Late Lessons, Jensen Huang and AI
Markdown version

Red team A (Huang’s advocate): D05, Innovation, infrastructure, trajectory and lock-in#

Reviewer’s role: find every place where D05 is unfair to Huang or to the engineering approach. File reviewed: working/synthesis/dimensions/D05-innovation-infrastructure-lockin.md (382 lines). What it was checked against: - the full transcript; - 02 §§7, 8.1, 8.2, 8.4, 9.1, 10.1 and 10.5; - 01 §§5.1, 5.5–5.8 and 6.1–6.2, and the lens entries D05 uses (L1–L6, K4, G9, S2, M3, M5, C3, C6, I8, T4); - T04 §§6, 7, 9, 12 and 13; - hindsight LL1-10 and LL2-06; - fact-checks C020, C051, C172, C173, C176, C206 and C214; - E1 and E3 on state pre-emption.

“l.” gives the line number in D05. Transcript quotations have stutters removed. The suggested fixes are worded so that D05 stays stand-alone: no article angles and as little project-internal commentary as possible.

Overall judgement#

D05 is one of the fairer dimension files. Its Mirror lines in 4.1, 4.4, 4.7, 4.8 and 4.10 are real, section 6 is substantial, LL2-22 is flagged, post-recording evidence is marked, and Huang is treated as sincere. Its unfairness is concentrated in four places:

Issues 1–6 would change the summary and section 5. The rest are local fixes.

High#

1. “Exits that lock-in makes expensive” misstates his conditions, misattributes a quotation, omits the best counter-evidence and has no Mirror#

Location: summary item 2 (l. 31); 4.4 Analysis (l. 191); §5 item 2 (l. 310); §7 item 3 (l. 339); §9 (l. 370).

Problem: - “His conditions are all exits” is inaccurate. His stated conditions (02 §10.5) also include: - adding regulation where a gap is shown [1:19:12]; - “regulation will come in” after harm [44:17]; - third-party auditors [51:20]; - tenfold evaluation compute [48:58]; - a US-first allocation rule [1:37:36]; - accepting community refusal [1:40:15].

In his model (02 §10.1, P5) the load-bearing parts are containment, verification and release discipline. The exits are the backstop. Calling them “the load-bearing parts of his safety model” (l. 310) inverts this. - “take a pause” is cited to [36:44]. It comes from Dreamforce on 15 September (02 §7.1, §10.5). At [36:44] he says “Don’t ship it” and “we have to shut the labs down”. - The central claim goes untested against the one natural experiment available. D05 says that each round of commitment makes the exits more expensive. In August 2026: - OpenAI paused reinforcement-learning training for two weeks and put its largest planned run on hold, “at great cost and delays”; - Anthropic moved about 150 engineers to security and paused external cyber evaluations; - Altman told the UN Security Council: “We have unilaterally slowed down in the past. We will do so in the future” (02 §7.3(b), §8.1 T4).

All of this happened while those labs held the industry’s largest commitments (OpenAI about $1.4 trillion; D05 l. 354). Judged ex ante, the smaller exits have so far stayed usable. - Fungibility cuts the other way for the smaller exits. Compute held back from a release can be moved to evaluation, which is exactly what he prescribes: “I want them to get more compute, but allocated towards evaluation” [1:16:05]. The lock-in argument has real force only for the drastic exits: a long, industry-wide pause or a shutdown, where lease and debt obligations bite. - The supporting evidence is weaker than presented. - Guidotti’s “room for them to turn around” is rated “partly held up… a suggestive mechanism, not an established one” (hindsight LL2-06, claim 8). It concerns legacy liability blocking cooperation on standards, not financial entanglement blocking a pause. - D05 itself says financial lock-in “has no precedent in the reports” (l. 187). Applying it here goes beyond the lens; it is not a lesson of the reports. - M3’s own limits note that “Organisations did reverse where they had less sunk commitment”. - The shutdown condition is a different kind of case. It is triggered by a lab’s own finding that the damage “is too great” [36:44]. The reports’ lock-in cases concern chronic, contested harm, where the cost of exit was weighed against uncertain harm. No case in the corpus resembles an admitted catastrophic failure of containment. - There is no Mirror in §5.2. The pacing advocates carry larger commitments, and their proposals (a pause, a coordinated slowdown) are exits too. The voluntary commitment D05 cites as dropped, OpenAI’s 20% pledge, was a pacing advocate’s, not Huang’s. - “The firms that would take these exits are financially entangled with Nvidia” (l. 310) implies that Nvidia’s stakes weigh against exits. Nvidia is not the party that would take them. And the shutdown condition runs against Nvidia’s commercial interest (02 §8.4, second table), though its expected cost is low.

