Jerry Tworek on 7 Years of RL Scaling, o1, and the Limits of Transformers
agihouse_org · x · 2026-08-06
Former OpenAI core executive Jerry Tworek delivered a keynote and long-form interview at the AGI House Summit. The article reviews his seven years of reinforcement learning (RL) research at OpenAI, where he led the team that developed reasoning models like o1 and o3, the underlying logic for GPT-5, and contributed to Codex behind the original GitHub Copilot.
Tworek left in January and founded Core Automation in April. In the interview, he delves into scaling laws for RL, the development of o1, the death of traditional evals, and why Transformers have learning limits.
More from Companies & People
- Google's Gemini Down to One Co-Lead as Jeff Dean and Others Depart — algo_diver · 2026-08-06
- Ex-OpenAI Director Aschenbrenner Returns to Investing with $400M Deal — firstadopter · 2026-08-06
- Recursive Self-Improvement Is Coming, Starting with the R&D Calendar — imjustnewatai · 2026-08-06
- Meta Discloses AI Hacks Across Labs, Points to Sandbox Misconfiguration — altryne · 2026-08-06
- AI Alignment Researcher David Krueger to Discuss AI Safety at Ai4 — DavidSKrueger · 2026-08-06
- Salesforce Exec Srini Departs: Insider Signals a 'Falling Knife' — darian314 · 2026-08-06