AI self-improvement loops gain evidence, but oversight remains
Exponential View reports growing evidence of recursive AI self-improvement: Anthropic says Claude led 26% of its AI R&D in August under human supervision, while OpenAI posted 719 math manuscripts produced by an internal model in October. Sakana AI and the University of British Columbia measured gains of 30 percentage points on SWE-bench and more than 15 points on Polyglot; AlphaEvolve reportedly cut a matrix-multiplication kernel’s runtime by 23%, and MiniMax reported a 30% internal-evaluation gain for M2.7 after more than 100 rewriting rounds. The evidence remains incomplete: these systems mainly improved surrounding code, harnesses, memory or kernels rather than their own weights, and an inconclusive agent comparison found run-to-run variation exceeded the difference between AI- and human-built results.
The report provides concrete but qualified evidence that AI-assisted research loops are improving components of scientific and technical workflows, while showing that recursive self-improvement has not yet been established as autonomous weight-level development.
This source provides an excerpt. Continue reading on the original site.
#Anthropic#Claude#OpenAI#Sakana AI#University of British Columbia#Google DeepMind#AlphaEvolve#Gemini#MiniMax#M2.7