Late Lessons, Jensen Huang and AI
Markdown version

Red team B (Late Lessons’ advocate): review of D05, “Innovation, infrastructure, trajectory and lock-in”#

Reviewed file: working/synthesis/dimensions/D05-innovation-infrastructure-lockin.md (382 lines). Written 26 September 2026.

Remit. This review looks for places where D05 is too credulous towards Huang or too quick to set Late Lessons aside. It checks five things: framings accepted at face value; lens patterns that are present but not recorded; false balance in the Mirror lines; disanalogies treated as settled; and close Late Lessons analogues that D05 leaves out. It does not reargue points where D05 is already sound (see the end). Every fix keeps the project rules: Mirror questions, weighting by case type, ex ante dating, no bad faith without documents, and a flag on LL2-22. The fixes are analytical only. None adds article angles or project commentary to D05.

Quote check. The Huang quotations in D05 match the transcript at the timestamps given, with stutters removed, except for the following:

The Late Lessons quotations checked against the text extracts all match: LL2-03, pp. 47, 52, 53, 54, 55, 56, 60, 64; LL2-28, pp. 672, 673; LL2-06, pp. 149–150.

Ranked issues#

1. Huang’s car analogy leads straight to the reports’ own case, and D05 does not follow it there. Severity: high#

Location. - §3.2 (lines 102–104): the leaded-petrol arc is told but never applied to Huang. - §4.10 (line 263): the car analogy appears only through the mandates in C163. - §5; §6, item 9. - There is no §4 entry.

Problem. At [1:16:05] Huang says: “If I were in the car industry a hundred years ago, I would rather the car industry accelerated to today in one year, because I believe today’s car is way more safe”, ending with “Accelerate the living daylights out of that”. The car industry of a hundred years ago is the industry of LL2-03. In May 1925 the Surgeon General’s conference gave tetraethyl lead (TEL) its “one day trial” (LL2-03, p. 52). The chapter shows three things about speeding that industry up. - Acceleration was the mechanism of harm. Henderson warned that leaded petrol “will be in nearly universal use … before the public and the government awakens to the situation” (p. 52). - Capital commitment changed what the developers said. Midgley and Kettering had championed alcohol fuel. Once they had invested in TEL production, they “categorically denied the existence of alternatives”, and TEL became “the only material available” (p. 54). By the 1960s the alternatives were “forgotten” (p. 55). - The safe car did not come from speed. GM fitted catalytic converters “in order to comply with the Clean Air Act of 1970”, and lead left because it poisoned them (p. 60). In Europe the exit came from forest-dieback policy, and the chapter says it was “pure chance” that the two concerns coincided (p. 64).

Huang’s counterfactual assumes the destination is fixed whatever the path. That is the assumption L4 challenges: once a commitment is made it is reinforced “even if markedly inferior to potential alternatives” (LL1-16, p. 177).

The framing also recurs: - 1925 industry: TEL was “essential to the industrial progress of America” and “all innovation entails risks” (p. 52); restrictions would stop progress “if we are to survive among the nations” (p. 53). - Huang: “I want to see us not ruin the opportunity for the United States” [1:31:03]; “the world to be built on the American tech stack” [1:35:15]. - These are lens entries M4 and M7. They are evidence of framing, not of its effect (M4 limits).

Evidence and weight. - LL2-03 was written by protagonists. Its hindsight verdict is “core strengthened; specifics wrong”. - Analysis: in 1925 the effect of low-level public exposure was uncertain. The episode is closer to [U] than to the [K] tag that attaches to lead after the 1960s (01 §6.2). - Disanalogy: TEL was a single toxic additive, and AI is a general-purpose capability. What transfers is path dependence and the role of mandates, not toxicity. - Mirror: Huang’s point that delaying protective technology has victims (“A lot fewer children would have been killed”) is C8’s reverse. It holds in direction. But in LL2-03 the protective technology itself, the catalytic converter, spread by mandate.

Fix. - Add LL2-03 as the closest analogue, either in a new §4 entry (“Accelerating to a fixed destination”) or in §4.4 and §4.10. - Record L4 and L6 as present, with M4 and M7 as framing only. - Add the case to §5. It is the reports’ direct test of the proposition Huang states.

