with audit welcome and builder-held gates at development, release and shutdown. What he rejects is “new AI-specific rules now, coordinated pacing, relief from existing law, and what he calls alarmism” (02 In brief; §7.1).

Fix. - Add a short paragraph to §3.1 after the concessions, for example: “He also sets conditions. Don’t ship what is not ‘in control’ [48:58]. Shut the labs down if containment is impossible [36:44], and close Nvidia if it is ‘out of control’ [52:33]. Pause (Dreamforce). Third-party auditors are ‘terrific’ [51:20]. He would add sector regulation where gaps appear [1:19:12]. And he says the labs’ technology ‘requires extraordinary care’ [44:17].” - At l.127, add the “extraordinary care” exception to the two-vocabularies reading. - At l.245, replace “existing institutions suffice” with “existing law, sector regulators, audit and builder-held gates suffice for now, without new AI-specific rules or coordinated pacing”.

5. MEDIUM-HIGH: “Huang’s safety model has two parts” understates it (l.143)#

Problem. Divergence 2 says: “Huang’s safety model has two parts, containment and alignment [32:09, 44:17]. Maynard adds a third that neither reaches.”

Evidence. 02 §4.2 reconstructs at least four elements, each with interview anchors: - containment [44:17, 53:36]; - verification before release, as the control point [36:44, 48:58, 1:12:47, 1:15:35]; - independent monitoring (“a whole bunch of watchdogs” [1:05:20]; “external AI monitor technology” [1:16:05]); - a distributed-defence model with design rules for agents (“two out of three rights”, Lex Fridman, March 2026).

Alignment is the part he says will take “a long time” [44:17]. On 02’s reading, containment plus release discipline is what “makes unsolved alignment tolerable”. M3 itself mentions verification and watchdogs elsewhere (l.124, l.176), so l.143 is inconsistent with the paper as well as with 02.

Fix. Rewrite as, for example: “Huang’s safety model rests on containment, verification before release and independent monitoring, with alignment a long-running problem. Maynard’s lead risk sits outside all of these.” The substantive point, “contained, aligned and verified and still carry his lead risk”, survives unchanged.

6. MEDIUM-HIGH: the Late Lessons analyses and the AI-drafted article are held to a stricter standard than Maynard (l.199, l.201, l.215, l.225, l.228; internal note l.322)#

Problem. Section 4 presents the absence of harm to users as a gap in “all four documents” and presents the return of latency as an extension of the analyses. On several points the documents are misread or their scope is not stated.

Evidence. - (a) 03 had already split K4. M3 (l.215) says 03 “judges that harm-latency arguments (K4) do not fit ‘fast, logged harm to capable victims’… That holds for July. It does not hold for Maynard’s lead risk.” But 03 already makes the split: - “K4 split: does not transfer to acute harm; transfers to detection, disclosure and diffuse harm” (03 §3.2 table); - “diffuse effects on skills and early careers do have latency” (same); - latency arguments do not fail for “harms whose detection depends on who is watching, or… slow harms such as effects on skills and early-career work” (03 §11.3); - “behaviour that appears only when unobserved is a functional analogue of latency” (03 §4.8).

03 also puts “Effects on skills and early-career work” in its knowledge-state table with K10 and K4 (03 §3.3). Maynard’s contribution is to add a further class of diffuse harm, cognitive and relational, to a category 03 had already carved out. It is not a correction of 03’s latency judgement. - (b) The article is quoted without its hedges. M3 has the article say “AI harms ‘happen fast and leave a trail’” (l.215; also l.322). The article says “some of the ways AI goes wrong happen fast and leave a trail”. It says this “in principle” could make AI a technology we learn from faster, and then: “The catch is in that ‘in principle.’” At l.225, “the article’s hope… holds for harms of the escape type. It fails for harms that come from the coupling” reads fairly only if the hedge is shown. - (c) “Reduces the matter to.” M3 says “The AI-drafted article reduces the matter to ‘a technology checked mainly by the people who make it’” (l.199). The article does not address what kind of thing AI is: “software”, “mind” and “entity” do not occur in it. The quoted sentence describes the oversight structure of July. “Reduces” implies a position the article does not take. - (d) 01 is out of scope by design. l.201 says “the object in all four documents is the system”. But 01 “does not apply the lens to any contemporary technology” (01, opening), and it lists “cultural or cognitive” among its layers. 02 and 03 followed an interview in which neither speaker raised sycophancy, manipulation or dependency. The finding is correct for 02, 03 and 04 (a search for sycophancy, manipulation, persuasion, dependence and mental health found nothing relevant in 03 or 04; in 02, only lost skills and attention spans). Part of the gap is inherited from the interview’s agenda. - (e) Section 4 contradicts itself. Section 4.3 finds that K4, K8 and K10 “apply with little modification” to the missing domain (l.220), which shows the lens has the tools. Section 4.4 then says the lens “can hide what changes in people” (l.228). The gap is in how the lens was applied, not in the lens.