Evidence: transcript [36:44], [1:16:05]; 02 §§7.1, 7.3(b), 8.1 T4, 8.4, 10.1, 10.5; D05 l. 187, l. 354; hindsight LL2-06 claim 8 and its verdict table; 01 §6.11 (M3 limits).

Fix: - Summary item 2: “Huang’s safety model rests on engineering practice (containment, verification, release discipline), backed by exits: ‘Don’t ship it’ and ‘we have to shut the labs down’ [36:44], and ‘take a pause’ (Dreamforce, 15 September). The reports show that commitment can raise the cost of exit (L4, M3). For the smaller exits, the 2026 record so far points the other way: OpenAI paused training and Anthropic redeployed staff while holding the largest commitments in the industry. The question is sharpest for the drastic exits, a long pause or a shutdown, and for financial obligations, a form of lock-in the reports do not cover.” - §5.2: make the same changes, and add a Mirror: “The same holds for the pacing advocates, whose commitments are larger and whose proposals are also exits; the voluntary commitment that was not honoured, the 20% pledge, was a lab’s.” - Delete the Guidotti sentence, or give its rating and note that its mechanism is different. - Replace “financially entangled with Nvidia” with a neutral statement of the stakes, plus: “his shutdown condition runs against Nvidia’s commercial interest”. - Confidence: medium for the drastic exits; low for the smaller ones. Change §9 to match.

2. Three quotations in 4.10 are cut so that they narrow or reverse his meaning, and the resulting “contradiction” does not hold#

Location: 4.10 Evidence (l. 263) and Analysis (l. 267); 4.12 row “L6 claims”.

Problem: - “Regulation is ‘the distraction’ [47:10]”. The full turn reads: “I’m not against laws and regulations. I’m not against laws and regulations. I’m against currently the distraction.” D05 drops the disclaimer, said twice, and the word “currently”. A statement about timing and focus becomes a statement against regulation. The Mirror in 4.5 (l. 205) uses “currently” correctly; 4.10 does not. - “State-by-state AI regulation would drag this industry into a halt” (December 2025) is quoted without the sentence that follows it: “A federal AI regulation is the wisest” (E1; E3). The claim is about fragmentation, and it comes paired with support for a federal standard. - The Analysis builds on the shortened quotation. It reads: “On those premises a binding verification requirement is the kind of rule least likely to ‘halt’ anything. His claim and his premises pull apart.” But the “halt” claim was never about a verification requirement, and with the full quotation there is no contradiction to report. There are two further problems: - T04 P7 requires “an available engineering… pathway”. Evaluation awareness (02 §8.1 T1) is precisely the doubt over whether such a pathway exists for frontier verification. The conditional is therefore less clearly met than D05 implies. - He accepts some binding rules: a US-first legal requirement is “no problem” [1:37:36], and third-party auditors are “terrific” [51:20]. - The car analogy is cited without 02’s charitable reading (T7). That reading is that his stated regulatory position, sector regulators plus engineering, “is closer to the historical pattern than his slogan”.

Evidence: transcript [47:10], [51:20], [1:37:36]; E1 l. 239 and E3 l. 194 (the full December 2025 quotation, and E3’s note that “I’m against currently the distraction” fits this record); 02 §8.1 T1 and T7; T04 §9.

Fix: - Evidence (l. 263): “‘I’m not against laws and regulations… I’m against currently the distraction’ [47:10]; ‘State-by-state AI regulation would drag this industry into a halt… A federal AI regulation is the wisest’ (December 2025)…” Add T7’s charitable reading to the car sentence. - Analysis (l. 267): “On Huang’s premises (few frontier labs, which ‘know how to do it right’ [44:17], facing an engineering problem), the reports’ conditional suggests that a single binding verification requirement would be among the rules least likely to stop development, provided a workable verification method exists, which evaluation awareness puts in doubt. His December 2025 objection was to a patchwork of state rules, and he favoured a federal one. The reports do not settle, either way, his general claim that new rules are unnecessary.”

