How Close Is Recursive Self-Improvement?

AI companies increasingly report that they are approaching or achieving recursive self-improvement, where AI systems help develop future AI systems in a recursive loop. In a new report, authors Simas Kučinskas and Josh Rosenberg examine reported data, public benchmarks, and existing forecasts to assess how close recursive self-improvement really is, concluding that while AI is a capable research assistant, humans still supply most research judgment. An ensemble of frontier LLMs used in the study forecasts about a 40 percent chance of a fast capability year by 2030. Open article →