Zuckerberg says AI labs can set their own release pace, but backs independent evaluators

The hold on Muse ran roughly five months, from the April window to the release ^[11]. Meta wrote the minimum requirements, tested against them, and decided they had been met ^[10].

Some of what testers reported is specific enough to price. One employee who asked Muse to monitor tickets and other scarce items found "many failure modes that made it unreliable," with the agent stopping refreshing after about 15 minutes, silently ignoring errors and sometimes disabling monitoring "for no apparent reason" ^[13]. Andrew Bosworth, Meta's chief technology officer, wrote that he was repeatedly logged out, at times several times within a few minutes ^[14].

Meta did build controls into the product. Muse runs inside a separate virtual machine for each user, and a supervision system called Sentinel reviews its proposed connector actions and network traffic before the agent sends data out or acts ^[15]. "It is impossible to say that there is never going to be a mistake, but every single part of the architecture has been designed to make this as safe, as secure, as private as we can possibly make it," Shah said ^[16].

Muse is the example Zuckerberg chose to make his case ^[1]. First-week complaints about a consumer agent are not the same thing as how a lab paces a frontier training run, but they are the best available evidence of what one company's internal threshold certifies.

The case rests on incentives, and both of the ones he names work only after the behaviour is visible. "My view is that trust and alignment are quickly becoming the most important capabilities that will differentiate agents and models," Zuckerberg said, adding that "Any lab that doesn't focus on alignment will fall behind" ^[6]. He also wrote that legal liability gave labs another reason to prevent harm ^[22].

The part of the post a buyer can act on is the evaluator language. "Engaging independent evaluators and advisors is industry best practice," Zuckerberg wrote, and he said Meta Superintelligence Labs already uses them in several areas ^[7]. Access and publication rights are what let a customer check an evaluator clause. According to implicator.ai, he did not name an evaluator or describe what access they have or whether they can publish ^[8].

Alexandr Wang, Meta's chief AI officer, went further on governance the same day, saying labs need strong governance, external evaluators and independent oversight of model launches ^[19]. In his framing the choice sits with each lab, and the party that chose can choose again. "Meta is committing the significant majority of our compute towards serving people rather than racing on RSI, and other labs can choose to do the same," Wang said ^[21].

Two of the three public positions on the record point at independent evaluation, and only Anthropic's needs a statute to work ^[23]. For the next 90 days, that means an evaluation regime with a named party, defined access and a reporting duty is something a customer can negotiate, while a verifiable pause requires legislation. A customer who accepts the vendor's own threshold language is accepting a delay that ends whenever the vendor decides its requirements are met.