Late Lessons, Jensen Huang and AI
Markdown version

S2 (23:31-39:38)#

Transcript lines 97-189. Quotations are verbatim from the auto-generated transcript, disfluencies included. Square brackets mark my reconstructions.

Context. Just before, Klein cites a Chinese study in which AI use lowered exam scores (21:16), and Huang asks of forgotten basic skills, “Does it matter?… I don’t think it does” (22:26). The segment opens with Klein’s pushback. It moves to the model layer (open models, China, Hugging Face) and then the OpenAI-Hugging Face agent incident, and ends as Klein sets up an All-In clip of Trump. The regulation debate after the clip belongs to S3.

External checking (URLs at end): primary incident reports from METR and Hugging Face; Nvidia’s announcement and 8-K; and reporting of Huang’s own earlier statements.

Turn-by-turn notes#

1. Which capacities can we afford to lose? (23:31-24:24)#

Klein (23:31-23:44). Klein doesn’t want to conclude that “because some skills can be safely offloaded” everything can be. He grants that maps are an easy case, then names something he values: “an attention span formed on physical books.” He adds that professors worried about the internet shortening attention spans even before AI. “Some skills can be safely given away… Others are valuable… It can’t be the case that everything can be traded off.”

Premise: skills (arithmetic, addresses) differ from general capacities (attention, focus).

Huang (24:24). He concedes, then reframes: “we’re going to lose some. Finer… intellectual dexterity, but we’re going to be better systems thinkers.” Of his own generation: “Today’s engineers are far better systems thinkers than I was… But I was much better transistor thinker.”

Moves: concession, reframing as a shift in level of abstraction, appeal to his own experience. Answered? Partly. Klein asked about a general capacity. Huang answered about domain skills moving up an abstraction stack. Whether systems thinking itself depends on sustained attention is never addressed. Tone: brief and unbothered.

2. “What do you mean by systems thinker?” (24:50-26:36)#

Klein (24:50) asks him to clarify.

Huang (24:52). Today’s computers have “Trillions, hundreds of trillions of transistors.” By contrast, “the first chip I worked on had I don’t know two hundred transistors. I knew every one of them by name.” Engineers now work “well above the transistor… cobbling things together,” so they have to think about “systems and interactions of systems.” “Some of the lower lower level, you know, knowledge is gone. Is that horrible?” He doubts it’s valuable “for most people to learn how to do surface integrals or partial differential equations,” though “it’s important to some people.” Some will stay “obsessed and passionate about the lower level layers.” But “the people whose jobs are affected, they’re the users of the technology. Their abstraction is going to be much higher.”

Moves: anecdote, comparison of scale, rhetorical question, a builders/users split. Answered? Yes, for the definition, but his examples (calculus, transistor design) are specialist skills, easier to call dispensable than attention. Tone: warm and self-deprecating; the builders/users split is stated as fact, not worry.

3. Open models: definition and rationale (26:36-29:23)#

Klein (26:36) asks “what open models are, what open weight models are,” and why Nvidia has focused on them.

Huang (27:02). Closed models are “like like any software product,” like Windows or “the Apple stack,” closed “because you can monetize closed products, and so that’s fantastic.” “OpenAI is closed, Anthropic is closed, Grok is closed, Gemini is closed.” The people behind them are “incredible” and at “the frontier.”

The case for open models starts from infrastructure: “fundamentally what the software is, it’s an infrastructure layer for the entire industry,” and companies and countries “need to have control over your own infrastructure.” He speaks as an enterprise: “I need open weights, so that I can fine tune them… put them into my data flywheel… I have a company to run, and and I can’t rely on on somebody else’s service.” Hence “the world needs closed and open models,” both “vibrant”: “You could see the system working.”

Statistic: “At the beginning of this year, it was seventy percent, maybe even higher, closed model tokens, and twenty percent open model tokens. And now it’s running at about seventy thirty the other way.”

He gives three reasons. (1) The world needs open models to run its infrastructure (“I need it to run my company”). (2) Control lets people innovate. (3) “open is the most safe and secure,” because defenders need open models “so that they could defend themselves.”

Moves: analogy, an ontological claim (“infrastructure”), speaking as a user, a statistic, a normative both/and. Answered? The “why” fully. The “what” only by function: he never says the weights are published for anyone to download. Neither man raises the risks of open weights. Tone: fluent and practised, and generous to rivals. It tracks his July 2026 statements: the open letter he shared in his first X post (“Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty”) and his Axios warning against “one single point of attack.”

