A company claims to have solved a millennium problem. Two researchers dispute the context. Agencies accuse foreign labs of siphoning off models. A music platform announces licensed models while facing more lawsuits than before. Four claims, four times the same question: what can someone else verify?
The announcement that overshadows everything
OpenAI claims an internal system has solved the Navier-Stokes problem, with around 10,000 agents coordinated over 88 hours and a proof formalised in Lean.
Nothing has been validated. The prize rules require publication and two years of acceptance by the community, as we explained in our article on these seven problems. The question that will decide it is short: does the formalised statement match the problem as posed?
And the announcement comes with a challenge. The two researchers who had published a neighbouring result on the Euler equations are raising questions about the context. We are not adjudicating, and we devoted an article to the general question: a research question belongs to no one; what needs discussing is the standards of conduct in a competition that has become highly asymmetric.
The verification thread
This is what ties the week together. Formal verification offers a concrete answer: you don't trust the model, you trust a verifier whose kernel fits in a few hundred lines.
By contrast, a text detector reported to have been bypassed is a reminder that a statistical signal always degrades more easily than it holds up. And model extraction shows that a system that answers well exposes its know-how.
Mistral raises €3 billion, the largest European tech funding round. China is targeting 9,800 exaflops by 2030 and preparing a cluster of 100,000 domestic accelerators. Qualcomm and Amazon are partnering on inference. The orders of magnitude remain incomparable, which led us to ask whether you have to be the best to count.
The environmental story
It may be the most important article of the week, and the least commented on. Data centres are leaving the electricity grid: a third are said to be fully self-sufficient by 2030.
They no longer draw electricity, they generate it — which sidesteps grid connection delays and local opposition, and removes a massive consumption from collective measurement mechanisms. Google locking down nuclear power in Finland is the best possible version of this, and it raises the same question: who decides how these facilities are powered?
And the rest
Suno launches licensed models and finds itself in more litigation than a year ago. ChatGPT Images 2.5 moves towards editing at an unchanged price, which led us to explain how an image is counted in tokens. A thousand engineers deployed at client sites to install agents. And an investigation into monetised illegal ads shows what automated review calibrated for volume produces.
On the thinking side: the dual use of security research, what deserves to be automated, who is the author of a discovery, and should you believe an extraordinary claim.
What it adds up to
This week produced more claims than established facts, and that is precisely what makes it interesting.
Faced with a stream of announcements that exceeds any capacity to verify, the useful stance is neither to believe nor to reject. It is to know, for each assertion, what would make it credible — and most often the answer is the same: someone else tried to break it and failed.
That is the conclusion of our article last night, and it sums up the week. We will know in a few months whether a millennium problem has been solved. Until then, saying we don't know is the only position that will remain defensible whatever the outcome.