Summary
Recursive self-improvement, or RSI, is one of the most discussed and most loosely used terms in AI. In this a16z conversation, two founders of a startup working on the problem explain what they think it actually means and why it is hard.
The guests, whose backgrounds include a physics and reinforcement learning PhD from the MIT Media Lab and continual learning and meta-learning research at DeepMind, draw a line between true RSI and what they call autocatalysis, where humans use better AI tools to build better AI. In their framing, the key question is who the optimizer is and what is being optimized. RSI means the model itself becomes the thing being improved, so gains are reinvested into the system’s own cognition. They discuss improvement at several levels of the stack, from prompts and harnesses to weights and even chips, and describe a multi-agent approach that has itself been recursively improved.
The discussion also covers why many published RSI papers fail to hold up when tested at scale on real benchmarks and economic problems, why simpler loops that close tend to win, and the value of grounding the work in a mathematical framework. It also touches on the compute economics of improving training efficiency and the interest in reducing energy use over time.
📺 Source: a16z · Published October 12, 2026
🏷️ Format: Interview
