Jensen Huang says AI doomer math is made up

Jensen Huang told All-In Podcast the 10 percent extinction math is made up.

The Nvidia founder, president and chief executive used the appearance to argue that narrow superintelligence is already here, that frontier labs remain in control, and that the current panic is not grounded in science. He was speaking on the show hosted by Chamath Palihapitiya, Jason Calacanis, David Sacks and David Friedberg that calls itself the number one podcast in the world.

Huang was responding to a weekend essay discussed as Dario’s essay and to a whistleblower case he referred to as Coxin. He said safety is paramount and whistleblowers must be taken seriously, but the scientific prediction layered on top was irresponsible. He then listed forecasts that did not materialize to make his point.

Those included a five-year prediction that artificial intelligence would replace all radiologists, a 6 to 12 month prediction that 90 percent of code would be generated by AI, and a 6 to 9 month prediction that 50 percent of entry-level jobs would be wiped out. He added older warnings that GPT-2 and Llama 3 would be too unsafe to release, and the claim that half of white collar jobs would vanish the next year. All proved wrong, he said, and AI has instead increased demand for radiologists while automating scan reading.

The timeline matters because Huang ties risk to compute concentration. Actual problems so far have come from frontier labs, he said, because only they have enough compute to push the frontier. A high school student or a typical startup will not, which narrows where controls need to apply. That is the before picture. The after picture, in his telling, is a transition from research to engineering. The labs are hair on fire, building culture and products at once, but each of the four incidents at one lab and the single large incident at another can be root caused and instrumented with sandboxes, runtimes and continuous monitors. The skeptical note is right there in his framing. Control is assumed to be achievable, yet the labs are still learning engineering discipline.

That assumption underpins his take on recursive self improvement. The news peg on All-In Podcast was Zhipu, the maker of GLM, raising about 5 billion dollars with plans to put 3 billion toward a run that automates AI building AI. Huang called RSI a catchphrase for sensible techniques that already exist inside companies, including in-context learning, skills, reflection, reinforcement learning, synthetic data generation and low-rank adaptation, or LoRA, that improves weights without retraining the base model. It will not spiral out of control on its own, he argued, because any product still has to be evaluated, tested for regressions and monitored before release. Internal iteration is allowed, external release is gated.

On open versus closed, Huang said the world needs both. He compared closed frontier models to bottled water, useful and worth paying for, while open models enable sovereignty, privacy and broad entrepreneurship. In the last six months 400 billion dollars of venture funding went into AI native companies and 80 percent of them use open models, he said. Much of today’s open source contribution comes from China simply because of scale, but once downloaded the model is yours to fork and own, as with Linux and Kubernetes. The race, in his view, is not who invents but who exploits the technology best, as the United States did with electricity and the internet despite European inventors like Maxwell, Volta and Ampere.

The broadcast then veered when President Trump called in on speakerphone. Trump told the crowd the doomer narrative is a hoax, that robots will not take over the world, and that data centers are the oil of the next 20 to 25 years, bigger than the internet, making dying communities wealthy. Huang agreed, added that AI is creating jobs and that reindustrialization depends on energy growth, and said he would tell the president that America needs every company, industry, state and researcher to win the AI race, not just a few labs. Trump repeated his line that whoever wins AI wins and said his administration will not let restrictions that block permitting and building happen, citing Google’s interest in Finland as an example he opposes.

What remains is measurement. Huang and guest Satya Nadella’s camp both pointed to basics first, standardization, evaluation and getting engineering right, before writing broad regulation. Huang favors third party evaluators modeled on financial auditors, ideally multiple firms to avoid capture. Until those evals and the engineering controls for frontier labs are in place, the policy debate will keep running on predictions rather than on demonstrated harms.