At a glance
WHAT IT’S REALLY ABOUT
Defining “real” RSI: closed-loop AI improvement, measured not imagined
- The speakers define “real” recursive self-improvement (RSI) as a closed loop where the AI system is both the optimizer and the object being optimized, distinct from ordinary autocatalysis where humans remain the primary source of improvements.
- They model progress as capability improvement rate being driven by a performance numerator and a cost/difficulty denominator, arguing RSI can sustain improvement when the numerator grows with capability and could produce takeoff only under superlinear scaling.
- They caution that closed-loop systems do not automatically imply runaway growth because nonlinear dynamics can lead to plateaus, oscillations, and reward-hacked or noise-driven false positives, making RSI fundamentally an empirical measurement problem.
- Their implementation is a multi-agent system that includes humans as agents and aims to co-optimize prompts/harness, weights, and potentially deeper stack layers, because any unoptimized component becomes a bottleneck.
- They focus on inference optimization as a recursive and commercially valuable domain where improvements reduce the cost of future iteration, and they claim they are already seeing measurable “glimmers” of simultaneous numerator and denominator gains.
IDEAS WORTH REMEMBERING
5 ideasRSI requires swapping the optimizer and the optimizee—humans-with-tools doesn’t count.
They argue most “RSI” today is just humans using better tools (autocatalysis). Their criterion is whether the AI model/agent is itself the object being optimized and is also part of the optimizer that produces the next improvements.
The numerator/denominator framing clarifies why closed loops can plateau, grow linearly, or “take off.”
Dhaval frames progress as the rate of capability improvement being driven by performance (numerator) over cost/difficulty (denominator). If improvement capability scales with capability, you can get sustained improvement; if it scales superlinearly, you could see takeoff.
Closed-loop self-reference is not the same as takeoff; dynamics must be measured experimentally.
They emphasize that a feedback loop alone does not imply exponential growth; nonlinear dynamics can yield plateaus, oscillations, and degenerate behaviors. They treat takeoff as an empirical question because the key exponents/relationships must be measured, not assumed.
They pursue RSI as a multi-agent, human+AI co-optimization problem across the whole stack.
Their “system” includes multiple AI agents plus humans as agents, and they try to optimize any component that could become a bottleneck. Leaving parts unoptimized (e.g., only harness or only weights) eventually caps gains via bottlenecks (Amdahl’s law).
Scaling RSI favors simple loop-closing mechanisms anchored in a unified framework, not elaborate one-off tricks.
They claim many RSI papers don’t replicate or don’t scale because they’re complex, introspection-driven “stamp collections” rather than grounded, unified frameworks. Their approach is to start from a mathematical map of the space, reduce complexity into core abstractions, then test what scales.
WORDS WORTH SAVING
5 quotesWhereas when now, w- for the first time, we could swap the optimizer as the optimizee.
— Owen He
But if you have a closed loop, it-- you still have to pour in resources, but the, the marginal gains ideally get either constant or larger over time, and that's a, that's the fundamental difference.
— Dhaval Adjodah
Anything that you don't optimize will later become your bottleneck.
— Owen He
It's not a belief. It's, it's empirical measurement.
— Dhaval Adjodah
One of, one of the internal jokes we have in the company is that the, the secret to RSI is somewhere in a garden hose company in Wisconsin.
— Dhaval Adjodah
High quality AI-generated summary created from speaker-labeled transcript.