3. S2 presents a point he concedes as an error he commits#

Location: 4.6 (ll. 211–221); 4.12 S2 row; §9, “High: that totals are outgrowing per-unit efficiency”.

Problem: - In the interview Huang does not judge per unit while totals grow, which is S2’s failure mode. He says the supercomputers are efficient “but they’re still going to use a lot of power” [1:40:15]. He forecasts computation up “a billion times” [1:21:05]. He says “we’re going to use a lot more fossil fuel” [1:40:15]. The Analysis (l. 215), “On his own model, efficiency will be spent on more computation, not less power”, is presented as a finding against him; it is his stated view. - “His test (‘how productive is it? Not how expensive’ [1:21:05]) is per unit; S2 asks who tracks totals” (l. 217) misreads the quotation. In context it concerns revenue per dollar of capital in an AI factory; the next sentence gives the $50 billion build cost and the rental figure. It says nothing about energy performance. - Hausfather’s Jevons line rebuts a claim Huang does not make in the interview: that efficiency will reduce AI’s energy use. Huang did make it in 2024 (“Accelerated computing is sustainable computing”). D05 reports that line in §2.5 but does not connect it here. If the 2024 record is the target, say so. - D05 classifies “Bring in your own power generation” two ways. 4.6 lists it as harm moved. Huang offers it to protect communities, alongside lower property taxes, setbacks and schools [1:40:15], and D05 §7 item 7 classes it as producer-pays (C6). The same proposal cannot be only displacement in 4.6 and producer-pays in §7. Its effect depends on the fuel. - “Refusals and moratoria push projects elsewhere” is listed as evidence about Huang. It is communities acting under his “so be it” concession, and belongs only in the Mirror (where it already appears).

Evidence: transcript [1:21:05], [1:40:15]; D05 §2.5 (l. 62) and §7 item 7 (l. 343); 02 §9.1 energy row.

Fix: - Evidence (l. 213): “Huang accepts the S2 point in the interview: efficient systems are ‘still going to use a lot of power’, computation will rise ‘a billion times’, and ‘a lot more fossil fuel’ will be burned first [1:21:05, 1:40:15]. His earlier message, ‘Accelerated computing is sustainable computing’ (2024), is the per-unit framing S2 warns against, and the analyst Zeke Hausfather’s rebuttal applies to that. The live S2 questions are therefore not whether totals grow, which he concedes, but who tracks them, which fuel meets them and where the costs land. ‘Bring in your own power’ protects ratepayers (C6), and moves emissions on-site if the power is gas.” - l. 217: delete the “productive… not expensive” sentence, or relabel it as an economic test. - 4.12 S2 row, “Present?”: change to “Conceded in the interview; present in his 2024 framing”. - §9: reword accordingly.

4. Energy lock-in: “no disanalogy”, “the record does not support” reversibility, and “Confidence: high” overstate a point he partly concedes#

Location: summary item 1 (l. 30); §5 item 1 (l. 308); §9 (l. 370). Also inconsistent with §6 item 9 (l. 330).

Problem: - Direction over magnitude (rule 6). Huang concedes the direction: more fossil fuel for “four or five years”. The disagreement is over duration and offset: whether gas built now locks in emissions for decades, and whether AI-driven clean procurement outweighs it. These are questions of magnitude, where the reports are weakest (01 §5.5 items 2–3). - The reports’ own energy record on reversibility is mixed. Hindsight LL1-10: UK sulphur dioxide emissions fell 98% after 1990, driven by the switch from coal and fuel oil to gas, and by power-station closures and conversions; structural change “did much of the work”. D05 §6 item 9 cites this as precedent for market-driven exits. “Assumes a reversibility that the record of energy and industrial lock-in does not support” (l. 30) contradicts that item. - Disanalogies remain, even at the energy layer. - Data-centre demand is new, additional load, so the comparison is between mixes of new build, not between an incumbent and an alternative. - Capacity lock-in is not emissions lock-in: gas plants can run at falling capacity factors, as back-up for renewables. - Chips are short-lived.