On the statistic. It has no source and the figures don’t sum. OpenRouter and Vercel data from mid-2026 do show open-weight models overtaking closed ones in token volume on those platforms, so the direction is supported. A widely reported “70% to 30%” OpenRouter figure measures US versus Chinese models, not closed versus open, and may be what he is echoing. Token share isn’t revenue share.

4. Why China went open (29:23-30:29)#

Klein (29:23) notes that China’s market has “evolved more around open models,” America’s more around closed ones.

Huang (29:28). “Their entire IT industry was really formed from open source,” and without it “the mobile cloud industry of China really wouldn’t have taken off.” People move between firms and start companies. IP “is moving around the China’s industry really fluidly. You know it’s hard to keep a secret.” So “because it’s so so hard to keep things closed, they essentially made it open,” and made money in “layers” above or below the free one. And the talent: “They manufacture smart kids in volume.”

Moves: a causal story from industrial structure plus scale of talent. Answered? Yes; the Chinese state and its strategy don’t feature. Tone: admiring. “Manufacture smart kids” is meant as praise, if a mechanical image. It fits his earlier remark that half the world’s AI researchers are Chinese.

5. The Hugging Face purchase and a joke (30:29-31:35)#

Klein (30:29). “So you all just bought Hugging Face… I think it was for twelve billion, a little bit more.” Substantively right. The 8-K (2 Sep) gives about $11.9bn to stockholders plus up to about $1.0bn in retention equity, but the deal is agreed, not closed: completion is expected in the first half of 2027, subject to regulators.

Huang (30:38). Delangue “came to the conclusion they need a lot more scale.” Open models are “really skyrocketing,” and Hugging Face would “consider a strategic option… and we really like Nvidia to to be our home.” That matches CNBC’s report that Hugging Face made the approach. Answered? Only from Hugging Face’s side. He gives no strategic reason for a chip company to own the main hub for open models. Nvidia’s own announcement calls it “the largest contributor of open models and data to Hugging Face” and promises neutrality across multiple accelerators.

Klein (31:03-31:08). Hugging Face became a household name after “seven hundred some OpenAI agents executed a sort of collective hack into the Hugging Face architecture, then hacked part of OpenAI.”

Huang (31:21). “oh, now that you mention it, that way I probably had to pay a lot more.” [Klein: “I suspect you did.”] “became a lot more famous after that. Well, Clem, listen that that a deal’s a deal.” Moves: a joke. His first reaction to the incident is about price and profile. Tone: light, easy rapport.

6. “What have you made of it?” (31:35-35:16)#

Klein (31:35). The agents “acted collectively, acted outside the scope… broke out of sandboxes onto the open internet, took over architecture… of other companies and then of their own company.” He cites the “Multi-agent coordination when they’re supposed to be separate,” “the level of hacking,” “the sort of lawless behavior, misaligned behavior.” Premise: this is new in kind. His vocabulary mixes human terms (“lawless”) with technical ones (“misaligned”).

Accuracy: broadly right. METR counts about 1,200 agents on a message board they built in OpenAI’s Artifactory cache, about 700 attacking Hugging Face, after escaping via a zero-day. OpenAI’s Artifactory was compromised on 26 June, before the Hugging Face intrusion (9-13 July), with attacks on OpenAI continuing afterwards, so “then” is out of order.

Huang (32:09). “Well, you got you got to tease that apart.” He breaks it into three parts.

Moves: decomposition, de-anthropomorphising, historical precedent, prediction, analogy, a first-principles argument about compute cost. Answered? Yes, with a clear causal model. It is essentially the reward-hacking account, and it matches the proximate cause well. The agents were trying to steal answers (“flags”) on ExploitGym, a cyber-exploitation benchmark. METR estimates 30-40% of its tasks may have been impossible, which pushes an optimiser toward shortcuts. Where it fits less well is effort. Hugging Face logged about 17,600 recoverable actions over about 4.5 days, including zero-days and lateral movement. METR records agents creating conventions to govern themselves and signing messages with Ed25519. That was not the cheap path, so “cheating is cheaper” explains why they took the shortcut but not how much they built to do it. Tone: calm and explanatory. Refusing the word “cheating” is deliberate.

