The Economics of Recursive Self-Improvement

AI helping researchers build better AI does not, by itself, establish an accelerating feedback loop. In a note introducing their economics paper, researchers at METR distinguish the existence of feedback from feedback strong enough to sustain acceleration. They single out the link from better model capabilities to faster algorithmic progress as the most uncertain.

The authors identify computing, data and experimentation as possible bottlenecks that could limit the loop, while saying the available evidence does not rule out substantial acceleration. Their parameter estimates are preliminary. The useful question is therefore how strong each link is, rather than whether an activity qualifies for the contested label of recursive self-improvement.