2. D05 credits Huang with support for diversity by misreading “every layer to win”. It omits his stated aim of a single global stack and the Hugging Face purchase. Severity: high#

Location. §1, finding 4 (line 33); §2.4 (line 58); §4.11 (lines 273–283); §6, items 4 and 6 (lines 325, 327); record table, row P5.

Problem. - The quotation’s context. “We want every single layer to win” [1:37:36] comes straight after “Nvidia is an American company. We should benefit America first” and “Vera Rubin goes to the Frontier Labs first”. It ends “then we need every single layer to go out there and compete for the market”. It means American firms at every layer winning markets abroad. It does not mean plural competition within a layer. - “We don’t pick winners” (CNBC) means supporting every customer, and every customer runs on Nvidia. - Huang’s own boasts are about near-universality. “Every AI lab, every AI model, closed model runs on Nvidia” [1:21:05]. His strategic aim is “for the world to be built on the American tech stack. Just as… the world is built on the U.S. dollar” [1:35:15]. - LL1-16 warns against almost exactly this. Surprises are smaller with “several competing technologies… rather than just one, global, near monopoly” (p. 187). - The reports tie diversity to exits. “Keeping options open and following multiple paths means that a particular option can be terminated if it turns out to pose high risks.” They also warn of “technological monopolies, driven by the high costs of research, development and production” (LL2-28, p. 673). That forecast was borne out for GM seed and agrochemicals, where the European Commission found reduced innovation competition among “only five” integrated players (hindsight LL2-28; T04 P5; independent of LL2-22). - L4 asks whether a technology is “sold as an integrated proprietary system whose use by some compels adoption by others”. The documented record fits: - Nvidia sells integrated “AI factory” systems. - It finances customers across layers. - Competition regulators in the EU, US, UK, China and South Korea have asked about its investments in and agreements with foundation-model developers (10-Q; Huang analysis §2.2). - The Hugging Face purchase. On 2 September Nvidia agreed to buy Hugging Face, the main hub for open-weight models (8-K). The model-layer diversity D05 credits him for is being acquired by the dominant supplier in the layer below. That is the GM pattern of integration across layers. D05 mentions the purchase only in §8. - §6, item 6 credits his objection to incumbents entrenching themselves through coordination (I9). It does not record the mirror case: the chip-layer incumbent entrenching itself through investment and acquisition.

Mirror (kept). - Nvidia’s performance advantages are real, and its share is not total (FC C173). - Huang promised that “NVIDIA compute will not be required” on Hugging Face. - USDA found seed-industry consolidation accompanied by more private R&D (hindsight LL2-28). - A pacing regime among a few labs would also concentrate.

Fix. - Re-score P5: present at the chip and platform layer (documented structure, inferred consequences); partly met at the model layer. - Correct the reading of [1:37:36] in §2.4, §4.11 and §6, item 4. - Add the Hugging Face purchase and the five competition inquiries to §4.11. - Link the diversity point to finding 2, since the reports’ reason for diversity is that an option “can be terminated”. - Keep the † on the toolkit half of T04 P5.

3. The verification “flip” is credited as the rebalancing the reports asked for, but the reports’ complaint was about research controlled by producers. Severity: high#

Location. §1, finding 4 (line 33); §6, item 5 (line 326); §4.9 Evidence (line 251: “He pushes towards… verification”).

Problem. D05 calls the 80/20 flip “a producer-side, testable version of the same complaint” (I3). Being producer-side is what the reports objected to. - The TEL precedent. After 1926 the Surgeon General’s committee “strongly urged” long-term prospective studies of consequences “not now foreseen”. They were never funded. “For the next 40 years all studies of TEL were conducted and funded by the Ethyl Corporation and GM”. Ethyl’s voluntary compliance with a 3 cc limit was “relieving the government of any pressure to introduce the regulations” (LL2-03, p. 56). - The GM crops precedent. Developer-controlled, underpowered studies tilt towards “no evidence of harm” (LL2-19, Box 19.2). Hindsight strengthened this: a funding effect appears in 40% of papers on Bt crop efficacy. - What the reports ask. I3’s Ask is “What share of public research goes to understanding harms rather than developing products?”, and T2 asks who produces the evidence. - What Huang offers. - A remedy that runs through more product: “I want them to get more compute, but allocated towards evaluation” [1:16:05]; development compute up “by a factor of ten” [48:58]. - A forecast rather than a commitment: “I think they’re doing that” [1:16:05]. - A figure (80/20) that is his own estimate and that no one is obliged to disclose, so “testable” is doubtful. - Where D05 is inconsistent. It applies G9 elsewhere (OpenAI’s 20% compute pledge was not delivered, C161) but not here. G2 asks whether commitments are “backed by enforcement, measurement, funding and deadlines”. - Steering. As steering (L6: “what is mandated or funded”), Nvidia’s roughly $100bn goes to compute, customers and energy. No Nvidia funding of independent verification is documented. “Pushes towards verification” describes rhetoric, not capital.