“Here the reports apply with no disanalogy” (l. 308) should read “with the fewest disanalogies”. - The surgery metaphor is over-read. It was a one-off reply to Klein’s point about subsidies, new in this interview (02 §9.1). On it D05 builds the claim that surgery “presumes a diagnosis, consent, and that the patient who bears the pain is the one saved” (l. 308). Two hedges are lost along the way: - the ellipsis in §2.5 drops “unfortunately” (“we have to unfortunately use renewable energy, use fossil fuel”); - §5.1 and the summary drop “hopefully” (“hopefully we can transition”).

Both words mark uncertainty about the transition, not a confident assumption that it can be reversed. - The point about who bears the costs (C3) is fair for the climate, but incomplete for ratepayers and neighbours. For them Huang proposes internalisation (own generation, lower property taxes, setbacks, schools, parks) and concedes a local veto, “so be it” [1:40:15]. §5.1 does not say so. - “Leaning on ‘market forces’ is leaning on a co-driver, effective but fragile in the reports’ record.” The reports’ co-driver exits held: catalytic converters, and the fuel switching that cut sulphur. “Fragile” is an inference drawn in the LL2-03 digest. The reversals listed in T04 §13.9 were policy exits (derogations, nuclear phase-outs), not market ones. - There is no Mirror. Klein wants energy built faster (“we need to build the energy faster”; subsidise it and “make it easier to build” [1:44:44]), and the labs calling for pacing have signed multi-gigawatt deals (§8). Faster permitting and building would speed up gas as well as clean power.

Evidence: transcript [1:40:15], [1:44:44], [1:44:52]; hindsight LL1-10 (UK SO2 down 98%; structural drivers); T04 §8(g) and §13.9; 01 §6.7 (L6 co-drivers); 02 §9.1 (“new in this interview… the surgery metaphor”).

Fix: - Summary item 1: “The energy layer is where the lessons transfer with the fewest disanalogies. Huang concedes the direction, ‘a lot more fossil fuel’ for four or five years [1:40:15]. The open questions are how long gas capacity built now will run, how heavily it will be used, and whether AI-driven clean procurement outweighs it. The reports’ record here is mixed: sunk energy investment delayed risk reduction (LL2-28, p. 672), but structural change later cut sulphur emissions fast (hindsight LL1-10).” - §5.1, surgery sentence: replace with: “The ‘surgery’ image [1:44:52], offered with an ‘unfortunately’ and a ‘hopefully’, treats the fossil phase as temporary. The reports give reason to test that, since energy capital outlasts the conditions that justified it, but not to assume it will fail.” - §5.1, other changes: - add his proposals for local internalisation and his “so be it”, and the Mirror; - replace “effective but fragile” with “effective, and in the reports’ market-driven cases durable, though contingent”; - split the confidence: high that the lock-in mechanism applies (and he concedes the direction); low on its duration and net effect.

5. “Appraisal after footprint” reads [48:58] and [1:11:19] against their context; the K4 evidence conflates personal adoption with exposure#

Location: §2.2 “When appraisal happens” (l. 49); 4.3 (ll. 164–170); §5 item 3 (l. 312); 4.12 K4 row.

Problem: - “This is very normal” [48:58] refers to issues that surface with use: “more use cases, more people using it, they’re going to get a lot more issues”. In the same turn he prescribes: - moving R&D to verification; - a possible tenfold rise in evaluation compute; - “if they believe they’re out of control, then the right answer is, don’t ship products until they’re in control”.

At [1:11:19], “unnecessary until now” refers to labs that six months earlier were trying to make something useful at all. His account is that testing effort should rise with deployment. That is the reports’ own lesson: “deployment scale should itself trigger scrutiny” (T04 P10; LL2-02 lesson 4, p. 35), and D05’s own L1 wording, “scrutiny in proportion to scale” (l. 134). Calling it “the reports’ scale-before-appraisal pattern stated as normal practice” (l. 312) reverses it. - The fair critique is stronger, and 02 already makes it (§8.1 T2). It concerns what triggers the testing: - in his account testing grows with commercial footprint, while the incident showed risk arriving with capability, before any product; - no pre-release gate can detect slow, diffuse harms, whether the gate is his, the labs’ or a public one. That needs sustained observation after release, which his model leaves to sector regulators and liability. - “Huang urges adoption as fast as possible [17:07, 20:17, 1:31:03]” (l. 164) runs together three different things: - [17:07] is advice to individuals worried about displacement (“so that you benefit from this transition… and not just be impacted by it”); - [20:17] is a prediction about graduates; - [1:31:03] is a goal of national diffusion.

