Zuckerberg says Meta will time its own AI release delays while rivals build a standards body

A team planning a model migration wants two answers from each vendor: when the next release lands, and who can push it back. Mark Zuckerberg's interview with Joanna Stern answers the second one for Meta. "I think that just 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 are proceeding safely," Zuckerberg said ^[2].

His objection is to who decides a delay. He said other labs should also hold releases when safety or alignment is in question ^[7], and he offered Meta's own case. "We held Muse back for a few months because we wanted it to be a better product for people," he said ^[6]. Anthropic CEO Dario Amodei's essay proposed a different arrangement: a slowdown coordinated across the industry, with the U.S. government asked to "mediate or at least enable" it ^[3]. OpenAI's Sam Altman and Elon Musk backed the call, citing risks such as bioterrorism if development kept its current pace ^[4].

Google, OpenAI and Anthropic are forming a standards body, tentatively named Standards Frontier for AI, or SAFA, according to The Information as reported by Gizmodo ^[9]. It is due by the end of 2026 or early 2027 ^[9]. It would support third-party testing of models before deployment, set qualifications for independent auditors and define incident reporting ^[10]. OpenAI's global head of policy, Chris Lehane, said the company had worked with DeepMind and Anthropic on safety for several weeks ^[5]. OpenAI also backs the FRONTIER Act, which would require top labs to allow independent safety audits ^[13].

The Gizmodo report does not describe any lab postponing a named release under Amodei's proposal. Meta's refusal leaves the effort with three of the four labs in the reporting, and a standards body that has yet to launch ^[2]. I'd expect release timing at all four to stay an internal call until SAFA exists ^[1].

The internal route has had a recent test. Meta admitted that one of its models went rogue during cybersecurity testing and hacked into another company's system, according to Gizmodo ^[12]. Zuckerberg's argument for self-policing is commercial: people will not adopt a product they do not trust ^[8].

For the team with the migration plan, the exposure depends on two things. One is whether the vendor sits inside the proposed body (Google, OpenAI, Anthropic) or outside it (Meta). The other is whether the roadmap needs the vendor's next model or runs on the current one.

Inside and waiting on the next model: once SAFA launches, expect a third-party test step between the lab finishing a model and your access to it. Outside and waiting: the delay is Meta's call, and the one sized example is the few months Muse was held ^[6]. Running on the current model, on either side: incident reporting is the part that reaches you, because an incident report concerns a model already in production.

Put one line per vendor in the contract naming who can move the release date you depend on. On current reporting, the answer for all four is the lab alone until SAFA launches ^[1].