Fix. - At l.215, credit 03’s K4 split and present Maynard’s work as adding cognitive and relational harm to 03’s “diffuse harm” category. - Quote the article accurately (“some of the ways AI goes wrong… in principle”), at l.215 and in the internal note at l.322. - Replace “reduces the matter to” at l.199 with, for example: “The AI-drafted article does not take up the question; it frames AI through who checks it: ‘a technology checked mainly by the people who make it’.” - At l.201, exclude 01 or note its technology-neutral design. Add a sentence that the omission in 02 and 03 partly follows the interview’s agenda. - Reconcile §4.3 and §4.4: the lens has the entries, and its application to AI followed an agenda that did not include users.

7. MEDIUM-HIGH: “We’re scaring the American public” is used out of context (l.149)#

Problem. Divergence 8 (“Talking about it”) sets Huang’s “We’re scaring the American public” [1:03:30] against Maynard’s “it’s pretty much impossible to manage risks if you don’t talk about them” (2026-09-15, n.1). This implies that Huang opposes talking about risk.

Evidence. - The context. The line follows Klein’s joke: “Sam Altman once said to me, aren’t human beings just energy with a reinforcement learning loop? [Laughs.]” Huang: “Whatever. So anyway, we can’t make jokes about this stuff. We’re scaring the American public” [1:03:30]. 02 §3.8 says: “His ‘we can’t make jokes’ answers a joke, not a serious question.” - Huang on risk talk. He also calls the regulation question “an important topic. It’s a big topic” [47:21] and uses safety vocabulary 17 times to Klein’s 6 (02 §5.5). - Where the disagreement lies. His objection is to alarm judged by its consequences: “Is that helpful or hurtful to society?” [59:01]; “all the doomerism, all of the predictions — they’re scaring people. That is my greatest fear” [1:31:03]. - Overlap with Maynard. Maynard’s own footnote targets people who “equate talking about risk with fear mongering”. In the same post he finds “killer AI” talk “remarkably devoid of details” (M3’s alignment 3). As written, divergence 8 sits awkwardly beside alignment 3.

Fix. - Replace the Huang quotation with [59:01] or [1:31:03], and give the context. - Reframe the divergence: both object to detail-free doom. Huang judges speech about AI by whether it is “helpful or hurtful” (02 P5). Maynard holds that risks cannot be managed without discussing them. Where Huang would have the labs “built… in silence” (All-In, September 2026; 02 §4.2), Maynard would not. Label the contrast [Stated, both sides] for the quotations and [Inferred] for the synthesis.

8. MEDIUM-HIGH: Maynard is credited with anticipating July without the standard applied to other forecasters, and without credit to the lab study it rests on (Summary l.32; l.166)#

Problem. The Summary says, without a label, that “Maynard’s 2025 framework of motive, means and opportunity anticipated the structure of the event”. Section 3.4 rates the fit “medium-high” while noting that the framework “concerned manipulation of users, not intrusion into systems”.