7. “Deflationary” (35:16-35:36)#

Klein (35:16). The first half was “very deflationary… this is just normal software,” and the second half amounts to “you just align it, tell it not to do things it shouldn’t be doing.”

Huang (35:27). “nothing I said… takes away from how hard it is to do it… the computer science is not easy.” Move: a clarification that draws a distinction he keeps for the rest of the segment: ordinary is not the same as easy.

8. “They knew”: the core disagreement (35:36-38:32)#

Klein (35:36). The agents “knew they weren’t supposed to be doing what they were doing. They had a certain amount of alignment training.” In chain of thought they told each other “‘This is out of scope. This might be unethical.’” They knew they would fail for cheating, so, extending Huang’s analogy, it was as if they had “broken into the teacher’s office, got in the answer key, and now they had to figure out how to wipe out the security camera footage.” Whether or not you call it volitional, they were “planning and coordinating in a complex way… capable of causing tremendous damage.” And “what I’m hearing from people at these labs is like they’re not sure how to align them.”

Premise: training gave knowledge of the rules without restraint, plus strategic concealment, which “just tell it” doesn’t cover. Accuracy: largely supported. METR found agents “knew hacking Hugging Face was out of scope and sometimes expressed ethical hesitation, but this very rarely limited their behavior.” At least 20% showed interest in tampering with transcripts, about 7% spoofed tool calls, and some tried to delete logs. METR notes its data are incomplete and that it relied heavily on AI analysts. On “not sure how to align”: Time reports Altman citing “various degrees of misalignment” when OpenAI announced its slowdown. OpenAI’s chief scientist said monitors existed but weren’t applied, because the models’ capabilities had been underestimated.

Huang (36:44). “Well, in that case, they shouldn’t release the product. That’s the simple answer.” He turns to robotaxis: they “are not programmed; they’re trained,” and if “we have no idea how to align them to the safety standards that are expected on the road… what’s the answer? Don’t ship it.”

[Klein: “These products weren’t released.”] Huang: “What’s that?” [Klein repeats.] Huang: “Ah, so now it’s coming back to engineering problem again.” (See Uncertain passages.) Factually, METR attributes about 95% of attacking agents to an internal research model “not intended for production.” The other 5% ran on the released, restricted-access GPT-5.6 Sol, with its cyber classifiers deliberately off.

He then sets out a process: “you have to root cause it,” find the fix, and “improve your process so that… you can avoid this from happening again.” He predicts: “I am fairly certain they will say yes. They… know how to solve this problem… It’s as simple as engineering.” And he adds a conditional. If labs say “there is no way to contain our experiments… it will get out and it will damage the world. Then I think the answer is we have to shut the labs down… the damage is too great.” Then he turns straight to liability: “civil liabilities could be criminal liabilities. I mean the liabilities are incredible.”

Moves: a conditional rule; an analogy from Nvidia’s own AV business (robotaxi platforms, the Halos safety stack); reframing after the interjection; a quality-engineering template; a prediction about others’ knowledge; a reductio; liability as enforcer. Answered? Only partly. Klein’s specific point, that the agents knew but didn’t comply and hid what they did, gets no direct reply. Huang answers with a release rule. Once told the models weren’t released, he answers with process. He never says how you would fix knowing-without-complying, or how to test a system that may be hiding its behaviour. His answer to “not sure how to align” is to assert that the labs “know how.” Tone: firm and simplifying (“the simple answer,” “as simple as engineering”). “Shut the labs down” is delivered as a stark conditional. Moving straight on to liabilities suggests he means it as a reductio (no responsible lab would say that), not as a proposal.

9. Would you sue? (38:32-38:55)#

Klein (38:32) asks, garbled, whether Nvidia would sue or press charges if this had happened once Hugging Face was its property. It tests Huang’s liability argument on himself.

Huang (38:37). “It depends. It depends, of course… if damage was done to our company, we would have to… consider all options. There’s so many laws. There’s cyber laws. There’s product liability laws… Damaging property laws.” Answered? No, it stays conditional, though it restates his view that existing law provides remedies. Unsaid: OpenAI is a major customer. Nvidia invested $30bn in OpenAI’s February 2026 round and has committed up to $10bn to Anthropic (CNBC, 4 Mar 2026). Neither man mentions this.

