Recursive Self-Improvement and the Future of AI
a16z · 2026-10-12 · 32 min
https://www.youtube.com/watch?v=3Q1mdn5ZGvMVideo summary
MakerMaker.AI’s Dhaval Adjodah and Owen He say their inference-focused RSI is showing gains in capability and efficiency.
On a16z General Partner Martin Casado’s interview, MakerMaker.AI co-founders Dhaval Adjodah and Owen He distinguish recursive self-improvement (RSI) from using AI to help humans build AI: in true RSI, the system being improved also drives the improvement. They describe a closed-loop, multi-agent approach spanning prompts, harnesses, model weights and, eventually, more of the stack, with human researchers still part of the system. They stress that a feedback loop alone does not guarantee runaway takeoff; outcomes depend on measured performance, costs and empirical behavior. Their startup focuses on inference because improving serving efficiency can both lower costs and free resources for more capability. They say many RSI papers fail to scale, track experiment costs on an internal “bleederboard,” and rigorously test results for noise and reward hacking. The team reports growing evidence that better results can come with less energy, while acknowledging physical limits and slower fields such as drug discovery, where wet-lab tests and clinical trials constrain iteration.
Chapters
- 0:00What actually counts as RSI: the model must become its own optimizer
- 5:10Capability as a numerator and denominator: fixed performance gains lead to a plateau
- 6:58Closed loop vs takeoff: the exponent must be measured empirically
- 8:32Humans as agents in the system: tracking experiment costs while co-optimizing the AI stack
- 13:39Why most RSI papers don't scale: simpler closed loops and unified system design
- 17:41Why they chose inference: optimizing model serving while preserving human expertise
- 24:00It's not a belief, it's happening: early signals require careful metrics and replication
- 25:42The Viceroy of Variance: stress-testing results amid noisy, domain-specific bottlenecks
- 29:39The Wisconsin Garden Hose Company: Overlooked Workflows Could Unlock RSI
What the full analysis includes
9 chapters · 40 key points · 9 section summaries · mind map
Preview of the first chapter · What actually counts as RSI: the model must become its own optimizer
The remaining 37 key points, section summaries and mind map are in the full analysis. Analyze this video to see everything →
This video is 33 min. The Free plan analyzes videos up to 30 min; plans from Plus cover videos up to 2 hours.
This is a Tier 1 public summary
Whether the chapter key points, section summaries and mind map are public is up to the person who shared it. Want the full analysis?Submit one yourself.
More from this channel
Hip Hop’s Origin Story, Told by the Legends Who Lived Ita16zKurtis Blow recalls building hip-hop from nothing in 1970s New York and mentoring Run before Run-DMC rose to fame.
What iPhone Packaging Reveals About the Robotics Opportunitya16zTutor Intelligence CEO Josh Gruenstein says its 100-robot Boston fleet trains AI while robots tackle factory work for about $16 an hour.
Why AI Could Turn Small Law Firms Into Billion-Dollar Businessesa16zEve CEO Jay Madheswaran says AI helped an LA law firm grow from 100 to 500 clients per attorney by automating plaintiff-law work.
Why Founders Are Betting Big on Factories, Mines, and Energya16za16z investors argue defense demand can rebuild U.S. factories, mines, and energy while creating commercial markets for robotics.
Mark Cuban on Why AI Could Turn Solo Founders Into Trillionairesa16zMark Cuban says AI could expose healthcare contract abuses, while Cost Plus Drugs bets $20 million manufacturing pods can lower drug costs.
Related analyses
Why Would AI Companies Want to Slow Down?a16zDatabricks CEO Ali Ghodsi says enterprise AI lacks organizational context, while cyberattack exploits now move from disclosure to weaponization in hours.
Anthropic S-1 Shocker, Greatest Business Ever? Google’s $500B Wipeout, Midterm PredictionsAll-In PodcastAll-In debates Anthropic’s $100 billion IPO, Google’s $500 billion selloff, AI infrastructure risks, and sharply divided midterm forecasts.
To Regulate AI Effectively, Focus on How It’s Useda16z Deep DivesMartin Casado argues AI laws should target harmful uses, warning regulatory uncertainty is pushing U.S. startups toward Chinese open-source models.
Datadog CISO on Securing AI Agents at Scale | Deep Dives with a16za16z Deep DivesDatadog CISO Emilio Escobar explains how 4,000 engineers use AI agents, while role-based MCP servers and an LLM judge help catch malicious code.
Inside the Race to Measure Frontier Intelligencea16zVals CEO Rayan Krishnan argues independent evaluations caught Llama 4 underperforming and help enterprises measure AI’s real-world value.