K4’s cases are involuntary exposure to persistent substances (MTBE, lead in air); a person using a tool is not exposure. The slow harms D05 lists (skills, labour-market pathways, dependence) are effects of use at scale and can carry K4’s question. The disanalogy (voluntary use rather than exposure; effects that depend on how the tool is used) belongs in §3.4 and in the Transfer line. - Weight by case type (rule 9). K4’s forward record is mixed, and the reports’ only digital case, mobile phones, is their clearest warning not borne out (K4 limits). D05 puts this in the Mirror, but it should also lower the transfer weight for an information technology. - Hamilton’s “how can you control the whole country?” (l. 166) is mapped onto the gap between containment in the lab and diffusion. Huang’s answer for the whole country is sector regulation at the application layer [1:19:12], plus “regulation will come in” [44:17]; D05 credits this in §6 item 8. The Transfer line should say so. - The June breach, disclosed only in September (l. 166), was a disclosure failure by OpenAI, a pacing advocate. It is a Mirror point, not only evidence about Huang’s model.

Evidence: transcript [17:07], [20:17], [48:58], [1:11:19], [1:19:12]; T04 §12 (P10); 02 §8.1 T2; 01 §6.3 (K4 limits).

Fix: - §2.2: “When appraisal happens. Testing effort should rise as products scale. Labs that six months ago lacked products found heavy testing ‘unnecessary until now’ [1:11:19]; with ‘so much market footprint’ they must shift R&D to verification, perhaps with ten times the compute, and ‘Don’t ship products until they’re in control’ [48:58].” - §5.3: retitle “Testing keyed to footprint rather than capability”. Text: “Huang ties the growth of testing to market footprint [48:58, 1:11:19], which matches the reports’ call for scrutiny in proportion to scale. The July incident shows risk arriving with capability before release. For slow harms to skills, labour-market pathways, dependence and emissions, no release gate, his or his critics’, will detect them. What the reports ask for is sustained observation after release, which his model leaves to sector regulators and liability.” Confidence: medium. - 4.3: separate voluntary use from exposure; add sector regulation to the Hamilton paragraph; add the disclosure Mirror.

6. The finding on how Huang evidences benefits rests partly on quotations about something else and on [K] precedents#

Location: summary item 3 (l. 32); 4.2 Evidence (l. 152); §5 item 4 (l. 314); 4.12 L2 row.

Problem: - “He treats adoption, venture investment and ‘offtake’ as proof that AI is useful [05:55, 1:25:12]” misreads both passages. - At [05:55] the “proof point” is for job creation: “Jobs are obviously being created from 500 billion dollars of new investment”. The usefulness claim is a separate one. - At [1:25:12] “offtake” is a condition he concedes: “if the AI services have no offtake, then obviously building computers for it is pointless”. That acknowledges that financed supply cannot stand in for end demand; it does not claim that demand proves benefit.

The fair point (C176; 02 §8.1 T13) is whether demand that Nvidia helps finance is independent evidence of usefulness. Say that. - “‘You could see the system working’ in token shares [27:02]” is not a benefit claim. The “system” is the balance between closed and open models: “the closed models are vibrant. The open models are vibrant. And you could see it.” - C172 is counted among benefit claims that “failed their checks”. The fact-check itself notes a “possible mishearing of ‘fourteen to fifteen’”, and D05 §9 (l. 382) flags the rental figure as a possible transcription error. It should not be counted. - The precedents are [K]-type cases where lack of benefit was already shown. DES was tested in 1953 and found not to work; growth promoters gave small gains; seed treatments were used “regardless of the presence and abundance of pests”. L2’s strength is “[K] strong (DES); [F] mixed”. For AI, D05 itself notes observable benefits: record numbers of radiology training posts, the defensive use of an open model in the incident, and no economy-wide displacement. Under rule 9 these precedents transfer weakly. - He does offer outcome evidence on benefits, in radiology and software engineering. Some of it failed its check (C011, and C013 on mechanism), and some holds (radiology demand). Report both.

