OpenAI's RSI report: pretraining shows earliest AI acceleration, 'research taste' is the bottleneck
ronbodkin · x · 2026-11-19
- Analysis of OpenAI's new research-acceleration report: AI's meaningful contribution first shows up in pretraining
- Method: 1-4 key contributors per area (pretraining, posttraining, safety monitoring) named the most important ideas behind GPT-6 Astra and GPT-6.2 Astra and explained AI's role in them
- Finding: across all three areas, researchers attributed the ideas entirely to humans; models helped implement them
- Key argument: developing and vetting conceptual research ideas ('research taste') may be the bottleneck for recursive self-improvement
- Links to the author's article on AI takeoff pacing, arguing progress on hard-to-supervise tasks may lag easily supervised ones
Related event: OpenAI report reveals exponential rise in internal AI research usage(25 posts)→
More from AGI Musings
- Open-weight models now handle 81.2% of OpenRouter tokens — the price of generic intelligence is open source — markjeffrey · 2026-11-19
- a16z Declares the B2A Economy: AI Agents Are Now the Internet's Main Users and Buyers — FinanceYF5 · 2026-11-19
- Bridgewater: Managing Money on a One-in-Three Odds of a Economically Destabilizing Cyber Attack — GaryMarcus · 2026-11-19
- Notion's AI Bet Isn't a Better Assistant—It's Becoming the Context Layer for Agents — ivanhzhao · 2026-11-19
- OpenAI ships three safety/research posts in a day, reigniting Anthropic transparency debate — daveholtz · 2026-11-19
- 'Moloch made me do it' is no excuse: defecting on humans is a choice — mimi10v3 · 2026-11-19