Recursive Self-Improvement and the Future of AI

影片總結

MakerMaker.AI says its AI agents are beginning to improve inference performance while reducing the resources needed, but not necessarily triggering a runaway takeoff.

a16z General Partner Martin Casado speaks with MakerMaker.AI co-founders Dhaval Adjodah and Owen He about what qualifies as recursive self-improvement (RSI). They distinguish a true closed loop—where the AI system being improved also does the improving—from humans using AI as a tool, and describe a multi-agent approach that includes people. Casado frames progress as improving both capability and its cost; the founders say their system spans inference, harnesses and model weights, and that inference is a practical starting point. They stress that a closed loop does not guarantee exponential takeoff: results must be measured empirically, with human effort and compute counted, and many published RSI ideas fail to scale. The team tracks experiment costs on a “bleederboard” and uses a “Viceroy of Variance” to challenge results. They report seeing early signs that better results can come with less energy, while noting physical limits, noisy evaluations and slow real-world tests can remain bottlenecks.

章節

1. 0:00What actually counts as RSI: the model becomes both optimizer and optimized
2. 5:10Capability as a numerator and denominator: higher-power gains could yield exponential takeoff
3. 6:58Closed loop vs takeoff: nonlinear outcomes require measuring the exponent
4. 8:32Humans as agents in the system: a hybrid team tracks experiment costs
5. 13:39Why most RSI papers don't scale: unified frameworks link harness and weight updates
6. 17:41Why they chose inference: multi-LLM agents target serving efficiency
7. 24:00“It's not a belief, it's happening”: growing results and replication checks
8. 25:42The Viceroy of Variance: stress-testing results amid GPU noise and slow trials
9. 29:39The garden hose company in Wisconsin: overlooked workflows may unlock full-system RSI