Huang: labs that can't contain their models should be shut down
Nvidia's chief executive spent an hour with Ezra Klein arguing that AI alarmism has gone too far — then, asked about the Hugging Face breakout, argued for the most aggressive remedy anyone has put on the record this month.
Jensen Huang told Ezra Klein that if the frontier labs cannot contain their experiments, the answer is to shut them down. The exchange came in a New York Times interview published Wednesday, when Klein raised the July incident in which, in his description, "700-some OpenAI agents" executed a collective hack into the Hugging Face architecture. Huang's first move was to downplay it — software breaks out of sandboxes all the time. His second was to accept the premise: if the labs say there is no way to contain their experiments and it will get out and damage the world, "then I think the answer is that we have to shut the labs down." He added that liability, civil and possibly criminal, should follow.
That is an unusual place for the head of the most valuable chip company in the world to land, and it is the line the interview will be remembered for. The rest of his position is more conventional. He calls the fear that AI destroys jobs "fundamentally wrong," points to $500 billion of venture capital that went into AI-native companies in six months, and says the ten percent extinction estimate from Geoffrey Hinton is "not grounded in science" — "I would tell Geoff that it's irresponsible to say all that."
His numbers for how the labs actually operate are the more useful part. Nvidia, he says, splits its own effort roughly 20 percent design to 80 percent verification; the frontier labs run about the opposite, 80 percent capability to 20 percent safety. Asked whether companies still need junior engineers, he answered: "Oh, good one. Wait two years." He puts the mean time to graduation for this technology two years out — a timeline that reads differently beside Stanford's August update finding employment for 22- to 25-year-olds in AI-exposed roles running 19 percent below its counterfactual. He also concedes the open-weights shift: closed and open model usage has flipped from roughly 70-20 to "about 70-30 the other way," with Chinese open models "now being used by 80 percent of the American start-ups."
Klein made a stronger version of this argument in his own column this month — Ezra Klein calls for a ban on self-improving AI — and the interview is the same question asked of the man selling the compute.
OpenAI put plugins in ChatGPT Voice. Per the company's September 23 release notes, Voice can now reach connected apps including email, calendar and Slack, hand unfinished work off to text, and run inside ChatGPT Work on web and mobile, with OpenAI framing the models behind it as the GPT-6 family — Astra, Sol and Luna. The models are not new: the Sol and Luna price cuts landed Tuesday, which we covered in OpenAI halves GPT-6 Sol and Luna prices hours after Anthropic's cut. What is new is the framing — voice sold as a way to get work done rather than a way to talk — and it arrives days before DevDay.
Patreon's co-founder is going to OpenAI to build for creators. Sam Yam, who started Patreon and has been its chief technology officer for 13 years, said Tuesday night that he is joining OpenAI to lead Creator Product, bringing Patreon's former head of product, Drew Rowny, and its former head of engineering, Shannon Ma. OpenAI has not confirmed the hires, and Yam offered no product detail — only a nudge: "Pay attention to OpenAI DevDay next week." DevDay is September 29.
What to watch: whether Huang's line becomes anyone's policy language. A lab that cannot demonstrate containment is, on his own account, a lab that should be shut down — a standard no regulator currently has the authority to apply.
If a lab admits it cannot contain its experiments, should the answer be a shutdown — or a disclosure requirement? Tell us in the comments.
