Top story
10,000 agents on Navier–Stokes: the breakthrough is the system, not a magic prompt
https://openai.com/index/navier-stokes-solution/
OpenAI says its internal system produced a proof that smooth three-dimensional Navier–Stokes flow can develop a finite-time singularity under smooth forcing while retaining finite energy, and it released a Lean formalization alongside the paper. The operational numbers change the meaning of the claim: roughly 10,000 agents, 130 billion output tokens, 88 hours to find the result, then another 17 hours for formal verification. On HN, minimaxir points at the missing line item—the likely cost at ordinary API prices, before accounting for a model more capable than GPT-6 Astra. I see strong evidence for research acceleration, not evidence that hard mathematics has become a button. A lab announcement still needs independent mathematical review, especially while the 512-comment thread disputes concurrent work and training-data provenance. For developers, the practical shift is that orchestration and verification now belong inside the capability boundary. Report total compute, provenance, and the verifier path; a model name alone no longer describes the system that produced the answer.
A one-petabyte genome lookup layer
AlphaGenome Atlas turns a model into queryable infrastructure. Google says it precomputed effects for nine billion single-nucleotide variants into a 1 PB dataset and AVI score; one study found 22% more non-coding associations. jFmDRz73 identifies the right boundary: this is a prediction cache, not a diagnosis. I would use it to rank hypotheses, then demand biological validation.
Images 2.5 should win on editing, not demos
The API release targets iterative production work. Images 2.5 adds Flare and Sunburst; Flare claims 50% lower latency, while both emphasize subject preservation and multi-turn edits. pelzatessa still spotted broken teeth and fingers in a showcase image. I would benchmark consistency on real assets before moving a pipeline.
— Tin