a16z podcast: how to tell real recursive self-improvement from AI-building-AI hype
a16z · x · 2026-10-12
a16z's Martin Casado talks with Erratic Intelligence co-founders Dhaval Adjodah and Owen He on how to tell when recursive self-improvement (RSI) is real.
Key points:
- Most RSI pitches are just people using AI to build better AI — tech has always worked that way (you needed computers to build chips; coal made steam engines better). Their bar is higher.
- True RSI requires the model to improve itself: its weights, its harness, even the prompt telling it what to work on next (optimizer vs. optimizee).
- Gains must hold up net of costs: the team tracks everyone's experiment spend on a "bleeder board," and one engineer, the "Viceroy of Variance," has the job of breaking their results.
- Over recent months they've seen early signs of their system getting better and cheaper at the same time.
Related event: a16z Podcast Discusses Recursive Self-Improvement in AI(2 posts)→
More from Research
- Sakana's self-improving MASS model hits ~70% win rate but can't see its own blind spots — The Decoder · 2026-10-13
- Four ETH Zurich lectures map embodied AI's gaps: hands, space, world models, data — le_james94 · 2026-10-13
- Harness code becomes a research battleground: CMU rewrites it with RL, Apple says it doesn't matter — burkov · 2026-10-13
- LiveBench 2.0 incoming: agentic long-horizon benchmark you can't benchmaxx — bindureddy · 2026-10-13
- Muse Wonder shows far better OOD generalization than Muse Spark 1.3 — armand_ruiz · 2026-10-13
- Tim Hwang: self-driving research fields in silico are on the edge of reality — CatAstro_Piyush · 2026-10-13