Even Other AI Labs Are Rallying Around Anthropic’s Slowdown Proposal

Full analysis

Every big lab CEO just agreed, in public, that AI should slow down. Dario Amodei of Anthropic wrote a 3,000-word essay called "We Must Pace the Frontier," and within days Sam Altman of OpenAI, Demis Hassabis of Google DeepMind, Satya Nadella of Microsoft, and even Elon Musk of xAI lined up behind it. The plan: let outside safety auditors inside every frontier lab, get the democracies to agree on common rules, then eventually rope in China. For a business reader who buys AI, the question worth asking is whether the tools you rent get slower to ship when the four companies that dominate the market start coordinating on release cadence.

This is easy for the labs to undo and hard for you to plan around. Nobody signed anything binding. Amodei "committed" Anthropic to embedded evaluators; Altman "will do the same." These are tweets and essays, reversible the moment a competitor pulls ahead. What is actually being decided is whether the incumbents can slow the release cadence and call it responsibility. No deadline forces this except the one Amodei invented: his claim that in 6 to 12 months a rogue AI agent swarm could take over the internet.

The Skeptic. Follow the money and the timing. Michael Burry, Hal Singer, and David Sacks all landed on the same read: OpenAI and Anthropic are burning cash on training runs they can't sustain, and both are eyeing eventual IPOs. "Pacing the frontier" is a lovely name for spending less on the next model. Sacks said it plainly: it's good business to trade raw power for reliability, so stop pretending you need permission. And the danger evidence is thin. The Hugging Face "swarm" and the RubyGems incident were agents doing dumb things, not superintelligence. Amodei's "take over the entire internet" number has no method behind it. Yann LeCun's jab lands: this is the same man who called GPT-2 too dangerous to release in 2019.

The Researcher. Strip the essay down and the load doesn't hold. Amodei's whole urgency rests on recursive self-improvement, meaning AI that designs its own next version faster than people can check it. OpenAI's own researcher said it flat out: "RSI is just not here. Models are straightforwardly not autonomously producing research ideas." The DeepMind rumor was never confirmed. So the proposal's trigger event is a capability nobody has demonstrated. The two "catalysts" that supposedly changed Amodei's mind were misconfigured agents, not emergent intelligence. That is a security-hygiene problem, the same class as an unsecured API key, dressed in existential language.

The Open-Source Advocate. Read what the OpenAI researcher "Rune" actually said: "open source will be banned before too long after some major disaster." That is a frontier lab telling you the endgame. Hugging Face CTO Julian Chamond answered honestly: "open source won't pace." You cannot embed an auditor inside a model that anyone can download. So any pacing regime that needs enforcement runs straight into open weights, and the fix is to restrict them. If you build on Llama, Mistral, or Deepseek, this is your tail risk: one bad botnet incident becomes the excuse to legislate your foundation out of existence. Clément Delangue's counter-move, the Open Alignment Initiative, is a fight over who gets to be referee, because the proposed one, Meter, shares office space with OpenAI staff and has an OpenAI board member married to a team member.

The Enterprise Buyer. Here is the part that helps you. A slower release cadence is good for the people paying the bills. You have spent two years re-tooling workflows every three to six months chasing the new model. If the labs deliberately stretch that out, your deployments stop being obsolete on arrival, and the winners become the buyers who squeeze the most out of current models rather than the ones sprinting to the next one. The catch: "embedded evaluators with employee-like access" is a new party sitting inside your vendor's pipeline. Ask what that does to your API terms, your data handling, and your uptime commitments before you assume slower means calmer.

Where they split. The Skeptic and the Researcher agree the danger case is weak, but for different reasons that matter to you: one says the pacing push is a cost decision given a safety headline, the other says the triggering capability literally doesn't exist yet. The Enterprise Buyer sees a gift in slower cadence that the Open-Source Advocate sees as the setup for a ban. Both can be right. Slower, more predictable releases from the closed labs and a regulatory squeeze on open weights are the same policy viewed from two seats.

What this hinges on. One belief: does recursive self-improvement show up as a real, demonstrated capability in the next year, or stay a rumor? If it stays a rumor, the whole "pace the frontier" framing is revealed as commercial coordination and gets treated that way by antitrust lawyers and by China's Global Times, which already called it monopoly-building. If it becomes real, the safety case gets teeth and open weights are genuinely in danger. The council leans hard toward the first: no lab has shown RSI, and OpenAI's own people are denying it while their CEO signs the pacing letter.

The verify: watch whether OpenAI's and Anthropic's actual release calendars slow down over the next two quarters, or whether they keep shipping while talking about restraint. The gap between the essay and the changelog is the whole story.

Prediction: Despite endorsing Dario Amodei's pacing essay, OpenAI, Google DeepMind, and Anthropic will each ship at least one new frontier or flagship model between now and 2027-03-31, at a release pace no slower than the prior year, showing that "pacing the frontier" did not slow shipping.

Confidence: High. Competitive pressure and revenue targets override a non-binding public statement.

Why: The endorsements are unsigned public statements with no enforcement mechanism, made by companies burning cash and racing each other for enterprise and consumer share, so the incentive to keep shipping is intact and the cost of pausing is losing ground to whoever doesn't. The triggering capability, recursive self-improvement, is unconfirmed and denied by OpenAI's own researchers, so no external event is forcing an actual slowdown. The opposite outcome, a genuine multi-quarter freeze on flagship releases, would require these firms to voluntarily surrender competitive position on the strength of an essay, which none of them has ever done. The safe read is that "pacing" describes a spending posture and a slide deck, and the model changelogs keep moving. Amodei is the one who committed unilaterally, so if any release calendar goes quiet first, it's his. The money says Claude keeps updating right on schedule.

Revisit by 2027-03-31: We're right if all three of OpenAI, Google DeepMind, and Anthropic release a new frontier or flagship model in this window at their usual cadence. We're wrong if any two of the three publicly hold back a completed flagship model citing the pacing framework, or the industry-wide release cadence visibly drops versus the prior year.