readonrisk

a16z (Andreessen Horowitz) · Oct 12, 2026 · 32 min

Recursive Self-Improvement and the Future of AI

A conversation with two founders building a recursive self-improvement (RSI) startup. They argue true RSI means the model itself is the optimizer, unlike human-driven autocatalysis, and that it must show gains in both performance and cost. They say they already see early empirical signs of this, mainly in inference optimization, while stressing noise, reward hacking, physical bottlenecks and eventual physical limits.

Watch on YouTube ↗More from a16z (Andreessen Horowitz)

Where they stand

AI

Very bullish

Recursive self-improvement is real and already showing early empirical gains in both performance and cost, though bounded by noise, physical limits and slow real-world domains.

Medium conviction · years

Key takeaways

  1. True RSI means the thing being improved is also the improver; human-using-AI-tools development is only autocatalysis. 2:44
  2. Simple model: capability growth is performance over cost. If performance scales with capability, improvement is constant; if it scales with a higher power, exponential takeoff follows. 5:28
  3. Whether RSI leads to takeoff is an empirical question, since there is no learning theory of deep learning and only empirical scaling laws. 7:39
  4. Claims of RSI must account for all resources, including compute, money and human experiment time, in the denominator; the company tracks experiment costs on an internal 'bleederboard'. 10:22
  5. Many published RSI papers fail to scale in their internal tests; simpler loop-closing approaches are the ones that scale. 13:42
  6. They optimize across the stack (harness, weights, eventually chips) and target inference efficiency, where both performance and cost improve. 17:04
  7. They claim to see early signs of better results with less energy and cost, but expect an eventual asymptote set by physical limits. 22:44
  8. Domains needing physical experiments, like drug discovery, will see slower RSI-driven acceleration than coding or math because of Amdahl's-law bottlenecks. 28:48

Calls and positions

  • Fields requiring physical experiments, like drug discovery, will be accelerated far less than coding or math because of evaluation bottlenecks.
    “those field probably won't be accelerated as fast as you know coding or mathematics”
    BearishPrediction: slower acceleration than coding or mathOct 12, 2026Watch at 28:48

Views expressed

  • Real RSI requires the model itself to be the optimizer, which makes human-with-AI-tools development merely autocatalysis.
    “that to us is just um autocatalysis”
    BullishOct 12, 2026Watch at 2:44
  • If performance scales with a higher power of capability, the result is exponential takeoff.
    “And then you get exponential takeoff.”
    Very bullishOct 12, 2026Watch at 6:33
  • Whether RSI takes off cannot be settled by speculation; it is an empirical question that requires hard testing.
    “it's almost not useful to speculate on what we think could happen because this is really an empirical problem”
    NeutralOct 12, 2026Watch at 7:39
  • Training GPT-2 on the same GPU went from weeks to hours at low cost through algorithmic improvements, many from auto-research.
    “a lot of them are auto research results”
    BullishOct 12, 2026Watch at 12:35
  • Many papers claiming RSI fail to hold up when tested at scale in their internal infrastructure.
    “a lot of the papers that claim to do RSI, when we test it in our internal infra, we find that they don't work”
    BearishOct 12, 2026Watch at 13:42
  • Inference optimization is an ideal RSI domain because it improves both performance and cost and is economically important now.
    “And one of those is the inference problem”
    Very bullishOct 12, 2026Watch at 18:11
  • It is already happening that the same energy or money yields better results over time.
    “it's not just possible. I think it's happening”
    Very bullishOct 12, 2026Watch at 22:44
  • Improvement will eventually asymptote because of physical limits on computation.
    “so it it will asmtote at some point”
    BearishOct 12, 2026Watch at 23:15
  • Human researchers will be hard for agents to replace, so the company treats humans as agents within a multi-agent system.
    NeutralOct 12, 2026Watch at 21:36#AI agents

Topics covered

More from a16z (Andreessen Horowitz)

Summary generated by AI from the video's captionsand checked against it; quotes are verbatim. Processed Oct 25, 2026.

Read on Risk summarises what other people said in public videos and podcasts. Summaries are generated by AI and checked by code against the transcript, but can still be wrong; follow the source links before relying on anything. This is not investment advice and nothing here is a recommendation to buy or sell. Disclaimer ·How this works