AI systems improving the research process that builds their successors

A new paper examines what happens when AI systems start automating the research needed to build better AI. Instead of merely helping with code or experiments, a system could improve the methods and tools used to create its successor, forming a feedback loop that accelerates capability gains beyond ordinary model scaling.

The paper argues that policymakers need visibility into how extensively frontier companies have automated model development and how close they are to recursive self-improvement. Its authors include Jakub Pachocki, Geoffrey Hinton, Yoshua Bengio, and Jack Clark, alongside researchers from academia, civil society, and frontier AI companies.