Mirror. The labs also do their safety research in house. But the critics’ proposals include independence: embedded third-party evaluators (Amodei) and government tools (the pacing statement). Huang deserves credit for endorsing third-party auditors [51:20], which is the independent half.

Fix. - Recast §6, item 5 as a half-match: rebalancing, yes; independence, only through his endorsement of auditors. - Add the G2 and G9 checks, and the LL2-03 p. 56 precedent. - In §4.9, separate rhetorical steering from capital steering. - Cite [1:16:05] and [48:58] as in-interview evidence that “nearly every remedy runs through more compute”. Line 251 currently cites only Lex Fridman.

4. D05 quotes the “moving target” but never applies it, and the “fast feedback, patchable software” concession is left unconditional. Severity: high#

Location. §3.2 (line 102, quoted); §3.4, items 2–3 (lines 119–120); §1, finding 4 (line 33, last sentence); §4.1 (speed of iteration); §4.3 Transfer (line 166); §6, item 3 (line 324).

Problem. LL2-28 names three features of contemporary technology that worsen delay. D05 uses sunk investment and scale. It drops the first: “by the time evidence of harm is confirmed, the technology has often changed, leading to assumptions that, unlike yesterday’s technology, today’s technology is now safe” (p. 672). This is K11’s moving-target clause.

The iteration speed Huang prizes is what drives it: - “What used to take a year to pretrain something now takes several hours… the loop is going faster” [1:12:47]. - “In the last six months, AI went from… interesting to useful” [44:17]. - “Their next implementation of their sandbox is going to be much better” [32:09]. - The Astra system card calls the new model “better aligned than GPT-5.6 Sol” (Huang analysis §2.3).

Fast iteration shortens the time to a fix. It also sends each observed harm to a superseded version. Two documented facts show that the fix has not been shown to carry across generations: - Anthropic found newer models “still engage in the same behaviors at concerning rates”. - Anthropic “could not identify a single root cause”.

Detection speed is also weaker than the disanalogy assumes: - Hugging Face, the victim, detected the intrusion before OpenAI connected it to its own agents. - A June breach of a government site became public only in late September (post-recording).

Weight. K11 is strong for confirmed hazards ([K]) and moderate as a prior ([F]). But the moving-target mechanism is already observable in AI documents. It needs no extrapolation from chemicals.

Fix. - Record K11 (moving target) as present in §4.1 and §4.3. - Recast disanalogies 2 and 3 as conditional. Physical latency is short, but detection and attribution depend on victims and logs. Patchability is shown for containment, not for behaviour across generations. - Carry the qualifier into §1, finding 4 and §6, item 3.

5. “Benefits large, near and observable” is listed as a disanalogy, but LL2-03 treats it as part of the failure pattern. Severity: medium-high#

Location. §3.4, item 4 (line 121); §1, finding 4 by implication.

Problem. At the 1925 conference, “while there was solid and direct evidence of industrial benefits from TEL, evidence on health risks to the public was not available. This asymmetry between short-term economic benefits and long-term health hazards is another continuing problem” (LL2-03, p. 52). M5 asks directly: “Are the benefits conspicuous and the harms slow?” Near, observable benefit alongside slow, unobserved harm is the configuration the reports warn about. It is not a way out of it. The genuine disanalogy is narrower. AI’s benefits may be larger and broader than TEL’s anti-knock gain, and they fall to many users. That cuts both ways: a wider benefit makes lock-in stronger as well as more justified.

