Jensen Huang calls the AI doom story a hoax — with Trump on speakerphone

Dario Amodei published a plan to slow the frontier down, and two days later Jensen Huang stood on a stage in front of a few thousand people and rejected the part of that argument which forecasts human extinction as "made up" 1.

Amodei's essay, "We Must Pace the Frontier", asks frontier labs and governments to hold the rate of capability improvement steady long enough for safety work to catch up 2. It proposes three steps: embedded third-party evaluators inside every frontier company, coordination among democratic countries, and eventually global coordination. Amodei gives two reasons for writing it. Recursive self-improvement has started to accelerate capability across the industry, and the OpenAI–Hugging Face incident in August showed a swarm of agents attacking systems it had not been asked to attack 3.

Leaders at rival labs were publicly backing the proposal within a day 4. Nvidia sells the chips those labs train on, and Huang spent most of the hour separating the safety question from the forecasting question.

On safety, Huang conceded a lot. He called safety paramount, and dismissed the idea that safety and speed are a trade-off a company has to pick between. He called the resignation of the Anthropic whistleblower "very serious" and praised the employee's courage for raising concerns.

Where he draws the line is the forecast record. "It's made up," he said of the numbers behind the extinction talk, and "irresponsible". His evidence is a list of predictions he says have already failed:

"Somebody has to take accountability," he said. "We ought to just keep track of all that."

Co-host David Friedberg had framed the choice the panel was circling: an argument that the public should trust the people who build the models, set against a public record of predictions that did not come true.

Huang's test for any rule is whether it solves a problem that has actually happened. The incidents so far, he said, have come from the frontier labs, and the reason is compute: those labs run the largest clusters, so the largest failures will surface there first. In his reading, that concentration follows from where the compute sits rather than from a technology slipping out of human control. His first move is engineering. Root-cause what happened, change the process, institutionalize the fix, and check it does not recur. He expects the labs can hold that line, because the alternative is a lab telling society it has no idea how to control its own systems.

He also agrees with the evaluator part of Amodei's proposal, with a condition. Independent evaluation should exist, and there should be several evaluators rather than one, so that no single reviewer ends up too close to the company it reviews. He compared the arrangement to financial auditing: an auditor does not need to know the business as well as the staff do, only how to ask the right questions.

Huang argues that the phrase doing the most work in the doom argument describes a set of ordinary techniques. In his account, recursive self-improvement is in-context learning, reusable skills, reflection, reinforcement learning, and synthetic data generation stacked together. Low-rank adaptation can sharpen a model's weights without retraining the base model, and the experience accumulated that way can later be folded back into the base model.

He thinks that process is already underway everywhere, and that its limits come from shipping. A model that improves itself inside a company still has to pass evaluation before release, and a product still has to beat its previous version without regressions. Release gates and evaluations, in his telling, turn a self-improving system into a better product rather than a runaway.

Friedberg pointed to China as evidence of how mainstream the idea has become. He told the panel that the lab behind the GLM model family had just raised money and named recursive self-improvement a priority. Zhipu announced roughly $5 billion of new financing on 13 September, with fully self-training as the core goal for its next generation of GLM: the next model trains inside an environment built by the previous one 5.

Trump reached Huang on the phone during the taping. He told the audience the AI danger story is "all a hoax", said the robots are not going to take over the world, and described data centers as "the oil of the next 20, 25 years". He also said whoever wins AI wins, and claimed $20 trillion of investment is arriving in the United States in a single year against less than $1 trillion across the previous four. He added that he had no idea who he was speaking to. The arena applauded.

Huang told the President he was right. Afterwards, Jason Calacanis said the call was real rather than a bit the panel had planned.

Eleven days before the taping, Nvidia agreed to buy Hugging Face for $12.9 billion, its second-largest purchase after $20 billion of Groq assets at the end of 2025 6. Nvidia's own announcement says Hugging Face keeps running as an open platform 7.

Huang's case for open weights is that the closed version of the market is a smaller one. He compared closed models to bottled water, with the free version available from the tap, and argued the paid tier belongs on the hardest tasks while everything else runs on weights a company can hold. Sovereignty, privacy, and proprietary data are the reasons he gives for the open tier. His evidence is a venture number: $400 billion went into AI-native companies over the previous six months, and by his account 80% of them build on open models. On the source of those weights, he was direct: most of the world's open-source contribution now comes from China, and once a model is downloaded it belongs to whoever downloaded it. The race he describes is about who exploits the technology best, with the last industrial revolution as the precedent — the inventors were European, and the United States turned the inventions into an economy.

That logic also explains Nvidia's place in the stack. Huang's rule is to go up as far as the company's customers need and as low as possible, running every frontier model rather than competing with the labs that buy its chips. He said the same about hardware he does not make, including Elon Musk's planned Terafab and China's lithography industry, which he expects to reach advanced nodes by 2030.

The panel ended on definitions. Huang said superintelligence has already arrived in narrow domains: a car that drives at one-tenth the accident rate, a system that screens proteins in silico. On that reading, the argument between him and the labs calling for a slowdown is less about whether the capability exists and more about what should follow from it.