Evidence: transcript [05:55], [27:02], [1:25:12]; factcheck C020, C051, C172, C176; 01 §6.7 (L2 strength); 02 §7.3(c).

Fix: - Summary item 3: “The reports ask that benefits be tested as hard as risks. Huang’s evidence of benefit leans on investment and demand: $500 billion of venture capital as the ‘proof point’ for jobs [05:55]. Nvidia helps finance some of that demand, so the demand is not independent evidence of usefulness (C176). He concedes that demand must be real (‘if the AI services have no offtake… building computers for it is pointless’ [1:25:12]). Some of his specific benefit claims failed their checks (radiology ‘superhuman’ detection, C011; clean-energy funding, C214); others held (demand for radiologists).” - 4.2: remove “You could see the system working” and C172. - §5.4: name the precedents’ case type, and lower the confidence to medium.

Medium#

7. Steering: Klein’s phrase attributed to Huang, verification research listed as starved by his steering, and no Mirror in §5.5#

Location: §2.3 (l. 54); 4.9 (l. 251); §5 item 5 (l. 316).

Problem: - “Single company industrial policy” is Klein’s phrase [1:27:32]. Huang replied “We’ve put a lot of money into this ecosystem. Yeah” [1:27:41]. That acknowledges the money; whether “Yeah” accepts the label is unclear. §2.3 (“he accepts”) and §5.5 (“‘Single company industrial policy’ answers to shareholders”) treat it as his own framing. - “Verification research, which his own 80/20 figure says is under-funded” is listed among alternatives that struggle for investment. In the interview Huang is its loudest advocate (“I want them to get more compute, but allocated towards evaluation” [1:16:05]; tenfold evaluation compute [48:58]). The under-funding is the labs’. - “Low-compute methods” are also listed. He calls Chinese open models “terrific” [1:33:51], and efficient open models can reduce demand for compute, against Nvidia’s interest (02 §8.4, second table; DeepSeek). Say so. - §5.5 has no Mirror, though 4.9 has one (coordination among a few labs; Klein’s own steering frame). Klein’s proposal for who would hold the gate is also never stated [54:44].

Fix: - §2.3: “Klein calls Nvidia ‘a single company industrial policy’ [1:27:32]; Huang answers, ‘We’ve put a lot of money into this ecosystem. Yeah’ [1:27:41].” - 4.9: drop verification research from the list, or add “which Huang himself urges the labs to fund”; note his support for efficient open models. - §5.5: add the Mirror sentence from 4.9.

8. G9: a relocated principle counted as a reversed commitment, and the pledge that failed was a critic’s#

Location: 4.5 Evidence (l. 201) and Mirror (l. 205).

Problem: - The recursive self-improvement example. D05 writes that his 2023 view that AI self-improving “in the wild… should be avoided” became recursive self-improvement as “a fabulous thing” [1:12:47]. In the same answer he says: “We can’t just have it recursively changing all the time… they have to test the product before they release it. We will test the product before we release it into operation.” 02 §8.1 T11 gives the charitable reading: - the constant is human evaluation before anything reaches the world; - the RSI he calls fabulous is narrower (skills, memory, the next release); - in 2017 he called AI writing AI “really incredible”.

This is at most a principle moved to a different point, not clear evidence of a protective commitment dropped. - “Nvidia’s 2023 support for licensing high-risk AI services” came from Nvidia’s chief scientist’s Senate testimony, not from Huang. 02 §9.1 rates his position on regulation “consistent in principle, hardened in practice”, with sector regulation constant since 2024. Licensing high-risk uses sector by sector is close to sector regulation. “Hardened” is fair; “reversed” overstates. - The protective commitment that clearly failed, OpenAI’s 20% pledge, belongs to a pacing advocate. The Mirror should say that G9 bears on voluntary pacing pledges, which are the labs’ proposed gates. - G9 is tagged [K] and [F]. The Strength line should say so and apply rule 9.

Fix: - Qualify or drop the RSI example, using T11’s reading. - Attribute the 2023 licensing line to Nvidia’s chief scientist, and use 02’s word, “hardened”. - Add to the Mirror: “G9 also bears on the pacing proposals: voluntary commitments like OpenAI’s 20% pledge are what the reports found easiest to drop.”