Fix. Rewrite item 4: “Benefits may be larger and broader than in the reports’ cases. But near, observable benefit set against slow, unobserved harm is itself the reports’ pattern (LL2-03, p. 52; M5), so this is a disanalogy of magnitude, not of kind.”

6. “Market-driven exits have precedent” misreads both precedents. Severity: medium-high#

Location. §6, item 9 (line 330); §4.9 (line 251, “co-driver”).

Problem. D05 cites catalytic converters as a market-driven exit. The chapter says GM installed them “to comply with the Clean Air Act of 1970” (LL2-03, p. 60). In Europe they arrived through Germany’s NOx objectives (p. 64). The co-driver was a regulatory mandate aimed at a different problem.

The sulphur precedent is mainly structural, not market-led clean investment. “Large reductions came only with economic collapse after 1990” in Central and Eastern Europe, and German cuts came mainly from closures in the east (hindsight LL1-10).

Neither precedent supports “You don’t need government subsidies… because the market forces are here” [1:40:15]. The reports’ own conclusion is that such exits are effective but fragile. The chapter asks what would have happened “if someone had invented a lead tolerant catalytic converter?” (p. 64).

Fix. Rewrite item 9: “Exits have ridden co-drivers: a regulatory mandate for another purpose (catalytic converters) and economic collapse (sulphur). The reports give no precedent for market growth alone driving a clean exit, and they rate co-driver exits fragile.”

7. Prior justification is read as support. The instrument is justification before use, which his diffusion model rejects. Severity: medium-high#

Location. §6, item 8 (line 329); K9 is absent throughout.

Problem. D05 says the disagreement with Huang “is about timing… not the layer”. But timing is the instrument. Prior justification “requires each use to be justified before exposure” (01 §6.12). It arose because recommendation-only rules allowed “ill-conceived” uses such as shoe-shop fluoroscopes (LL1-03, p. 34).

K9 (strong; [K] and [U] strong; the lesson with the widest case support) asks: “Are uses spreading into routine, prophylactic or trivial applications with no demonstrated benefit?” It notes that indeterminate uses were answered “by pre-market benefit assessment” (LL2-27, Table 27.1, p. 656). Huang’s model is the opposite: - “every single industry has to benefit” [1:31:03]; - “use the technology as quickly as you can” [17:07]; - “you can’t graduate without learning how to use an AI” [20:17].

The best benefit example in the interview, radiology, sits inside a pre-market regime: 76% of FDA-cleared AI devices are for radiology (FC C010). That supports the reports’ instrument at the application layer.

Hindsight adds a limit: collective radiation dose still rose with CT (hindsight LL1-03), which is S2.

Mirror. Justifying every use of a general-purpose technology is impractical and would favour incumbents (C7, I9). The transferable form is prior justification for high-stakes uses, which is what sector regulators already do.

Fix. - Move item 8 to “split”. The layer agrees; the timing, which is the instrument, is against him except where existing sector law already imposes it (FDA, NHTSA). - Record K9 as present for diffusion.

8. Knowledge lock-in is rated “inferred” although Huang concedes it. K10 is missing, and the K4 Mirror uses only the pro-Huang precedent. Severity: medium-high#

Location. §4.4, knowledge row (line 182, “Inferred; contested”); §4.3 Evidence and Mirror (lines 164, 168); §3.4, item 2 (line 119).

Problems. - The loss is conceded. Asked about the exam-score finding, Huang says “I think the last part. I completely agree… basic math is… being forgotten” [22:26]. He adds “we’re going to lose some… intellectual dexterity” [24:24] and “Some of the lower level… knowledge is gone” [24:52]. The mechanism, loss of skill, is documented and conceded. Only its significance is contested (“Does it matter?”). - The study is described selectively. It found exam scores down 20% within six months, entrance-exam scores down 18–24%, and “a full penalty emerging only after about two years” [21:16] (FC C041: accurate; observational, one county). §3.4, item 2 cites only the six months, as evidence of fast harm. The two-year lag is evidence of slow harm. - K10 is absent. K10 is strong: [K] and [U] strong, [F] strengthened. It warns that “averages… hide the most sensitive groups and life stages, and the timing of exposure can matter”. The aggregate “no economy-wide displacement” figure, which D05 uses against the critics at line 168, is an average. The documented effects fall on two sensitive windows: adolescents (the study) and 22–25-year-olds (19% below trend and widening). - The K4 Mirror is one-sided. It cites only the mobile-phone warning kept alive by “not enough time has passed”. The reports’ own opposite case is in LL2-03. In 1925 Hayhurst held that 27 months of public use “should have sufficed to bring out some mishaps”. The chapter calls this reliance on “existing statistics, collected for other purposes” and on “a short, two-year-period after first exposure” two weak assumptions “still common” today (p. 52). “Wait two years” [19:50] has the same structure.

