John Schulman on dead ends, scaling RL, and building research institutions

This podcast features an interview with John Schulman, where he answers questions about OpenAI's early days, including the possibility of creating ChatGPT earlier with full hindsight, the ragtag nature of early OpenAI, and unsuccessful projects like Universe. He also discusses ideal research management, compares research environments like early OpenAI and Thinking Machines, and shares his thoughts on value functions, continual learning, and brittle generalization in AI. Schulman explains his personal AI usage, describes his research process, and reflects on changes in research skills and idea generation. He also touches on academic publishing versus internal coordination in AI companies, the evolving talent pool in the field, the future of RL research, AI lab coordination, AGI timelines, and Thinking Machines' Tynker API and future plans.