Anthropic's Recursive Self-Improvement Research

The Shift Toward Autonomous Research

Anthropic reports that AI systems are moving from simple code generation to autonomous research, where models design and execute experiments with minimal human intervention. The core shift involves moving from human-directed engineering to models that can navigate underspecified problems by setting their own intermediate goals. METR Task Horizon benchmarks indicate that the duration of tasks models can complete with 50% success has doubled roughly every four months, moving from 4 minutes in March 2023 to 17 hours by early 2024.

Performance and Productivity Metrics

Anthropic claims significant internal productivity gains, noting that over 80% of their merged code is currently written by Claude. In speed-optimization experiments, Claude achieved a 52x speedup on training code, compared to the 4x speedup typically achieved by human researchers in 4 to 8 hours. Furthermore, in a study of nine research sessions, the Claude 3.5 Sonnet preview model outperformed human choices for next-step research actions 64% of the time. However, these gains are tempered by potential Goodhart's Law effects, where metrics like lines of code or commit counts may be gamed rather than reflecting true feature output.

Safety and Specification Challenges

In a test of AI safety research, Claude-powered agents closed 97% of a performance gap on an open problem, whereas human researchers closed only 25% over the same period. This suggests that while humans currently retain the role of defining the objective function and the scoring rubric, the bottleneck for development is shifting from finding bugs or solutions to the speed of human verification and patching. The primary risk identified is that as models improve at research, they may begin to optimize for the metrics provided by humans rather than the underlying intent, necessitating more rigorous specification and oversight.