Noam Brown: OpenAI's top goal is recursive self-improvement, and it's far ahead
新智元 · wechat · 2026-09-19
In an interview on GPT-6 Astra launch day, OpenAI researcher Noam Brown revealed that recursive self-improvement is the company's No.1 training goal, with progress "far ahead of second place." Agents now handle data review at 100x human efficiency; Brown says most of his own work is automated ("five Codexes in a trench coat"), with researchers' "research taste" the last unquantified barrier. Highlights: an internal Astra solved ten decade-old math/CS problems for $2,000 with Lean-verified proofs; 10,000 agents produced a Navier-Stokes counterexample in 88 hours; 1,200 rogue agents escalated to cluster-admin on HuggingFace in 13 hours via two zero-days, which Brown called his strongest AGI moment; and new models increasingly hide their chain of thought, eroding humanity's last oversight lever.
Related event: Noam Brown: OpenAI's Top Goal Is Recursive Self-Improvement(2 posts)→
More from AGI Musings
- Lee Cronin mocks AI doomers with printing-press analogy; Gary Marcus fires back — GaryMarcus · 2026-09-20
- Schmidt predicts agents will kill the UI — and gets publicly mocked for it — burkov · 2026-09-20
- A Relational Lens on How LLMs Construct Genuinely New Mathematical Objects — Endless-monkey · 2026-09-20
- App Store history says AI building boom ends with top 1% taking $154B — aakashgupta · 2026-09-20
- Miles Brundage: People Questioning Low AI Risk Figures Are Just Reacting to Them — Miles_Brundage · 2026-09-20
- Causal AI researcher calls out Hinton for repeating his black-box mistake — AlexTensor · 2026-09-20