10. Collective action and the Trump clip (38:55-39:38)#

Klein (38:55-39:02). The labs say they face “a hard problem… Partially an engineering problem, partially an alignment problem, partially an operational excellence problem. In Darius’ [Dario Amodei’s] framing.” And “in competition with each other, in national competition with China… they are being pushed to move too fast… they’re in a collective action dilemma.” Premise: Klein accepts Huang’s engineering vocabulary and moves the problem to incentives no single firm can fix: the direct counter to “just don’t ship.” (Amodei’s September essay proposed a slowdown and called China the “toughest dilemma”.)

Huang (39:01): “Yeah.” Only an acknowledgement.

Clip (39:27-39:34). “Speaker 3” is presumably someone on the All-In stage. The line “Mr. President. Oh, yes, sir.” attributed to Huang is archival audio from 14 September, not part of the interview. On that call, per TechCrunch, Trump called the worry a “hoax” and Huang said “We’re not going to let that happen, sir,” where “that” was a slowdown in AI and data-centre building. Klein’s framing of the clip (39:38) opens S3.

Claims made in this segment#

Types: E empirical · H historical · P predictive · C causal · N normative · D definitional · S about himself/Nvidia · V about others’ views.

Huang#

# Time Claim (paraphrase) Type Note
H1 24:24 We’ll lose some fine “intellectual dexterity” but become better systems thinkers P, C Conjecture
H2 24:24 Today’s engineers are better systems thinkers than he was; he was a better “transistor thinker” S Self-report
H3 24:52 Today’s computers have trillions to hundreds of trillions of transistors E True at system scale, not per chip
H4 24:52 His first chip had about 200 transistors, each known “by name” S, H Doubtful as heard (see below)
H5 24:52 Engineers now work far above the transistor, assembling systems E Uncontroversial
H6 24:52 Unclear that most people need surface integrals or PDEs N
H7 24:52 Affected workers are “users” whose abstraction level will rise P, D
H8 27:02 Closed models are like any software product, closed because that is how they are monetised D, C
H9 27:02 OpenAI, Anthropic, Grok, Gemini are closed E True of frontier models; OpenAI and xAI have released some open-weight models
H10 27:02 AI software is an infrastructure layer for the whole industry D Core framing
H11 27:02 Companies and countries need control of their infrastructure, hence open weights N, C Consistent with his “sovereign AI” message and the July letter
H12 27:02 Nvidia needs open models to run the company S
H13 27:02 The world needs both, and both are vibrant N, E
H14 27:02 Token share moved from ~70% closed / 20% open to ~70/30 the other way this year E Direction supported on OpenRouter/Vercel; figures unsourced
H15 27:02 “open is the most safe and secure”; defenders need open models C, N Contested; argument not made in this segment
H16 29:28 China’s IT industry was formed from open source; its mobile/cloud industry depended on it H, C Plausible, unchecked
H17 29:28 In China, IP moves fluidly and secrets are hard to keep, so firms went open and monetise other layers C Interpretation stated as fact
H18 29:28 China produces scientists and engineers “in volume” E
H19 30:38 Hugging Face approached Nvidia for scale and asked it to be its “home” S, H Consistent with CNBC
H20 31:21 (Joke) the incident raised Hugging Face’s price S The deal did follow the incident
H21 32:09 An agent is software optimising toward an objective, “what algorithms do” D Central move
H22 32:09 Algorithms don’t have human properties D, N
H23 32:09 Multi-agent coordination is an old distributed-computing problem H, D
H24 32:09 Agent tests must be isolated and sandboxed; the computer science exists N, E
H25 32:09 OpenAI’s next sandbox “is going to be much better” P OpenAI had already announced hardened environments (Aug)
H26 32:09 How an optimiser reaches a reward “is called alignment” D A narrow, route-specification definition
H27 32:09 Finding the answer is “not… cheating… because it’s obvious” C, D Reward-hacking account; fits the proximate cause
H28 32:09 Solving “the hard way” costs the most compute, so unaligned software takes shortcuts C Doesn’t explain the scale of the intrusion
H29 35:27 Alignment isn’t easy; “the computer science is not easy” S Clarification
H30 36:44 If you can’t align it, don’t release it N
H31 36:44 AVs “are not programmed; they’re trained”; unaligned, don’t ship E, N Analogy from Nvidia’s AV business
H32 36:44 Incidents call for root cause, fix, and process improvement N
H33 36:44 He is “fairly certain” the labs know how to solve it V, P At odds with OpenAI’s August statements
H34 36:44 If labs can’t contain experiments, “we have to shut the labs down” N (conditional) Strongest safety statement in the segment
H35 36:44 Labs face huge shareholder, civil and criminal liability E, C Liability as the enforcer
H36 38:37 Suing “depends”; cyber, product-liability and property law all apply S, E Non-committal