9. Knowledge lock-in: arithmetic in schools mapped onto forgotten industrial alternatives, and a stretched inference about evaluation#

Location: 4.4, row “Knowledge and skill” (l. 182); 4.4 Analysis (l. 191).

Problem: - [22:26] is about children’s long division and multiplication tables. The quotation (“Does it matter?… I don’t think it does”) omits what follows: “but there must be some set of skills that matter… But maybe not those. We’re going to discover new ones.” At [24:52] he adds that “there are many people who are still going to be obsessed and passionate about the lower level layers”. Specialists remain. - The reports’ knowledge lock-in is industrial. Producers “forgot” the alternatives to tetraethyl lead (LL2-03, p. 55); seed networks and agronomic skills were lost (LL2-19). Offloading arithmetic in general education is a different thing. His own comparators, calculators and PCs, are cases where offloading did not lead to lock-in harm, and the Mirror should test that comparison. - The Analysis stretches. It says that knowledge lock-in erodes the capacity to evaluate systems, and that his gate relies on “humans… in the loop”. The evaluators in his model are specialists, whom he expects to persist. That general deskilling erodes specialist evaluation is not shown. 02 §8.2 A7 makes a narrower point, that the capacity to scrutinise concentrates among builders. Use that.

Fix: - Quote [22:26] and [24:52] in full. - Set the row’s status to “Inferred; contested; weak fit to the reports’ cases”. - Replace the Analysis sentence with 02’s narrower A7 point.

10. L3 is applied to energy supply where nothing has been restricted#

Location: 4.7 Evidence and Transfer (ll. 227–229); 4.12 L3/I8 row.

Problem: - L3 is an “after restriction” entry: it concerns substitutes chosen for a restricted incumbent. Gas for data centres is new supply for new load; nothing is being phased out and no incumbent is being replaced. - “Directly for energy” therefore overstates the fit. The relevant entries are L4 (capital) and S2 (totals), which D05 already applies. The “no alternative” point belongs to L6 and is handled fairly there (C206: mostly accurate). - His claim that clean energy is being funded as never before bears on whether gas is the only drop-in. C214 rates it misleading, but notes that AI is “a real clean-power buyer”.

Fix: - Move the energy point to L4 and S2. - Keep in L3 only the cases where it genuinely applies: - displacement driven by restriction (export controls; the defenders’ switch to an open model); - substitutes Huang favours (Chinese open models). - Transfer line: “with modification, for substitution driven by restriction; not for energy supply”.

11. The leaded-petrol narrative invites an Ethyl analogy without saying how the cases differ#

Location: §3.2 (l. 104); 4.3 Transfer (l. 166).

Problem: - The only extended case narrative in D05 is leaded petrol. Ethyl’s president’s “apparent gift of God” and a consultant’s “if we are to survive among the nations” sit just before the comparison. Readers will map them onto Huang’s “magical thing” and his “American tech stack”. - LL2-03 is a case unlike Huang’s in the respects that matter: - a known poison dispersed into the air; - deaths at the plants in 1924; - documented producer conduct, including refusal to supply dealers selling alcohol blends and forty years of research by industry alone.

D05 elsewhere treats Huang as sincere (§2.7, M1). Nothing in the text separates the two cases.

Fix: add after l. 104: “The case illustrates how lock-in proceeds, not how the actors behaved. Tetraethyl lead was a known poison dispersed into air, and the chapter documents producer control of research and supply. None of that is claimed for AI here.” Alternatively, use a lock-in example from a genuinely uncertain [U] case, such as MTBE or CFCs.

12. Section 5 draws its conclusions without the Mirror#

Location: §5 items 1–5 (ll. 308–316).

Problem: Rule 0 asks for the symmetry checks “first, and again before concluding”, and rule 2 makes the Mirror the minimum. Every §4 entry has a Mirror, but §5, which readers will take as the findings, has none. The missing checks:

§5 item Mirror to add
1. Energy lock-in Critics want faster building too
2. Exits The labs’ commitments are larger, and pacing proposals are also exits
3. Appraisal No pre-release gate detects slow harms
4. Benefits The claimed benefits of pacing are untested
5. Steering Coordinated pacing is also a few deciding for many

Fix: add one Mirror sentence to each item, drawn from §4.

