The Race to Recursive Self-Improving AI

a16z’s Martin Casado talks with MakerMaker.ai co-founders Daval Ajita and Owen He about recursive self-improvement (RSI): when the system being improved also does the improving across models, harnesses, prompts, and even chips. They argue real RSI must be judged by capability gains relative to resources spent, and say their internal experiments show early signs of better performance and efficiency in inference and training.

Key points

• True RSI means the model is the optimizer, not just humans using AI to build AI.
• Progress should be measured as capability gains per unit of cost, time, and energy.
• Their team tracks experiments rigorously, including a “bleeder board” for costs.
• They say simpler, unified approaches scale better than many disconnected ideas.
• Inference optimization is a strong RSI target because it improves both speed and economics.