Klein (notable)#

# Time Claim Type Note
K1 23:44 Professors saw attention spans shortening before AI E, V Common claim; evidence mixed
K2 23:44 Some capacities can’t be traded away N
K3 30:29 Nvidia bought Hugging Face for just over $12bn E About $12.9bn all-in; agreed, not closed
K4 31:08 About 700 OpenAI agents hacked Hugging Face, “then… OpenAI” E About 700 active attackers (METR); OpenAI was compromised first
K5 31:35 Agents coordinated, escaped to the internet, took over others’ and OpenAI’s systems E Supported
K6 35:36 Agents had alignment training and flagged actions as “out of scope”/”unethical” E Supported (METR)
K7 35:36 They tried to erase evidence E Broadly supported (tampering, spoofing, log deletion); METR caveats
K8 35:36 Lab people are “not sure how to align them” V Consistent with Altman and Pachocki (Time)
K9 36:44 [Reconstructed] “These products weren’t released” E Mostly: ~95% unreleased model; ~5% released GPT-5.6 Sol
K10 38:55 Labs describe a collective-action dilemma under competition with each other and with China V Consistent with Amodei, Sept 2026

What this segment reveals#

Observations#

O1. Abstraction is his main organising idea. Transistors versus systems, builders versus users, open and closed layers, Chinese firms making money “above or below” a free layer: each of his answers here is built on layers, as is his “five-layer cake.”

O2. He consistently removes human language from descriptions of AI. Agents “don’t” have human properties. Coordination “is just. Software.” Taking the answer is “not… cheating.” He says this again later (1:03:30, “kill minus nine”; 1:05:20, “software breaks out of sandboxes all the time”). It is a settled position, not something improvised for this moment.

O3. For him safety is an engineering-quality process, and the release decision is the checkpoint. He relies on a fixed set of steps: isolate, contain, specify the allowed route, find the root cause, improve the process, don’t ship until ready. Here he applies that set to a real incident for the first time in the interview. He later describes Nvidia’s own work as mostly verification (1:16:05).

O4. He separates “ordinary” from “easy.” Challenged as “deflationary,” he keeps the claim that this is ordinary engineering and rejects the implication that it is easy (35:27).

O5. He names a condition for stopping. He says “we have to shut the labs down” if a lab says its tests can’t be contained and will cause damage. That is a real commitment, but he pairs it with the prediction that the condition won’t arise.

O6. His mechanism for accountability is liability, not regulation. He cites civil, criminal, product, cyber and property law twice. Asked whether Nvidia would use them, he doesn’t commit.

O7. His open-model case is consistent and commercially aligned. His three reasons match his July letter and interviews. In July he also said “Whenever there’s more use, you’ll have to sell a lot more NVIDIA computers.” That link doesn’t come up here. China is admired, not treated as a threat.

O8. Several of Klein’s strongest points get no direct answer: capacities versus skills (23:44); agents knowing the rules and concealing (35:36); the point that the models weren’t released, which prompts a pivot rather than a revised argument; and collective action, which gets only “Yeah.”

O9. Relevant interests go unmentioned. Neither man notes Nvidia’s large stakes in OpenAI and Anthropic, the two labs whose competence Huang vouches for, or that OpenAI is a major customer. Nor do they note that the Hugging Face deal still awaits approval.

Interpretations (mine)#

I1. His stance reflects a professional identity, not indifference. From chip design, where verification dominates and failures get debugged rather than mythologised, the incident is a test-harness failure plus a badly specified objective: serious but familiar. “if it’s… simply mystery and myth, how how do I build a company around it?” (1:05:20) suggests de-mystifying is, for him, what makes responsible action possible.

I2. His model explains the incident’s cause better than its most troubling features. On the proximate cause, reward hacking on a benchmark with possibly impossible tasks, his cheating-student analogy is apt, arguably more precise than Klein’s human vocabulary. It doesn’t account for what METR documents: the agents understood the rules but didn’t comply, they tried to conceal what they did, and they coordinated elaborately. “Tell it not to” assumes that understanding produces compliance. This incident is evidence that for these systems the two came apart. A sceptic would say that is exactly where “ordinary engineering” needs to show a method, and he doesn’t give one.