13. “Wait two years” is treated like an open-ended latency argument, but it has a date#

Location: 4.3 Evidence (l. 164) and Mirror (l. 168); §7 item 2 (l. 338).

Problem: - D05 says “Wait two years” keeps a benefit forecast alive against early contrary evidence “in the same way” as the mobile-phone argument that not enough time had passed. That argument was objectionable because it was open-ended. Huang’s forecast has a horizon, about late 2028 (02 §10.5), and D05’s own open question 5 uses it as a test. - §7 item 2 asks “what would count against ‘Wait two years’, and by when?” He has already said by when. - His forecast concerns the next cohort of AI-native graduates, not today’s 22–25-year-olds, so current data do not contradict it directly. The employment gap is still fair as context.

Fix: - Mirror: “‘Wait two years’ differs from open-ended latency arguments: it has a horizon, and the early-career employment gap in AI-exposed occupations (19% below trend) offers a test of it by about late 2028. What it lacks is a statement of what result would count against it.” - §7 item 2: drop “and by when”.

Low#

  1. The open-weights analogy in L1 (l. 140).

    • D05 calls open weights the “nearest AI analogue” to chemical persistence and mobility. The two differ: a chemical’s hazard persists, while an open model’s relative capability declines as the frontier and defences advance. Released weights also add defensive capacity, as in the incident. Add one sentence.
    • Klein’s “mutate to take on new jobs” (l. 137) is his hypothesis, not evidence; label it as such.

  2. The dependence row (l. 184) and 4.4 Analysis. His “I can’t rely on somebody else’s service” [27:02] is about owning rather than renting infrastructure, and hardware that is bought is owned. The concentration point (more than 80% of accelerators) is valid but distinct: it is about concentration of supply, not control. State the distinction and keep the question.

  3. The defensive-adoption row (l. 185). Farmers planted dicamba-tolerant seed because herbicide drifting from neighbours’ fields physically damaged their crops. Competitive pressure to adopt a productivity tool is general (PCs, which Huang cites) and does not discriminate between cases. Relabel it “competitive adoption; weak fit”.

  4. The capital row (l. 180). Chlor-alkali plants aged 42–83 years beside a $105 billion, 20-year lease invites a false equivalence. Put the lifetimes in the table: chips about 2–6 years; buildings, grid connections, generation and leases, decades.

  5. M5 (l. 152).

    • “The most consequential companies of all of all time” [1:11:19] is part of his defence of the labs, not wonder at the technology.
    • “The magical thing” [03:52] refers to the user’s experience; “Nothing magical about it” [32:09] describes the mechanism. 02 lists this pair among the tensions that dissolve on inspection.
    • “Lean into AI” on climate has partial support: C214’s “real clean-power buyer” and Hausfather’s conditional case.

    Keep M5 at “language only”, and add the counter-quotations.

  6. §2.7 (l. 76).

    • “A disposition to treat harms as phases (‘digestion’, ‘transition’, ‘surgery’)”: digestion refers to market oversupply and transition to the labs’ change of focus; only surgery concerns harm.
    • For symmetry with the list of Nvidia’s stakes, add his positions that run against its interest (02 §8.4): the shutdown condition, the concession on digestion, “so be it”, and support for efficient open models.

  7. §8, Energy (l. 362).

    • “Follows the administration’s framing more than his own pre-2025 record” implies political accommodation. 02 §9.1 rates his record “evolved”, with optimism about clean energy “continuous”; the surge in demand also changed the facts.
    • D05 notes that the lab leaders’ energy views are not documented, so this is not a divergence from them.

    Rephrase it, or move it out of “Where he diverges”.

  8. Omissions from §6 (ll. 322–330). Credit:

    • his concession of the S2 direction;
    • his concession of local consent and his proposals for internalising local costs, which match C3 and C6 (only C6 appears, in §7);
    • his dated, self-set tests, which answer M2’s question “Has anyone said what evidence would change the view?”: no glut for two to three years, “two years” for graduates, tenfold evaluation compute.

  9. The record table (4.12). Align it with the fixes above:

    • S2: “Conceded in the interview; present in 2024 record”;
    • K4: “Yes for slow effects of use at scale (inferred); exposure disanalogy”;
    • L3: “Substitution driven by restriction only”;
    • L2: “Yes; precedents mostly [K]”.

What D05 gets right (keep these)#