Daily Brief
2026-09-09 · Morning · 12 items
-
OpenAI fully rolls out Astra to Plus, Pro, Business, and Enterprise users
Astra is now available to four paid tiers, with a live demo released by OpenAI.X：OpenAI (@OpenAI)
-
OpenAI agent team announces solution to Navier-Stokes Millennium Prize Problem
The problem has been open for about 90 years; a group of agents using a next-generation model produced the solution.X：OpenAI (@OpenAI)
-
Mistral completes €3 billion Series D round, valuation exceeds €21 billion
Mistral says the round is the largest-ever equity financing for a European tech company, and it was founded just 3 years ago.Mistral AI：News（网页）
-
Runway releases free plugin for Premiere Pro/After Effects
The new panel embeds in the timeline to generate videos/images and can rerender footage at its original length via Aleph 2; free to download.Runway：News（网页）
-
Tom Tunguz breaks down OpenAI’s 3x productivity gain: each researcher runs 4 agents
Internal OpenAI data ties one 8-hour researcher shift to 3.14 agent workdays, with 4 agents usually running in parallel.Tomer Tunguz 博客（VC 分析）
-
Sam Altman responds to Navier-Stokes dispute, says he was threatened with plagiarism claims
Altman says the other side had only the Euler result, coordination failed, and they threatened a plagiarism claim.X：Sam Altman (@sama)
-
Noam Brown says OpenAI’s agent effort on Navier-Stokes cost millions
OpenAI’s proof cost millions, but Noam Brown expects such costs to fall quickly.X：Noam Brown (@polynoamial)
-
NYU mathematician accuses OpenAI of unfair competition on Millennium Prize problem
They published three preliminary results and said OpenAI, after a leak, raced along their route with massive compute.TechCrunch：AI（RSS）
-
Anthropic explains cutting costs and boosting performance with Claude Platform
Key point: better cache hits and prompt cleanup can cut costs without hurting quality.X：Claude Devs (@ClaudeDevs)
-
Dwarkesh Patel study: pretraining gains mostly from data improvements
Key finding: data upgrades gave 3.24x the compute-efficiency boost of model changes.Dwarkesh Patel：Podcast & Blog（RSS）
-
Berkeley RDI launches CUA-Lite, an open platform for computer-use agents
Key point: one framework covers 15+ benchmarks, 30k+ tasks, and 10+ datasets.Berkeley RDI：Blog（AI 安全与评测）
-
GPT-6 Astra + Blender test: Computer Use burns half Pro quota in 4h
Building a Temple of Heaven model with Computer Use took four hours and burned nearly half of the $200 Pro membership.公众号：数字生命卡兹克