Last week’s debate had a catch, and the catch was antitrust law. If labs agree to go slower, in any other industry that is collusion. OpenAI asked Congress about it, and David Sacks then ruled out an antitrust exemption, which left the second stage of Amodei’s plan hanging in mid-air.
Sam Altman has now answered in writing. The short version: we’re not waiting.
What he wrote
«We welcome a federal framework that sets consistent safety requirements for frontier AI. But we do not believe we need to wait for an anti-trust exemption or legislation to begin the work of providing this confidence.» Independent auditors get a direct mention as a good idea.
And one line that is sharper than anything OpenAI has said on the subject so far: «Every frontier lab must deliver on this, and there is no reason any of us should come to work if we cannot.»
The concrete part
The mechanics are more interesting than the stance. Altman concedes that Responsible Scaling Policies and Preparedness Frameworks are aimed at finished models, at the moment of release. What happens during development isn’t covered.
So OpenAI says it now formulates an explicit «safety case» ahead of frontier reinforcement learning runs it expects to raise capability significantly. That sits on top of the work that already happens before a release. He wants other companies to copy the approach or propose their own, and argues that shared standards for misalignment, monitoring and safety produce better outcomes.
Plus a clarification on the word itself: «When we talk about ‘pacing’, we do not mean ‘stopping’.» Progress stays fast, just slower than it could be. Safety cases and monitoring cost something, and he thinks the cost is worth it. His words: no amount of American competitive pressure justifies recklessness.
Where he does want government is international coordination. Everything else, labs can start on their own.
Washington’s objection just lost its target
Sacks delivered the sharpest line of the debate on Saturday: the easiest way not to build superintelligence is to agree not to build it, and nobody needs permission for that. He meant it as a refusal. Altman has taken it as a plan.
That settles the question of whether Washington writes an antitrust exemption. It leaves a less comfortable one: whether the labs actually do this when nothing forces them to. A safety case before a training run is checkable, but only from the inside. Altman writes himself that independent auditors would be a good idea. He does not write that OpenAI has any — unlike Anthropic, which announced exactly that as stage one.
Which leaves the sentence he can be measured against in six months: that there is no reason for any of them to come to work if they can’t pull this off.
Sources: