AI Economics

Concept

Recursive self-improvement

Recursive self-improvement is the process by which AI capabilities contribute to research that improves subsequent AI capabilities; it becomes self-sustaining only when progress accelerates without growing external inputs.

Explanation

Using AI in AI research is already a feedback effect, but it is not automatically an intelligence explosion. R&D automation, self-sustaining acceleration, and capabilities growing without bound are different claims and require increasingly strong conditions.

The loop can be represented as connected production relationships: better algorithms raise capabilities, capabilities help produce further algorithmic improvements, and economic output may finance more compute and data. The strength of a multi-step loop depends on the responsiveness of every link, so one weak elasticity can limit the entire cycle.

Human judgment, experimental compute, inference compute, data, and diminishing returns can break or weaken the feedback. Acceleration may also remain narrow: systems could improve at AI benchmarks or algorithm optimization without becoming equally capable at open-ended research or economically valuable work.

Examples

  • A reinforcing loop

    A coding agent improves a training system, the resulting model becomes a better coding agent, and the next training system improves faster.

  • A loop that stalls

    Agents quickly find easy optimizations but cannot judge novel research directions or run enough large experiments, so repeated use produces diminishing gains.

Sources