
Hendrycks calls for an indefinite pause on the largest AI training runs at DealBook
By Zain·
At the New York Times DealBook Summit, Center for AI Safety director Dan Hendrycks urged a pause on frontier training runs. Neuralink co-founder Max Hodak said enforcement would require a surveillance state.
Dan Hendrycks, the director of the Center for AI Safety, told a New York Times DealBook Summit panel that the United States should consider an indefinite pause on the largest AI training runs, and Neuralink co-founder Max Hodak answered that enforcing one would take "an exhaustive surveillance state on the internet." The exchange, reported by The Next Web on November 21, 2026, is a clean statement of the two positions now dividing the debate over a frontier AI pause. This piece reports the panel; our dated log of every public call to slow down sits in a separate tracker.
What Hendrycks proposed
According to The Next Web's account, Hendrycks argued that the US midterms showed strong voter feeling on AI, which opens room for politicians to discuss pausing development. He said he once favored narrow technical rules, but companies blocked them for years. "So now it's, we're going to need a more blunt intervention. That's how it is" he said.
His proposal was an indefinite pause on the largest training runs, with carve-outs for narrow models such as AlphaFold and voice models, and with inference sales allowed to continue. He accused Anthropic of having no safety plan: "What they're doing is reckless and they don't have a safety plan. They know it." The Next Web adds that he said some safety groups privately agree with the criticism but will not sign on for fear of losing funding or access, and that moderator Michael Barbaro noted Hendrycks had been a safety adviser to several AI companies until hours before the panel. We have not seen a response from Anthropic to the panel remarks, and the claim about private agreement is unattributed.
The case against, as Hodak put it
Hodak's objection was practical. A training ban, he said, would require an exhaustive surveillance regime to enforce. "I don't think that the genie is going to be put back in the bottle" he said. He argued that current top models already meet his own definition of artificial general intelligence, naming Mythos, Fable, Astra and Sol; that governments will keep a "superintelligence frontier" while others get an "industrial intelligence frontier"; and that antitrust law, limiting frontier labs to selling intelligence rather than entering drugs, cars and other products, could do most of the work without a new regulator.
“So now it's, we're going to need a more blunt intervention. That's how it is”

The rest of the panel
The Next Web says the rest of the panel sat between the poles. Mati Staniszewski of ElevenLabs agreed a pause could work with global alignment but said whether that is achievable "is the big question mark." Lila Ibrahim, Google DeepMind's chief AI readiness officer, said the choice is not either pausing or something else, argued safety cannot be added as an afterthought, and said she had never seen global trust in technology as low as now. Clem Delangue of Hugging Face, who joined with 18 minutes left, said his main concern is concentration of talent, compute and wealth, and said Hugging Face came out of the OpenAI-agent attack more optimistic: "People only see that we got attacked by AI and they rarely see that we defended ourselves with AI." Barbaro's summary, per the report, was that the consensus on stage was no moratorium and less concentration of power, with Hendrycks the exception, and he raised China's likely non-compliance with any moratorium.
Where the argument stands outside the panel
A pause is no longer a fringe ask. A statement titled "Pacing the Frontier" lists 1,386 signers from frontier AI companies, who say the world lacks the tools to deliberately pace the frontier and ask the US government to back an international effort to build them. Defense One reported on November 19 that METR's assessment of AI research acceleration at Anthropic and updated forecasts have led some researchers to shorten timelines. Several of those threads are covered in our reporting on the nonprofits suing to halt frontier AI development.
Note what a "pause" means in these debates. Hendrycks's version is narrow in one way, only the largest training runs and not inference sales, and broad in another, indefinite. The Pacing the Frontier statement does not ask for a halt; it asks for tools that would give governments and industry the option to slow down. Mixing the two is common and misleading.
What the evidence supports
On enforcement, Hodak's point has weight but is not decisive: compute is physically large and concentrated, which is why pause advocates point to it, and the panel record contains no technical analysis either way. On Hendrycks's charge, "no safety plan," the panel provided assertion, not a document; Anthropic's own filings and published frameworks would be the test, and this report does not evaluate them. The politics claim, that midterms opened room, is Hendrycks's reading, and we did not check polling for this piece.
“I don't think that the genie is going to be put back in the bottle”
What is solid is narrower: a leading safety researcher publicly called for an indefinite pause on the largest runs, a founder of a neurotechnology company called it unenforceable, and the moderator found a majority against.
Why the venue matters
DealBook is a business conference, and the audience is investors and executives, not legislators. A call for a pause delivered there is aimed at the people who fund and sell frontier models, at a moment when the largest US IPO on record, Anthropic's, was about to start trading. That context is the moderator's, not ours, but it explains why the exchange turned so quickly to enforcement and to who profits from concentration.
Our take
The panel did not move policy, but it shows the pause question has shifted from whether to who can verify. Watch three things: whether any member of Congress turns Hendrycks's carve-out model into bill text, whether the 2027 regulatory fights he predicted are visible in committee calendars, and whether the labs name a verification method that does not depend on trust. Without one, "pause" and "no pause" will both remain slogans.
Frequently asked questions
Should frontier AI development be paused?
At the DealBook panel, Dan Hendrycks said yes for the largest training runs, indefinitely, with carve-outs. Max Hodak and most of the panel disagreed, citing enforcement and China. Moderator Michael Barbaro summarized the consensus as no moratorium.
What exactly did Hendrycks propose?
An indefinite pause on the largest training runs, with carve-outs for narrow models such as AlphaFold and voice models, and with inference sales allowed to continue, according to The Next Web's report.
Why did Hodak say a pause cannot be enforced?
He said it would require an exhaustive surveillance state on the internet. He also argued antitrust law could limit frontier labs to selling intelligence.
What is Pacing the Frontier?
A statement listing 1,386 employees of frontier AI companies. It says the world lacks tools to deliberately pace the frontier and asks the US government to support an international effort to build them.
Did Anthropic respond to Hendrycks?
We have not seen a response. The Next Web notes Anthropic was not on the panel and that its IPO filing says it does not expect Claude to fully build its successors without meaningful human supervision.
Sources
What each one is, and whose it is.
- 1
Dan Hendrycks calls for a pause on frontier AI, Max Hodak pushes back, The Next Web (November 21, 2026)
Press reportIndependent of the vendor - 2
Pacing the Frontier, Employees of frontier AI companies (November 22, 2026)
OtherIndependent of the vendor - 3
'I was seriously wrong': Experts say superintelligent AI may arrive as soon as next year, Defense One (November 19, 2026)
Press reportIndependent of the vendor