John Schulman of Thinking Machines, Beren Millidge of Zyphra, and Charlie O’Neill of Baseten debate on Dwarkesh Patel’s podcast whether recursive self-improvement in AI is imminent. They discuss technical bottlenecks such as sim-to-real gaps, sample efficiency, and whether current reinforcement-learning methods can discover genuinely new paradigms. The panelists disagree on whether progress depends primarily on data availability or architectural innovation, with one noting research automation is ‘nowhere near the ceiling’ while others suggest models may plateau before reaching human-level AI research capability.
