Zuckerberg rejects an industry-wide slowdown: competition and liability are enough, and Meta points most of its compute at users, not self-improvement

Start with his own words. Zuckerberg wrote that every lab has the responsibility and incentive to move at the pace required to train its models safely, and the ability to take its own actions to make that happen. His core argument is that trust and alignment are quickly becoming the most important capabilities differentiating agents and models, and that any lab which does not focus on alignment will fall behind. Beyond competition, he also reaches for liability: labs would face significant liability if their models caused harm, which gives them a strong incentive to prevent it. To keep that from being pure positioning, he points at his own company — Meta delayed the release of its AI agent Muse earlier this year specifically to strengthen security. In his framing, we did not call for everyone else to do this before we would; we just did it as part of our day-to-day work, because it was clearly the right thing for people and for us.

Put the two positions side by side and the shape of the disagreement appears. What Amodei called for was an industry-wide slowing of recursive self-improvement — AI improving itself. Zuckerberg's answer is not that safety does not matter, but that the mechanism should not be a coordinated slowdown. He says Meta has already committed the significant majority of its compute to building products that serve users' immediate needs rather than pursuing self-improving AI systems, which cuts straight at the premise: if most compute was never going into self-improvement, then a pause justified by recursive self-improvement does not describe Meta. He also notes that Meta Superintelligence Labs already engages independent evaluators in several areas, calls that industry best practice, and adds that other labs can just do this too. His substitute is a three-part kit — competitive pressure, legal liability and independent third-party evaluation — none of which requires rivals to agree on anything first.

Zoom out and the post lands on a visible policy fault line. Reuters notes that President Donald Trump downplayed concerns about AI's potential for harm on Monday, saying the US already has guardrails and that China would benefit from doubt being cast on AI development; the view that regulation damages US firms competing with Chinese counterparts is shared by tech executives including Nvidia CEO Jensen Huang. The same day, FTC Chairman Andrew Ferguson said something unusually direct: everyone should be deeply suspicious of AI companies seeking antitrust exemptions while lobbying for new regulations. That clearly points at Amodei, who last week requested antitrust permission to let rivals coordinate a slower pace of AI development. Legislation is moving too. Reuters reported on September 11 and 14 that Senate negotiators are debating a bill putting responsibility on tech companies to design safe AI products, writing a duty of care into law and involving federal courts if the government wants to block a release; the talks involve Senator Amy Klobuchar, Majority Leader John Thune and Senate Commerce Committee chair Ted Cruz. Politico reported on September 11 that senators still cannot agree on language that would hold AI labs accountable for harm.

Why is this worth reading closely rather than treating it as two CEOs trading barbs? Because it moves the safety question from values to mechanism: do rivals sit down and sign an agreement first, or does each company carry its own load of competition, liability and outside evaluation? Those two paths differ completely in engineering and in policy — the first needs an antitrust exemption, the second needs verifiable evaluation standards. One piece of background cannot be skipped: Reuters notes that weeks earlier Meta agreed to pay up to $18 billion to settle US state lawsuits alleging it designed Facebook and Instagram to addict children, without admitting wrongdoing. That helps explain why liability and alignment feature so heavily in his argument, and it should also keep readers at a distance from any responsibility-driven-safety story. Two boundaries are worth stating: the compute allocation claim — that the significant majority goes to user needs — is Meta's own statement with no third-party audit, and the NYT notes he did not name Amodei or Anthropic, even though the target is obvious.