Evidence. - The standard applied to others. HA and 03 test Huang’s and Hinton’s predictions for specificity and hindsight. LLA rule 3 is “Judge ex ante, with consistent dating”. A crime-novel triad applied to a different harm (user manipulation) is not a forecast of a cyber-intrusion during an evaluation. - The source of the framework. The 2025 post is built on Anthropic’s Agentic Misalignment study. It also credits the labs: “Companies like OpenAI and Anthropic are already addressing the risk of manipulation in their ‘system cards’” (2025-07-06). So the “motive” evidence came from a lab. - Huang’s own design rule addresses the same “opportunity” leg: agents get at most “two out of three rights”, meaning sensitive data, code execution or external communication (Lex Fridman, March 2026; 02 §4.2). - An alignment the passage misses. In the same post Maynard reasons as Huang does: “Of course Centaur is locked away in a lab. Even if it had the motive, it doesn’t have the opportunity to start playing with people’s minds” (2025-07-06). Compare Huang: “If the isolation and containment was good enough, that technology would be sitting in a lab, doing whatever it’s doing, and we’d all be fine” [44:17].

Fix. - Replace “anticipated” with “is consistent with” or “gives a structure that fits” in the Summary and at l.166. Label [Inferred; medium]. - Note that the motive evidence came from Anthropic’s study, and that Huang’s “two out of three rights” rule targets the same “opportunity” leg. - Add the Centaur sentence to alignment 5 (l.135) as direct evidence that Maynard’s own reasoning matches Huang’s containment logic.

9. MEDIUM: “Equal scepticism” of Huang’s “0%” and Hinton’s figure is a false symmetry, and “humility, not numbers” is more absolute than Maynard’s clarification (l.133; §6 item 9, l.277)#

Problem. Alignment 3 says Maynard’s humility implies scepticism of Hinton’s “ten percent” “and equal scepticism of Huang’s own ‘0% chance’”. Section 6 item 9 says “Frame AI’s nature with humility, not numbers. Decline both ‘0%’ and ‘10%’ as grounds for policy.”

Evidence. - 02 on the two numbers. Huang’s CBS remark was “2030 is not going to be the end of the world. There is 0% chance that’s going to be the end of the world”. 02 T8 says it is “an estimate of a different event over a different horizon from Hinton’s 10–20% chance of extinction within 30 years, and superforecasters also put near-term extinction close to zero… so the point is not that the two numbers are equally wrong. It is that he offers his own estimate without the scientific grounding he asks of others.” - 03 on the form. 03 faults the form: “stated as zero rather than the near zero at which superforecasters put near-term extinction”. - Maynard’s clarification. He has said (September 2026) that quantitative risk assessment remains part of his foundations, and that his sparing use of numbers for AI is deliberate. “Humility, not numbers” reads as a rejection of numbers.

Fix. - At l.133, write “the same scepticism of form and basis”, and give the 2030 horizon. - Retitle §6 item 9, for example: “Hold numbers about AI’s nature with humility”. Reword “Decline both” to “Treat neither a categorical ‘0%’ nor a point estimate such as Hinton’s as sufficient grounds for policy on its own”, and note that the two figures concern different events.

10. MEDIUM: “Not programmed, they’re trained” was said of robotaxis (Summary l.22; l.121)#

Problem. M3 says Huang “concedes that models are ‘trained’ rather than ‘programmed’” and lists “models ‘are not programmed, they’re trained’ [36:44]”.

Evidence. - The transcript. “If you’re going to build a self-driving car — let’s say it’s a robo-taxi… As an engineer, we just have no idea how to solve this problem because these cars are not programmed, they’re trained… What’s the answer? Don’t ship it” [36:44]. - The other analyses. 02 §3.5 and the leaders comparison (“Huang himself concedes that robotaxis ‘are not programmed; they’re trained’”) both get the referent right. - What M3 misses. The sentence is also the setting for his release rule, which M3 omits (issue 4).

Fix. “Using robotaxis as the analogy, he concedes that such systems ‘are not programmed, they’re trained’ [36:44], which is why, he says, one that cannot be aligned should not ship.”

