Dwarkesh Patel has released a new approximately 96-minute discussion featuring John Schulman, Chief Scientist at Thinking Machines and co-founder of OpenAI, Beren Millidge, CTO of Zyphra, and Charlie O’Neill, Head of Model Training at Baseten. The conversation ranges from "how to find the strongest counterarguments for recursive self-improvement" to timelines.

The discussion opens with a counterpoint question: if there is no superintelligence that significantly transforms the world by 2036, what are the most likely technological reasons, excluding war, regulation, and political shocks? The three repeatedly distinguish between two types of tasks—cumulative tasks and non-stationary tasks. Discoveries like attention, mixture of experts, and GRPO can be directly inherited by later models once documented; however, tasks like law firm assistant work require constant adaptation to changing relationships, implicit processes, and local knowledge. O’Neill presents a stark conclusion: recursive self-improvement happens to be easier than being a law firm assistant.

Progress in Chinese laboratories is discussed separately. Schulman points out that routing and proxy services aimed at Chinese users will collect and sell traffic to access cutting-edge models in the U.S. (primarily programming-focused), and this prompt distribution serves as a ready-made textbook for distillation. Millidge states that leading laboratories purchase data from large data companies, and Chinese laboratories can do the same; if data overlaps, distillation can bridge the generational gap. He also mentions that the Chinese side will use online interactions for distillation, with the next generation of models directly consuming the deployment trajectories of the previous generation.

How automation researchers train and whether long-term reinforcement learning can lead to general intelligence are placed on the same chain. Millidge believes that a complete closed loop—where the model decides its own data and training—is what counts as recursive self-improvement, with the key variable being how much human information is still needed. The environment can challenge any human, but that does not mean trajectory distillation can surpass the teacher's upper limit. Schulman emphasizes that much research is not about climbing a clearly defined goal but turning the intuition of "where the model should improve" into verifiable settings. O’Neill points to the bottleneck of how far Mercor and internal laboratory teams can expand the complexity of reinforcement learning environments.

Reinforcement learning is reinterpreted. Patel mentions that a year ago, academia was generally pessimistic about the scalability of reinforcement learning; Schulman himself wrote in a paper that models learn about one bit per round. Millidge breaks down recent successes into underappreciated mid-term training: many so-called reinforcement learning victories actually come from synthetic reasoning data and warm-up environments, equivalent to a round of pre-training before entering reinforcement learning. The gap between simulation and reality, catastrophic forgetting during continuous learning, and loss of plasticity are cited as reasons why long-term judgment and goal setting still require human input. In a rapid Q&A, treating models as replaceable remote white-collar workers or timelines that could increase AI research efficiency by an order of magnitude is estimated to take about 2 to 10 years; Schulman gives a judgment of 3 to 4 years for stronger intelligence.

In terms of market mechanisms, this is a narrative trade rather than an immediate order. The buy side bets on recursive self-improvement, willing to channel computing power from inference back to training in leading laboratories and their cloud vendors; the sell side still relies on human researchers, data procurement, and the expansion of evaluation environments in open-source and mid-sized training stacks. Funding flows into two branches: one continues to purchase prompt distributions, synthetic reasoning data, and reinforcement learning environments; the other flows towards talent brands on the "open-source side" that can speak publicly and reduce information asymmetry. Beneficiaries include companies like Thinking Machines, Zyphra, and Baseten, which can discuss roadmaps on camera, as well as latecomer laboratories catching up through distillation; those under pressure are closed-source giants that treat recursive self-improvement as a scheduled narrative and application layer buyers unable to write non-stationary real work into scalable rewards. The event-driven nature comes from the discussion itself: it transforms the question of "Is RSI around the corner?" from a slogan into a falsifiable task classification.

The program is titled "AI researchers debate how close we are to recursive self-improvement," with timestamps covering counterarguments, Chinese laboratories, automation researcher training, long-term reinforcement learning, simulation to reality, data contribution ratios, why reinforcement learning works, AlphaGo's 37th move and entropy collapse, and rapid timelines. The full video can be watched on video platforms, with a transcript released simultaneously.

Source: Public Information