What he said
Mark Zuckerberg told NBC News chief tech analyst Joanna Stern that leading AI companies can slow or pause their own development if they judge the technology unsafe, and that nothing beyond that is needed.
“I don’t think that we need some kind of industrywide coordination,” he said in the interview published on Thursday. “I think that each lab needs to take the time, and when it sees that there are issues, you just take the time that you need internally to basically make sure that you’re proceeding safely.”
His argument for why that is sufficient is commercial. “I happen to think that there’s plenty of commercial incentive to get this right,” he said, and: “If you’re building a product, people aren’t going to want to use it if it doesn’t do what they want.”
What he is breaking with
The remarks land against a specific proposal. Anthropic chief executive Dario Amodei published an essay calling for intercompany efforts to “pace the frontier”, which Sam Altman endorsed alongside Google DeepMind co-founder Demis Hassabis and xAI’s Elon Musk. Amodei and Altman briefed the United Nations Security Council on Wednesday, arguing for cooperation on slowing the rate of progress to buy time for safety and monitoring work.
The immediate trigger for the debate was the resignation in early September of Jacob Coxon, a former Anthropic researcher who left over fears that AI systems could escape human control within a few years.
Zuckerberg is not the only sceptic. Critics of the pacing proposals have argued they are a Trojan horse that would let the largest AI companies shut out competitors and take control of how the industry is regulated. That objection does not depend on thinking the risk is fake; it depends on who writes the rules.
The product underneath the argument
Meta has a reason to want speed. Its Muse agent, released on 8 September, is now the most popular free app on both the Apple App Store and Google Play. It books flights, replies to emails, compiles research briefs and cancels subscriptions on a user’s behalf.
Zuckerberg’s answer to the safety question for that product is architectural rather than industry-wide: Muse is several agents, and a second Meta agent called Sentinel polices the main one. If Sentinel sees Muse potentially exceeding its remit, it can alert the user and ask for permission before anything further happens.
“The AI agent needs to be trained from the ground up to have a sense of what information is OK to share and what isn’t,” he said. Asked about the privacy exposure of an agent wired to email accounts and bank cards, he pointed to Meta’s record and to WhatsApp’s encryption: “We’ve shown that we have the ability to build those kinds of world-class, leading security systems.”
The position, plainly
On the larger question he is unambiguous. “There’s a lot of rhetoric that is filled with doom,” he said. “I’m quite optimistic about the fact that this is going to be a very positive future for everyone.”
That is a real disagreement between the people running the largest labs, stated in public, about whether the current pace is a problem at all. It is also, for now, the position of the company whose agent most people have actually installed.
