一个 AI 代理发现了 Karpathy 20 年来错过的 20 个机器学习改进

Andrej Karpathy released autoresearch on GitHub last week, showcasing the unique aspects of its architecture. The framework consists of 630 lines of Python code that runs an AI agent in a loop: reading a training script, forming a hypothesis, modifying the code, running a short training job (five minutes), evaluating results against a scalar metric, and repeating the process. On Karpathy's own ML training setup, the agent conducted 700 experiments over two days on a single GPU, achieving an 11% improvement in training speed. This outcome indicates that AI agents can significantly optimize processes in the field of machine learning, advancing research progress.