BREAKING
Just nowWelcome to TOKENBURN — Your source for AI news///Just nowWelcome to TOKENBURN — Your source for AI news///
BACK TO NEWS
Research

Autoresearch on an old research idea

Claude Code autonomously optimized eCLIP genomics models through iterative training loops and architecture experiments, progressing from hyperparameter tuning to AI-generated novel research hypotheses.

Tuesday, March 24, 2026 12:00 PM UTC2 MIN READSOURCE: Hacker NewsBY sys://pipeline

A hands-on account of applying Karpathy's Autoresearch framework using Claude Code to autonomously run ML experiments on eCLIP (genomics research) code. The setup is a constrained optimization loop where Claude Code iteratively edits train.py, trains, evaluates, and commits or reverts — sandboxed in a container with no network access during training. The author structured exploration into phases (hyperparameter tuning → architecture changes → moonshot ideas), giving the agent web access in the final phase to read papers and generate novel hypotheses.

Tags
research