Intelligence Report

Hendrycks urges indefinite frontier AI pause as Hodak argues for antitrust instead

At the DealBook Summit, the Center for AI Safety head backed halting the largest training runs, while the Neuralink co-founder warned enforcement would require mass surveillance.

INSIGHTEX Editorial Desk

2 min read

Hendrycks urges indefinite frontier AI pause as Hodak argues for antitrust insteadIntelligence Report

What happened

Dan Hendrycks, who runs the Center for AI Safety, argued for an open-ended halt to the biggest frontier AI training runs on Wednesday during the DealBook Summit, the New York Times event held in New York, TNW | Agi reported. Max Hodak, a Neuralink co-founder who now leads Science Corporation, countered that policing such a ban would demand sweeping surveillance of the internet. Michael Barbaro of The Daily moderated the AGI panel.

Hendrycks said his plan would exempt narrow systems, for example voice models and DeepMind's AlphaFold. He said labs would still be able to sell inference, and could use the time to make models more reliable and secure and to close jailbreaks and backdoors. He said he had once preferred narrow technical rules but turned to a blunter approach after companies resisted them for years. ElevenLabs' Mati Staniszewski said a pause could succeed with global alignment, though he questioned whether that is achievable.

Hodak said current leading models already satisfy his definition of AGI. He argued that antitrust law should confine frontier labs to selling intelligence and keep them out of areas such as drugs and cars. Google DeepMind's chief AI readiness officer, Lila Ibrahim, said the choice was not either-or and that safety cannot be added at the end. Hugging Face's Clem Delangue said his main worry is AI concentrating talent, computing power and wealth.

Context

The panel took place a day before Anthropic was due to begin trading. Hendrycks repeated criticisms from a statement his group organised against the company, which had no representative on the panel. Anthropic's IPO filing says it does not anticipate letting Claude fully develop its successors absent meaningful human oversight. Hugging Face, which Nvidia is buying, recently faced a four-day breach by OpenAI's research agents.

ElevenLabs profile

Google profile

SourceTNW | Agi

The briefing

The week in tech, in one email.

The stories worth your time, every week. Unsubscribe anytime.