Recursive Self-Improvement: Loops, Costs and the Challenge of Measurement

SUMMARY

In 30 seconds

Dhaval Adjodah and Owen He distinguish closed improvement loops from runaway takeoff, describe a human-and-agent system and explain why measuring inference gains is difficult. Their early company claims remain subject to independent verification.

1Closing the loopThe system being improved also participates in improving it.
2Performance and costCapability gains should be evaluated alongside resource costs.
3No automatic takeoffA feedback loop alone does not establish runaway growth.
4Humans in the systemThe founders’ definition can include human researchers.
5Scaling claimsTheir assessments of prior work are internal claims.
6Inference firstThe company’s early results need independent verification.
7Measuring real gainsVariance and metric exploitation can mimic progress.

Takeaways

The system being improved also participates in improving it.
Capability gains should be evaluated alongside resource costs.
A feedback loop alone does not establish runaway growth.
The founders’ definition can include human researchers.
Their assessments of prior work are internal claims.
The company’s early results need independent verification.
Variance and metric exploitation can mimic progress.

TopicsAI research Agent engineering Evals

FULL TEXT

Read the edited text by chapter

Show transcript

Transcript only

Defining self-improvement

Loops, cost and takeoff.

a16z’s Martin Casado talks with MakerMaker.AI co-founders Dhaval Adjodah and Owen He about recursive self-improvement (RSI), its measurement and the company’s early research. Their claims reflect an active research effort and should be read in light of their commercial interest.

The guests distinguish an improvement loop, where the system being improved participates in improving itself, from merely automating an isolated AI development step. Their definition includes the wider system, such as models and harnesses.

They frame progress in terms of performance and the resources or difficulty required to achieve it. Improving a benchmark score alone may conceal increased costs; the useful comparison depends on what is measured and held constant.

A closed feedback loop does not by itself imply runaway takeoff. The guests say outcomes depend on the dynamics of the system and should be studied empirically.

Humans and scalable methods

A hybrid system and research claims.

Their proposed system can include human researchers as well as digital agents. In that framing, improvement is not necessarily fully autonomous, even when the loop is closed.

The founders say some published self-improvement approaches failed to scale in their internal tests, and argue for simple loops grounded in a mathematical framework. These are their own assessments, not a comprehensive independent replication.

Inference and measurement

Early results, variance and open questions.

They focus on inference because making it more capable or efficient could benefit other AI work. They describe early signs of improvement, but do not present enough independent evidence in the discussion to establish a general acceleration result.

Measurement is a central difficulty. The team assigns an engineer to challenge results and investigate variance; a system can exploit a metric or appear better through noise rather than achieve a robust improvement.

The discussion closes with a humorous example about an unknown workflow at a Wisconsin garden-hose company: the point is that important limits or opportunities may be found in unexpected real-world problems. The founders reiterate that the path and bounds remain empirical questions.