a16z podcast: recursive self-improvement is happening, but closed loops don't guarantee takeoff
a16z · youtube · 2026-10-12
a16z General Partner Martin Casado talks with MakerMaker.AI co-founders Dhaval Adjodah and Owen He about recursive self-improvement (RSI).
Key points:
- True RSI ≠ using AI to automate parts of AI development; it means a closed loop where the improved system does the improving, across models, harnesses, and more of the stack
- A closed loop doesn't guarantee runaway takeoff; measuring progress requires treating capability as both numerator and denominator
- Most RSI papers don't scale; they chose to start from inference
- Early research suggests self-improving systems can raise performance while improving efficiency
- Proving real gains is hard: variance, reward hacking, distinguishing signal from noise
Related event: a16z Podcast Discusses Recursive Self-Improvement in AI(2 posts)→
More from AGI Musings
- Breiman's 2001 'Two Cultures' paper resurfaces as AI makes black boxes the norm — akbirthko · 2026-10-12
- AI filings are flooding courts — researcher proposes a parallel AI adjudication track — latticecut · 2026-10-12
- jd_pressman to Richard Ngo: MacAskill's opinion on AI doesn't matter, would SBF have skipped Anthropic? — jd_pressman · 2026-10-12
- Perplexity CEO: $750M annualized revenue, AI-native SMBs fueling new millionaire boom — AravSrinivas · 2026-10-12
- Richard Ngo: Elites ignored AI not from stupidity but because it seemed weird and cringe — RichardMCNgo · 2026-10-12
- Why the messy LLM training pipeline may still lead to superintelligence — akbirthko · 2026-10-12