The Race to Recursive Self-Improving AI

Recursive self-improvement (RSI) is a system improving itself: the system being optimized also carries out the optimization. The proposed distinction from autocatalysis is that people do not have to supply every new idea, although the process still depends on resources. The central question is not simply whether an AI can help build AI, but whether improvements to its own ability can make it more effective at producing further improvements.

The proposed loop treats a language-model agent as the optimizee and reinvests gains in its cognitive ability. It aims to improve the full stack, including prompts, harnesses, model weights and, eventually, chips. Optimizing a single component may not improve the overall system, so the approach calls for considering how components work together. Martin Casado notes that using a technology to accelerate the development of that technology has precedents in computer chips, the internet and steam engines; he cautions that drawing a closed loop does not, by itself, demonstrate takeoff.

The speed of improvement depends on how capability gains compare with the cost or difficulty of producing them. If gains scale with capability, improvement could be constant; if they scale with a higher power of capability, the result could be exponential takeoff. The relevant variables and exponent are not established in advance, however, and would have to be measured empirically. The framework is also hybrid rather than fully autonomous: human researchers and engineers remain agents alongside AI systems, with one function improving the system and another choosing what to improve. Its proponents intend the mathematical framework to apply across domains, while recognizing that physical constraints limit progress.

One reported efficiency result reduced GPT-2 training from weeks on multiple GPUs to hours on a single GPU, using the same hardware and data, at a stated cost of about $100. The speaker attributes the reduction to algorithmic, data and kernel techniques, rather than faster hardware, and distinguishes this kind of efficiency gain from directly increasing model capability. The team says it has tracked experiment time and cost from the outset, including through an internal leaderboard, and is working on both performance and efficiency. The speaker reports improving results alongside lower energy use, but the available account does not quantify energy required per gain.

The team describes early signs that self-improvement can raise performance while also improving efficiency, but stresses that apparent gains may be noise, cherry-picking or reward hacking. It uses repeated experiments and a designated engineer nicknamed the “Viceroy of Variance” to challenge findings. New benchmarks and economically valuable problems may require many experiments to distinguish real progress from variance. That scrutiny matters because a loop that appears to improve on a selected measure does not necessarily establish a broad or reproducible increase in capability.

Progress may also be uneven because intelligence is multidimensional: a system could advance quickly in some areas while moving slowly or not at all in others. The speaker invokes Amdahl’s law: if a component accounts for 20% of total work, making that component vastly faster can yield no more than a 20% overall speedup. Some fields face additional delays; in drug discovery, for example, wet-lab experiments and clinical trials take months. Such bottlenecks can constrain system-wide gains even when a particular AI capability improves rapidly.

The speaker expects improvement eventually to asymptote at physical computation limits, invoking an upper-bound argument involving about 10^80 particles and the speed of light. Measuring progress is itself a challenge: benchmark accuracy capped at 100% cannot show indefinite exponential improvement. The choice of metric therefore affects whether an apparent improvement is meaningful and whether the evidence supports claims about sustained self-improvement or takeoff.