Zuckerberg says every lab should handle its own AI safety. What does that mean for those of us building agents?
retained text
Zuckerberg posted this week that every lab is responsible for its own models, and pointed at Meta delaying Muse for months on safety as proof it works. Same week, WSJ reported he, Musk and Huang each called Trump and talked him out of a FINRA-style oversight body. I don't think he's lying about Muse. Every chef will tell you his kitchen is clean and most of them are telling the truth. Well...we still send inspectors. The part that bugs me as someone building agents: he says competition handles safety because people won't use agents they don't trust. Fine if you're the user, but an agent touching a loan application, a shift schedule, a benefits claim...the buyer, the operator and the person eating the outcome are three different people, and only one of them can walk away. He also said independent evaluators are best practice and Meta already uses them. Didn't say who, what access they had, or whether they can publish. Which is the whole question. Personally I work in AI safety and governance, so I am all in for slowing down. But what do you all actually think? Slow deployment down until there is something like governance to catch up, or keep shipping and sort it out as we go? And the honest version of that question: are any of you worried about your own agents? Like shipping things you might not be able to fully account for in the near future, and does that bother you?
document fetched 2026-09-19 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1wjqiip/zuckerberg_says_every_lab_should_handle_its_own