Cursor John Schulman on dead ends, scaling RL, and building research institutions
10 snips
Dec 17, 2025 
Join John Schulman, co-founder and leading researcher in reinforcement learning, as he delves into the early days of OpenAI and the potential for rapid advancements in AI like ChatGPT. He discusses the evolution of research management styles and the importance of effective team structures in fostering innovation. Schulman also tackles pressing issues like the future of reinforcement learning, the impact of continual learning, and how AI tools enhance daily researcher workflows. Plus, his insights on AGI timelines are sure to spark debate!
AI Snips
Chapters
Transcript
Episode notes
How ChatGPT Could've Been Built Earlier
- John Schulman says with full hindsight a few talented people could have built a ChatGPT-3.5-level system by 2018–2019 using networked GPU boxes.
- He cites NanoGPT as an example of extreme simplification where one person produced a working system in months.
Early OpenAI Was Ragtag And Exploratory
- Schulman recounts OpenAI's early ragtag phase where many small projects ran like academic labs alongside bigger engineering efforts.
- He gives Universe and early robotics as examples that were correct ideas but premature or unwieldy at the time.
Big ML Projects Need Two Layers
- Large ML projects combine environment integration and bespoke training systems that are often not fully decoupled.
- Successful efforts require both solid environment hooks and robust distributed training infrastructure.