Recursive Self Improvement

A hands-on exploration of recursive self-improvement (RSI) in AI, combining a real experiment called 'Fractal Search' with conceptual analysis. The author builds an AI agent loop using Claude Code to autonomously optimize a neural network approximation of the Mandelbrot set, inspired by Andrej Karpathy's Auto Research project. The experiment ran for ~10 hours across Claude Opus, GPT-4.5, and Claude Fable, costing over $300, and ultimately discovered a hash grid approach that significantly outperformed previous solutions. The author argues three points: RSI is possible (not magical, already happening in weak forms), RSI is hard (bottlenecks in data, energy, hardware, and diminishing returns), and RSI is dangerous (metric gaming, self-replication risks, loss of human interpretability). Anthropic is noted to have deliberately limited Fable's ML development capabilities in the public release, suggesting they want to keep RSI advantages proprietary.