A lab publishes that its models direct 26% of its own research. Two labs allow outside evaluators inside their walls. A parliamentary chamber votes 417 to 3 to make data centres pay for their grid connections. After months of unverifiable commitments, three things that can be observed.
The figure of the week
Claude directs 26% of the work that builds Claude, against less than 1% in February. Thirty thousand agents run simultaneously, and the monitoring systems block roughly one action in 47,000.
What matters is not the level but the trajectory, and above all the fact that it has been published. We had been calling for months for denominators rather than anecdotes. Here they are.
These figures arrive six days after the essay asking the industry to slow down, which itself triggered within four days three responses from executives, a proposed ban, a presidential comment and a riposte from the Chinese state press.
What we were calling for has arrived
Independent evaluators are to work inside the labs, with employee-level access. This is exactly what was missing from every announcement this summer.
The test will be simple and we will follow it: what happens the first time an evaluator says no?
In parallel, the standards body of the three major labs is confirmed — and immediately labelled a cartel by a competitor. Both positions hold, and what will need watching is who sits on it and who can challenge it.
A vote of 417 to 3 requires data centres to pay for the grid reinforcements they cause, instead of passing them on to households. That score in an assembly divided on everything says something: the questions that get settled are the ones that hit a bill. Effective regulation of this industry will probably come less from texts targeting AI than from those targeting electricity, water and land.
Infrastructure and geography
A leading open model runs on 100,000 Chinese accelerators, at a cost per token its operator calls comparable. An open agent of 744 billion parameters beats a commercial model on web navigation. And 517 billion in compute commitments were disclosed the very week quotas were tightening.
On the European side, a company moved its headquarters to Germany to escape the Cloud Act. It is the most encouraging sign in recent months, and it comes from neither funding nor a supercomputer: a binding framework, when it is credible, eventually becomes a label.
And the rest
Apple handed Siri over to Gemini. Cloudflare blocks dual-purpose bots. Mathematicians declared their objectives misaligned with those of AI companies. A group is campaigning against model withdrawals. And Salesforce gave its agents first names.
On the thinking side: slowness as a method, the politicisation of a controversy, the question of rights and the race metaphor.
What this is shaping
A month ago we wrote that AI was becoming an industry like any other. This week adds the next step: an industry that is starting to be held to account.
Not out of virtue, and not completely. Unaudited figures, evaluators appointed by those they evaluate, a standards body whose governance remains to be defined. But for the first time, there are things a third party can observe rather than believe.
This is the condition that was missing, the one we stated in every article: a commitment no one can verify is not a commitment. There is now something to verify, and that is what the coming months will have to be measured against.