A few days ago, Dario Amodei published an essay calling on all leading labs to slow down. Within four days, the heads of Microsoft, OpenAI and xAI had each responded publicly, a US senator proposed simply banning superintelligence outright, the US president commented on social media, and Chinese state media attacked the author. This is the first time the question has been debated publicly among those who can actually decide it.
Yesterday we reported that Microsoft was joining this position, noting that a convergence between competitors signals a norm in the making. The full sequence is more interesting still.
What makes this moment unusual
Debates about AI safety have until now taken place in three separate spaces: research papers, carefully calibrated corporate communications, and parliamentary hearings where nobody says much.
Here, four leaders who effectively control the pace of development are answering each other publicly, within a few days, on a subject they had every interest in handling in private.
That is not trivial. We wrote about the slowdown announced in August that individual caution is punished as long as it is not collective. A public debate is precisely what makes it possible to break out of that deadlock: it makes commitments comparable and costly to betray.
The substance of the disagreement
The positions expressed are not identical, and the nuance matters.
Some argue for a deliberate pace — taking the time needed for alignment to keep up with capabilities. That is what the code of conduct published by Microsoft covers, with principles of human control, of safety taking priority over task completion, and of no autonomous goals.
Others dispute that slowing down is possible or desirable, putting forward the argument we have often reported: a Western slowdown does not slow down the world. This week we documented a Chinese cluster of 100,000 domestic accelerators and a five-year plan worth 532 billion. That argument is not a posture.
When a president comments on a social network and a foreign state press attacks a company leader, the discussion has changed in nature. It is no longer a controversy between specialists about capability thresholds: it has become a political object, with all that implies in terms of simplification, partisan alignment and commentary. That shift was inevitable, and it has a cost: technical arguments become positions to be defended rather than hypotheses to be examined.
The proposed ban
The most concrete element of the sequence is a legislative proposal aimed at banning superintelligence.
It immediately raises the problem we encountered with the word AGI: there is no shared definition of what would be banned. A law that targets an object without a legal definition produces either complete unenforceability or considerable discretionary power for whoever interprets it.
That takes nothing away from the intention. It simply signals that turning a legitimate concern into applicable text remains entirely to be done — and that is exactly what we have been calling for for months: capability thresholds that are defined, measurable and public.
What is still missing
This public debate is good news, and it changes nothing about what we keep repeating.
A commitment to slow down is only worth anything if it is verifiable by someone else. Nobody outside can currently establish that a lab has actually slowed down, or measure by how much.
And incident reporting is only worth anything if it is protected. We wrote about the inventory requested by Congress: if reporting exposes you, nobody reports.
One item reported this week illustrates that limit well: part of the cybersecurity community believes it was kept out of the preparation of the alerts that these same labs published. Discussing the pace among leaders is not the same thing as building a system with those who will have to apply it.
What to take away
It is worth welcoming that this question is finally being discussed openly, and worth being wary of what the public format does to the arguments.
A debate conducted between four people who run the labs concerned, commented on by politicians and caught up in rivalry between states, produces attention and rarely criteria.
What would make it possible to settle the matter has not changed since we started writing it: capability thresholds defined publicly, evaluations by third parties with resources, and a reporting obligation that protects the person reporting. None of those three elements has advanced this week, and that is the measure by which the real significance of this episode will have to be judged.