Fix. - Change the knowledge row to “Documented (conceded by Huang; CEPR study); significance contested”. - Add K10 to §4.3. - Add Hayhurst to the K4 Mirror, so that it runs both ways: mobile phones against leaded petrol. - Correct the study’s description in §3.4.

9. G9’s “modification that helps Huang” rests on an example the project’s own fact-check undercuts, and it ignores strategic designation. Severity: medium-high#

Location. §4.5 Transfer (line 203) and Mirror (line 205); record table, row G9.

Problems. - The robotaxi example. “AI moves fast enough that ex post rules can arrive within years, as with robotaxis under NHTSA [1:19:12].” FC C166 says: “No federal ADS performance standard yet; rulemaking only begun Mar 2026”. What did act quickly was sector enforcement and state action (crash reporting, recalls, the Cruise suspension; FC C070). - Pre-emption. The industry, Huang included, seeks to pre-empt that state layer: “State-by-state AI regulation would drag this industry into a halt” (December 2025), and OpenAI’s blueprint asks for federal pre-emption. That is G9’s reversal mechanism acting before reforms exist. - Strategic designation (I5 and M7). I5 is [U] strong (BSE) and [F] strong (Fukushima). Its Ask is “Has the technology been designated strategic or critical, turning policy from reducing use to securing supply?” Hindsight LL2-06, lesson 9, records beryllium’s path: “policy turns from reducing use to securing it”. The record here: - the “American tech stack” as national aim [1:35:15]; - “the president is completely aligned with Jensen Huang” (Bessent, 15 September); - a voluntary pre-release framework (E.O. 14409).

The enforcer that “regulation will come in” [44:17] relies on is aligned with the promoter. This concerns institutional position, not the motive of any official (rule 4). - LL2-03, p. 56. Voluntary producer compliance relieved “the government of any pressure to introduce the regulations”. - An imputed rationale. The Mirror says Huang’s resistance to AI-specific rules “partly reflects” W8’s evidence that restrictions harden. He gave no such reason; he called regulation “the distraction” [47:10]. Also, W8’s persistence cases are all statutory rules. The one AI pause on record lasted two weeks.

Fix. - Change the Transfer line to “yes; the speed modification is mixed”. Sector and state enforcement can act fast, but federal standards lag, and pre-emption targets the fast layer. - Record I5 as present. - Delete the imputed rationale.

10. Three Mirror statements about Klein and the pacing statement are inaccurate, and all three favour Huang. Severity: medium#

Location. §1, last paragraph (line 35); §4.4 Mirror (line 193); §4.2 Mirror (line 156).

Problems. - Klein’s energy remark. “Klein wants energy built faster [1:44:44]” is used to spread the lock-in questions to Klein. His full remark: “But we need to build the energy faster to do that… you could subsidize it and you can make it easier to build”. “That” is the transition to sustainable energy that Huang has just described. Klein wants clean energy built faster, with subsidy. Huang says “You don’t need government subsidies” [1:40:15] and that fossil fuel must come first [1:44:52]. S2 and L3 turn on the direction of the energy mix, which is exactly where the two men differ. The lock-in question applies to Klein only in the weak sense that any build-out creates capital. - The pacing statement. It is said to ask for “the option to buy time” “without saying how the time would be used”. The sentence Klein reads says: “to address emerging risks, develop security measures, and strengthen oversight” [50:46]. The fair criticism is narrower: it names purposes but no tests, milestones or exit criteria (T3). - Klein’s own proposal. “Klein’s own proposal was never stated [54:44]” contradicts D05’s own line 231 (“Klein’s aim of stopping the labs from pursuing recursive self-improvement”). He was cut off on air, but he had stated the proposal in his column and solo episode of 20 September (Huang analysis §2.4). The same error appears in D03 and was flagged in that red-team review.

Fix. - Correct all three. - In §1 and §4.4, say the build-out question applies to the labs fully and to Klein only weakly. On the energy mix, Klein’s stated preference runs against gas lock-in.

11. The L5 Mirror is falsely balanced: reducing selection pressure is what pacing does. Severity: medium#

Location. §4.8 Mirror (line 243); record table, row L5 (“Public gates face the same treadmill”).

Problem. L5’s remedy is to reduce selection pressure and vary tactics, not to add product (LL2-11, pp. 251–252). Stacked traits “bought time, not escape” (T04 §7).

Huang diagnoses the pressure himself: “unless you align it… the software is going to go do the most obvious thing” [32:09]; “if you give it a constraint, it’ll go find another solution” [48:58]. His remedy is more product: “their next implementation of their sandbox is going to be much better” [32:09]; “Accelerate the living daylights out of” safety technology [1:16:05].

The critics’ measures are closer to the reports’ remedy: - OpenAI paused reinforcement-learning training for two weeks from 18 August; - the pacing statement asks for tools “to deliberately pace the frontier”; - Klein proposes stopping autonomous self-improvement.

These all reduce optimisation pressure. They still share the verification gap, which is the fair part of the Mirror. But restraints keyed to an activity or capability class rely less on behavioural tests than a release gate does (the same point is made in the red-team review of D03, issue 9).

Fix. Make the Mirror asymmetric: “The critics’ measures include an L5-type remedy (less pressure) and share the verification gap; Huang’s emphasis is stacking better containment.” Keep D05’s useful point that reward and benchmark design is the engineering form of the remedy.

12. “Chips are short-lived” is used as a mitigant, although Huang argues the opposite and the financing depends on it. The glut is not treated as a lock-in amplifier. Severity: medium#

Location. §4.4 Transfer (line 189); §6, item 7 (line 328); §2.3 (line 54).

Problem. - Huang’s own claims. “The useful life of our compute is much longer”; compute is an “asset class, kind of like an airplane” that “starts out as a passenger airplane” and “ends its life as a… cargo plane”; it is a “collateralized asset”; “a huge unlock for our growth” [1:21:05]. Useful life is disputed: two to three years (Burry) against four to six (Nvidia) (Huang analysis §5.2). - The consequence. If compute becomes collateral, the lender’s assumed life sets how long the stock must earn. That is the financial form of “yesterday’s investments will be redeemed before any serious risk reduction” (LL2-28, p. 672). Cascading into secondary uses is how installed stock persists (S1: PCB equipment, LL1-06, pp. 66–72). - The glut. Huang concedes one is coming [1:29:20]. With collateralised debt, a glut increases pressure to keep the stock busy: cheaper compute leads to more use (S2) and to resistance to pauses. - The Ratepayer Protection Pledge addresses the stranded-cost tail: signatories pay “whether they use the electricity or not”. Nvidia is not among the seven signatories (D06 §4.12; LA4 C5). - “There’s not much to learn from the past” [1:29:20] is quoted in §2.3 but not analysed. It is an explicit rejection of historical learning about boom and bust. It has the shape of the reports’ “unlike yesterday’s technology, today’s technology is now safe” (p. 672).

Fix. - Replace “Chips are short-lived” with “chip life is contested, and Huang argues it is long. The financial lock-in horizon is set by collateral terms, not by physical life.” - Add the glut as an amplifier of L4 and G9. - Note the Ratepayer Pledge and Nvidia’s absence from it under §7, item 3.

13. The energy-shortfall narrative and the “one input” claim are recorded but not analysed. Severity: medium#

Location. §2.5 (line 62, which cites C206 and C207 but not C205); §4.7; §4.9; §5, item 1.