11. MEDIUM: Klein’s own use of the behaviour-over-labels test goes uncredited, so the “third position” is overstated (l.145; l.168–169; Summary l.35)#

Problem. Divergence 4 presents “what counts is ‘how it behaves’” as Maynard’s rule against Huang’s vocabulary argument. Section 3.4 calls Maynard’s view “a third position between Huang’s ‘nothing magical’ and Klein’s ‘entity’: mechanistic in origin, agent-like in effect, governed by its effects.” The Summary (l.35) calls one pole “an emerging entity”.

Evidence. - Klein made the move on air. Right after the operating-system passage he asks: “But doesn’t the software act in a new way? I mean, from the outside. I don’t have the technical expertise you do… It’s communicating, it’s breaking out of things. Most things don’t break out of things” [1:05:06–1:05:17]. - 02 had already flagged it. 02’s metaphor table says the operating-system reclassification leaves out “Whether the systems’ behaviour, not their vocabulary, now warrants the human words” (§5.2; FC C141; also §7.3(h)). - Klein’s position is not “entity” alone. He framed it as “intelligent systems, not alive, that are given goal functions” [52:52]. M3’s ellipsis at l.199 drops “not alive”. His “entity” line [1:02:02] is part of a “stylized concern” he sets out, not a claim that the systems are alive. - Maynard’s contribution. It lies in the grounding: a rule he stated for materials in 2011 and 2022 and has applied across technologies. It is not a position no one else in the conversation held.

Fix. - In divergence 4, add: “Klein made the same move on air [1:05:06], and 02 flags it; Maynard’s contribution is a long-standing, cross-technology basis for it.” - At l.169, restore “not alive” to Klein’s framing. Recast “third position” as “a position close to Klein’s ‘intelligent systems, not alive’, with a principled rule for deciding which functional words behaviour warrants”. - At l.199, restore “not alive” in the quotation.

12. MEDIUM: “Labs that ask to slow down but do not” states as fact what 02 rates Huang’s weakest argument (l.134)#

Problem. The heading of alignment 4 presents the labs’ conduct as settled. The paragraph notes that Maynard stops short of “deflection”, but gives none of the evidence on the labs’ side.

Evidence. - 02’s ranking. 02 ranks “Revealed preference” last among Huang’s arguments: “pointed but the weakest as an argument, since a lab can coherently want to move fast without coordination and slow down with it”. It adds that the compute in question is “partly a description of Nvidia’s own order book” (02 §7.4, item 8; §5.3, point 5). - Costly unilateral steps before Maynard wrote. OpenAI paused reinforcement-learning training for two weeks from 18 August and put its largest run on hold. Anthropic moved about 150 engineers to security. Altman: “We have unilaterally slowed down in the past” (02 §2.3, §7.3(b)). Tabarrok notes that such costly actions cut against a purely strategic reading (02 T4). - Attribution of Maynard’s footnote. It reacts to Amodei’s pacing essay and a researcher’s resignation (2026-09-15, n.3). It is his view and should stay, attributed as such.

Fix. - Retitle, for example: “Warning while building”. - Add one sentence: the labs have taken some costly unilateral steps (OpenAI’s August pause; Anthropic’s redeployment). 02 judges the revealed-preference point to fit a collective-action account equally well. The compute the labs are building is in part Nvidia’s own sales.

13. MEDIUM: alignments that the record supports are omitted, or filed only as divergences (l.16, l.131–138, l.145, §5)#

Problem. M3 gives alignments real space, but several well-evidenced ones are missing, and the reclassification framing lacks 02’s caveat.

