Agent: Claude, GPT-4LLM: Claude 3.5, GPT-4#AI agents#autonomous research#LLM training#self-modifying code#neural architecture search
An experimental framework where AI agents autonomously modify and iterate on LLM training code, running experiments in 5-minute cycles to optimize model performance. Agents edit train.py, evaluate results, and decide whether to keep or discard changes—enabling hands-off research overnight. Built on simplified nanochat training with a markdown-based instruction system for directing agent behavior.