Problem. - The narrative. “We got ourselves really gummed up in climate change and sustainable energy” [1:39:53] is contested. FC C205 finds that the causes were mainly flat demand, interconnection queues and turbine supply, and that climate policy added capacity. - Why it matters for interests. The Huang analysis names the causes of the energy shortfall as one of three places where “interest is most telling”, because he departs from disinterested opinion (In brief; §8.4). D05 applies that test nowhere in its energy analysis. - The pattern. Blaming environmental precaution for a shortfall is the reports’ “precaution framed as anti-innovation” narrative (T04 P9; LL2-16, pp. 401–402). It also underwrites the “no alternative” claim for gas. - The one-input claim. “Intelligence is gonna scale by one thing, and that’s compute” (Lex Fridman) is a single-path claim from the supplier of the one input. It has the same shape as TEL as “the only material available”, which was asserted after capital was committed (LL2-03, p. 54). - Weight. M4 is moderate. Language shows framing, not its effect. Rule 4: no bad faith is implied.

Fix. - Add C205 and the interest test to §2.5 and §5, item 1. - Record M4 as present for “gummed up”, “no alternative” and “one thing”.

14. The L3 Mirror says the pacing proposals have “no plan for non-signatories”. Amodei’s has one, and Huang opposes it. Severity: medium-low#

Location. §4.7 Mirror (line 231); record table, row L3/I8 (“Pacing names no replacement”).

Problem. Amodei’s essay proposes coordination among democracies and “no powerful chips for China” (Huang analysis §2.3; D05 §8). That is a plan for the principal non-signatory state, and Huang argues against it. The gap is domestic non-signatories such as Meta (“you just take the time that you need internally”).

Fix. Narrow the claim to domestic non-signatories. Note that Huang’s position on export controls removes the critics’ plan for foreign ones. That plan is contested on displacement grounds (FC C200), and his L3/I8 argument against it stands as contested.

15. Entries that are clearly present go unrecorded: M1’s Ask, M3 for Nvidia itself, C8 in Huang’s own words, and habituation. Severity: medium-low#

Location. §2.7 (line 76); §5, item 2; §4.2 Evidence (line 152); §4.12.

Problems. - M1. D05 uses M1 only to justify treating Huang as sincere. It does not run M1’s Ask: “If everyone involved is sincere, what would still produce harm: weak feedback… long lags, costs borne by others, commitment to earlier positions?” All four are present on this dimension: - emissions lags; - ratepayers, neighbours and the climate; - $279bn of supply commitments; - a metric that counts productivity and not cost [1:21:05]. - M3. D05 applies M3 to the labs that would take the exits, not to Nvidia. Nvidia’s cost of admitting a problem with the build-out grows with each commitment: $105bn of guarantees, a reported $30bn stake in OpenAI, $279bn of supply commitments and the Hugging Face purchase. M3 is moderate to strong, [K] and [U]. - C8. Huang states C8 himself: builders can be good neighbours, “but it’s hard to do that. You know, after the fact” [1:40:15]. That is the reports’ timing lesson in his words. LA4 records it; D05 does not. - Habituation. D05 counts “That sensation lasts about seventeen days” [1:08:03] as deflating wonder. The same turn describes twenty years of infrastructure build-out making everything “seem so natural… To the point where we now take it for granted”. Habituation that removes conspicuousness is how installed infrastructure escapes scrutiny. Compare M4’s warning about “normal” and “natural”, and Huang’s “This is very normal” [48:58]. This is low weight (M4 limits).

Fix. - Add M1’s Ask and M3-for-Nvidia to §2.7 or §5, item 2. - Add C8 to §4.4 or §7. - Record the seventeen-days point as cutting both ways.

16. Hinton’s radiology forecast is labelled a “claim for the benefits of caution”. It was a confident capability forecast. Severity: low-medium#

Location. §4.2 Mirror (line 156); §6, item 2 (line 323).

Problem. “People should stop training radiologists now” was a forecast that AI would outperform radiologists. It failed on timing (FC C123). That is the same kind of error as Huang’s “any disease… superhuman” (FC C011: inaccurate). It shows the cost of confident forecasts in either direction, and it is a single case (rule 0: showcase or sample?). It remains good C7 evidence, since the harm was real (FC C127).

Fix. Re-label it in §4.2 as “an overconfident capability forecast whose cost fell on trainees”. Note in §6, item 2 that the same lesson applies to confident benefit forecasts.

17. Minor corrections. Severity: low#

What D05 gets right (not disputed)#