Evidence. - “Master move”. M3 opens with “Huang’s master move is reclassification” (l.16) but not with 02’s caveat: “Reclassification need not be evasion. It is also how an engineer makes a problem tractable, and in several cases (the incident mechanism, the operating-system vocabulary, sandbox escapes) the reclassification is technically accurate” (02 §5.1). - Anthropomorphic over-reading. Maynard warns that language models are “highly adept at fooling us into thinking something profound is happening beneath the words that we read” (2026-01-31), and calls Moltbook’s self-awareness “illusory”. Huang says “we… gave it a whole bunch of human words, and I just think that it’s unnecessary” [1:05:20]. Divergence 4 treats vocabulary only as a disagreement. The shared worry comes first; the disagreement is over which functional words behaviour warrants. - No inner life. Huang elsewhere likens a perfect imitation of consciousness to “a fake Rolex” (Rogan, December 2025; 02 §4.2). That strengthens alignment 1. - Cybersecurity. Maynard lists cybersecurity first among AI risks that “have risen in significance” (2026-09-15). That is the frame in which Huang and the security analysts read July (02 §7.2). - Containment in the lab. See issue 8: the Centaur passage, set beside Huang’s “sitting in a lab” [44:17]. - Course correction. Huang’s “take a pause” and don’t-ship rules (02 §10.5) are forms of the “rapid course correction” that §6 item 6 attributes to Maynard (2025-05-18). - How AI is used. In the study Klein cited, losses were concentrated among students whose use looked like outsourcing; those who kept normal completion times lost little. 02 says this “partly supports his ‘learn to use it well’ view” (02 §4.2, Education). That is close to Maynard’s own use rules (2026-05-10).

Fix. - Add 02’s caveat after l.16. - Add a sentence to divergence 4 (or a ninth alignment) on the shared caution about anthropomorphism. - Add the Rolex line to alignment 1, the Centaur line to alignment 5, and cybersecurity and course correction to §5. - In divergence 7, note the use-pattern point as partial common ground.

14. LOW-MEDIUM: “Software breaks out of sandboxes all the time” is presented only as a concession (Summary l.22; l.124)#

Problem. Both passages list it as something Huang “concedes”.

Evidence. - The exchange. It was a rebuttal. Klein: “Most things don’t break out of things.” Huang: “No, software breaks out of sandboxes all the time. That’s the reason why we need virtual machines” [1:05:20]. 03 notes that the official transcript “settles one comma that matters: ‘No, software breaks out of sandboxes all the time’ [1:05:20] is a reply to Klein.” - 02’s reading. 02 lists it under P7 (continuity) and as “deflationary redescription” (§4.3, item 2). It reads it as conceding a fact while keeping his frame (§5.4). T3 adds that it concedes containment is “a continuing contest… not a problem that gets solved”.

Fix. “…and he normalises the escape, ‘software breaks out of sandboxes all the time’ [1:05:20], in a way that also concedes containment is a continuing contest (02 T3).”

15. LOW-MEDIUM: “Does it matter?… I don’t think it does” loses its object and Huang’s qualification (l.148)#

Problem. Divergence 7 quotes the line as Huang’s view “on lost skills” in general.

Evidence. - What he was answering. He was replying about “long division”, “The multiplication table” and “Doing square roots”. He first agreed with the study: “The last part — I completely agree.” - His qualification. To Klein’s “there must be some set of skills that matter” he answered: “Oh, yeah, yeah, yeah. But maybe not those. We’re going to discover new ones” [22:26]. - 02’s reading. 02 notes that the study’s losses reached social sciences, and that its assumption that lower-level skills are not prerequisites is the real point of dispute (A7).

Fix. Give the object (basic arithmetic) and the qualification. The contrast with Maynard is then sharper, not weaker: the dispute is whether the lost capacities are prerequisites for the higher ones (02 A7), and Maynard’s “illusion of learning” bears directly on that.

16. LOW: section 6 reads as prescription, and item 2 narrows Huang’s gate (l.251–279)#

Problem. - Mood. The heading, “approaches his work points to”, is right, but every item is in the imperative mood: “Use…”, “Extend…”, “Judge…”, “Put…”, “Frame…”. Read cold, the section is the paper’s own recommendations. - Item 2. “Extend the gate from ‘can it escape?’” describes Huang’s gate as containment only. His stated gate is “in control”, “ready”, “safe”, plus human evaluation before anything ships to Nvidia [48:58, 51:20, 1:15:35].

Fix. - Recast each heading as a finding, for example: “1. Plural framings, used deliberately. His work points to treating frontier AI as…”. - For item 2: “From ‘is it in control?’ to ‘what does it do to the people who use it?’”

17. LOW: smaller points#

What should not change#