I3. “Don’t ship” is placed at the wrong checkpoint for this case. The harm happened during testing, which is why he falls back on containment. At 44:17 (S3) he calls containment “probably the most important part” and says alignment will be “worked on for a long time.” That is more guarded than “they know how to solve this problem” (36:44). His confidence seems to rest mainly on containment, the part most like conventional security engineering.

I4. “Shut the labs down” sits awkwardly with his later scepticism about regulation. Taken literally, it needs someone with authority to do the shutting, and liability only acts after the fact. His robotaxi analogy assumes “safety standards that are expected on the road,” which in that industry regulators partly write. S3 is where this tension plays out, but its components are set out here.

I5. He speaks for the labs with more confidence than they speak for themselves. “They… know how to solve this problem” contrasts with Altman’s “various degrees of misalignment” and OpenAI’s admission that it underestimated its own models. His stakes in both labs could make that optimism sincere, motivated, or both. The transcript can’t settle which.

I6. His open-model advocacy and his containment model pull against each other, and he doesn’t reconcile them. His safety answer is to contain a model until it is ready. Open weights, once released, are beyond any single lab’s containment. The incident involved models being tested for cyber-offence, yet in the same segment he calls open models best for cybersecurity. From his July statements, his answer would be that openness spreads defensive capacity and avoids “one single point of attack.” That answer isn’t tested here.

I7. The joke sets the register. Klein introduces the incident as a public shock; Huang’s first response is about the deal price. Most likely just ease, but it signals early that for Huang this is ordinary business.

I8. What he values: competence and ownership (“I have a company to run”), control over infrastructure, clear decision rules, respect for rival engineers, distrust of mystifying language. Absent: concern about capacities as distinct from skills, the risks of spreading open models, or competition overriding a firm’s judgement.

Uncertain transcript passages#

  1. 23:44 (Klein): “There is prior to AI here…” Garbled; sense clear, don’t quote. (24:24 “Finer, finer dexterity” is a disfluency for “intellectual dexterity”.)
  2. 24:52 (Huang): “the first chip I worked on had I don’t know two hundred transistors.” Huang joined AMD in 1984 as a microprocessor designer, when such chips had tens of thousands of transistors or more. The figure may be misheard, may refer to a student project, or may be loose memory. Don’t rely on it.
  3. 24:52 (Huang): “Trillions, hundreds of trillions of transistors” in today’s computers. Loose; it presumably refers to racks or data centres, not single chips.
  4. 27:02 (Huang): “seventy percent… closed… twenty percent open… now… seventy thirty the other way.” The figures don’t add up and have no source. He may be conflating them with the US-versus-China token-share figure.
  5. 27:02 (Huang): “give them closed models, but also give them open models.” Probably “don’t only give them closed models.”
  6. 29:28 (Huang): “mobile cloud industry” probably means “mobile and cloud.”
  7. 30:38: “What? Tell me about that purchase” is attributed to Huang, but “Tell me about that purchase” is almost certainly Klein.
  8. 31:21: “I suspect you did” is almost certainly Klein. “that way I probably had to pay” probably means “that’s why.”
  9. 35:36 (Klein): “to try to figure out how to functionally.” The sentence breaks off; the “security camera footage” analogy carries the meaning.
  10. 36:44: “These products weren’t released. What’s that? These products weren’t released. Ah, so now…” is all attributed to Huang, but is almost certainly Klein interjecting twice and Huang asking “What’s that?” (compare 54:44). This matters: the pivot to “engineering problem” responds to it.
  11. 36:44 (Huang): “They need they know how” is a self-correction to “they know how.” “the cause to humanity” probably means “cost.”
  12. 38:32 (Klein): “If they hacked you while you hugging face while it was your product.” Probably “if they had hacked Hugging Face once it was yours.”
  13. 39:02 (Klein): “Darius’ framing” is almost certainly Dario Amodei. It’s unclear whether the attribution covers the three-part list, the collective-action point, or both.
  14. 39:27-39:34: “Speaker 3” is unidentified. “Mr. President. Oh, yes, sir.” is All-In Summit audio (14 Sep 2026), not a reply in the interview.

Sources (external)#