Summary
Recursive self-improvement, or RSI, is the idea that AI systems can accelerate their own development, and frontier labs are starting to measure it. In this episode of a16z’s Working Theory, general partner Sarah Wang talks with Mark Chen, chief research officer at OpenAI, and John Schulman, co-founder and chief scientist of Thinking Machines and an OpenAI co-founder.
Both researchers explain how they define RSI: not models building new models entirely on their own, but AI used to speed up the research process, from pre-training and reinforcement learning to evaluation. Chen describes today’s systems as AI research interns, measured by how much compute is managed end to end by Codex to run experiments and automate parts of the research pipeline. Schulman walks through his own workflow, in which a model works on an idea for days and produces reasonable experiments, but human guidance still makes the difference in the final results.
The conversation also covers where researchers’ intuition comes from, and whether models can learn it through approaches like reinforcement learning from human feedback. It also covers how to verify open-ended research progress. Pre-training has a clear metric in perplexity, while downstream evaluations have increasingly short shelf lives and require constantly building harder benchmarks. The discussion closes on alignment implications and the most credible reasons for optimism about AI’s future.
📺 Source: a16z · Published November 17, 2026
🏷️ Format: Interview







