Altman puts OpenAI's pacing on paper — and drops the legal caveats
The week of frontier-lab statements about slowing down finally produced the thing skeptics asked for: a written position from OpenAI, published by the CEO himself, with a named mechanism attached — and a refusal to wait for Congress before using it.
Sam Altman published a long post Sunday night laying out how OpenAI will "pace" itself, and the headline is that the company says it needs no permission. "We welcome a federal framework that sets consistent safety requirements for frontier AI," Altman wrote, "but we do not believe we need to wait for an anti-trust exemption or legislation to begin the work of providing this confidence." The concrete commitment: OpenAI now formulates explicit safety cases in advance of frontier reinforcement-learning runs it expects to significantly increase capability — a pre-training gate, not just the pre-release reviews its Responsible Scaling Policies already required. He also defined the contested word directly. "When we talk about 'pacing', we do not mean 'stopping'... But it should be slower than it otherwise could be; interventions like safety cases and monitoring have significant costs," he wrote, adding that "no amount of American competitive pressure should justify recklessness."
That is a direct answer to the criticism that has dominated the debate since Anthropic's Dario Amodei called for slowing the frontier: David Sacks' argument that the labs are already free to pace themselves and asking for a legal cartel anyway reeks of regulatory capture. Altman concedes the first half — do what we can ourselves, government help is needed only for international coordination — while keeping the request for national rules. The bet is that a published, auditable safety-case practice deflects the "AI theater" charge better than the vague sign-on letters of the past week; the risk is that self-authored, self-published cases with no external verifier are exactly the costume critics already say the labs wear. We tracked the internal version of this shift Friday — Altman tells staff OpenAI would slow down — if rivals move first — and Amodei's auditor-heavy template it now answers — Amodei's pacing plan puts outside auditors inside Anthropic.
While American labs debate who should referee progress, CosmosMind — with Stanford, Berkeley, MIT, Tsinghua and Peking University — released MetaRSI-v1, a paper that formalizes self-improvement itself. The claim is that every recursive self-improvement system to date is one fixed program repeatedly editing the model; MetaRSI abstracts the loop into a shared "kernel" with three typed operators — generating verified training data, editing the agent scaffold without touching weights, and bounded training that internalizes capability — plus a meta-layer that rewrites the scheduling policy, recursion one level up. The target model (Qwen3.5-35B-A3B) runs every role in its own loop with no external teacher, evaluated on Terminal-Bench 2.1, SWE-bench Pro, GPQA-Diamond and AIME. Its "five laws" are quietly a safety document: self-improvement only amplifies capability a model already holds, and inside a closed loop, genuine gains and relaxed success criteria produce identical scores — undetectable from within. Related reading: Zhipu spoils GLM-6.0: a fully self-training model, leaked by a filing.
What to watch: whether OpenAI ever publishes the safety cases themselves, and which lab matches the mechanism rather than the sentiment — Microsoft's Satya Nadella today promised first-party MAI models with a "Code of Conduct," due, he says, tomorrow.
If the only auditor of a lab's pacing is the lab, is a published safety case evidence — or a receipt nobody verified? Tell us in the comments.
