Summary

Jensen Huang on the All-In stage, mounting the most direct public rebuttal yet by a principal to the AI-doom position, prompted by an essay from Anthropic's Dario Amodei and the Coxen whistleblower episode. The venue is friendly to the point of sycophancy - the hosts call him GPU Jesus, President Trump telephones mid-interview and is put on speaker - and Huang's commercial interest in AI proceeding at speed is total. Both things are true at once, and the arguments are still the sharpest available on that side.

• Huang's position on quantified extinction risk is blunt: 'we shouldn't because it's made up', and putting such a number out is 'alarming and troubling' and 'irresponsible'. He separates this from the whistleblower, saying Coxen showed 'great courage' and that any whistleblower must be taken seriously.
• His argument rests on a track record of failed predictions, listed: radiology would be entirely taken over within five years - the world now needs more radiologists, though automated scan reading did arrive; 90% of code generated by AI within 6-12 months; 50% of entry-level jobs wiped out within 6-9 months; GPT-2 and Llama 3 too unsafe to release. 'We have to take account for all of the stupid predictions that were made.'
• The structural claim about where danger actually sits, and it is a good one: every actual problem so far has come from the frontier labs, because they have the most compute. A high school student cannot cause one; nor can a startup. He offers this in the labs' defence, not as an attack.
• His prescription is engineering rather than legislation: root-cause each incident, institutionalise the fix, and he 'would bet money' each was within the labs' control to prevent. He names the alternative explicitly - that they analysed the incidents, concluded they don't know what happened and cannot control it - and says he doubts it.
• On recursive self-improvement he is dismissive of the spiral framing: RSI is a set of sensible existing ideas - in-context learning, skills, reflection, RL, synthetic data, LoRA - and the reason it will not run away is mundane. 'You could RSI all day long inside your company, but when you release a product, you've got to evaluate it, don't you?'
• He does support third-party evaluation, by analogy to financial auditors: they need not be as expert as the company, they need to ask the right questions, and there must be multiple so no single evaluator can be captured.
• On open versus closed models: 'closed models is kind of like bottled water... water is free.' Both are needed, for sovereignty, privacy and proprietary reasons. The figure cited is $400bn of venture funding into AI-native companies in six months with 80% using open models - 'if not for open models, how could they build their dream?'
• On Chinese open models he declines the national-security framing: most of the world's open-source contribution today comes from China because they have more engineers, 'and once you download it, it's yours. We fork it. We improve it. We make it ours.'
• His definition of the race is the strongest idea in the interview: the last industrial revolution's inventors - Maxwell, Volta, Ampere - were not American, and America won it by exploiting it better than anyone. 'The race is really about who exploits the technology best.'
• On why the pessimistic narrative travels: he says China's is simply more practical - AI as a technology that advances the economy - with nobody there saying it ends civilisation.
• President Trump telephoned mid-interview and was put on speaker, calling AI fear 'a hoax' twice, saying 'the robots are not going to be taking over', and repeating 'whoever wins AI wins'. Recorded because it happened live, not because it constitutes evidence.
• Nvidia's strategy, stated plainly: 'go up as far as we need to and as low as possible.' He claims Nvidia is the frontier model in five domains, and justifies each by customer need rather than disruption - Alpamo for carmakers too small to build a self-driving stack, protein models because Lilly and Merck need them.
• On China's domestic lithography his answer is a flat date: 'They're going to get there by 2030', with the reasoning that China is very good at high-volume production and this is a matter of time.
• Asked whether we are at AGI he says 'I think we're already there', and at superintelligence too when segments are taken narrowly - his example being a self-driving car at a tenth of the human accident rate. 'I don't want you to make me an omelette, I just want you to drive the car.'

Why it matters

This is the best-argued version of the accelerationist case the Handbook holds, and it deserves recording as an argument rather than as a position, because two of its moves are genuinely strong regardless of who is making them. The compute argument locates danger where it actually is - at the labs, because nobody else has the compute - and does so in the labs' defence, which is a harder thing to dismiss than the usual dismissal. The exploitation argument reframes 'winning AI' away from who invents and towards who deploys, with a historical precedent that holds up. Against that, the failed-predictions list is selective and Huang's interest in the technology proceeding without restraint is as large as any interest in the industry, so the confidence should be discounted accordingly. Note also that he and Musk, interviewed the same week at the same event, converge on third-party testing from opposite premises - Huang from 'they have this handled', Musk from 'the models are dangerous'. When those two agree on a mechanism, the mechanism is worth watching.

S7CrlFLAmEA-transcript.txt