Every week we run a quick tally of what actually moved and what got pushed back on in AI — not the loudest headlines, the ones with a real before-and-after. This week had three genuine wins and one pointed piece of resistance worth sitting with.
The wins
OpenAI reported a second Millennium Prize-adjacent problem tackled using a swarm approach — roughly ten thousand coordinated agents working the problem in parallel rather than a single model reasoning alone. Whatever you think of the framing, coordinating that many agents on one hard problem without the process collapsing is itself a nontrivial systems achievement.
On the tooling side, Cursor's newer "Projects" feature reportedly pushed merged-PR rates up by around 30% for teams that adopted it — a meaningful jump for a workflow change rather than a model upgrade. And Cognition, the company behind the AI coding agent Devin, disclosed an annualized revenue run-rate approaching $900 million, with a claim that Devin now writes roughly 90% of its own code changes. That's a striking number for a two-year-old product category, even accounting for how loosely "writes its own code" gets defined in press materials.
The pushback
The counterweight this week came from academia. Roughly 770 mathematicians, many from Caltech, signed an open letter objecting to an AI-run mathematics competition, and OpenAI subsequently pulled its sponsorship from the event. The objection wasn't to AI doing math — it was about what an AI-judged, AI-sponsored competition does to the incentives and culture of a field that has historically valued human peer review and transparent methodology over benchmark chasing.
Why both stories matter together
Taken separately, these are just two AI news items. Taken together, they're a decent snapshot of where the field actually is: capability is climbing fast enough that a swarm of agents can meaningfully attack an unsolved problem, and the people closest to that specific problem — working mathematicians — are simultaneously the ones most uneasy about letting AI set the terms of how their field gets evaluated. Both things are true at once, and neither cancels the other out.