autoresearch

autoresearch by karpathy is a pioneering AI agent research tool designed to automate the exploration and optimization of Large Language Model (LLM) training. It empowers AI agents to autonomously modify model architecture, hyperparameters, and core training logic within `train.py`. Operating on a strict 5-minute training budget, the system evaluates improvements using `val_bpb` and iteratively refines configurations. Guided by human-programmed `program.md` files, autoresearch aims to discover optimal LLM setups on a single GPU, significantly accelerating frontier research and enabling efficient, unsupervised overnight experimentation.

• Autonomous AI Agent Optimization: Agents autonomously modify model architecture, hyperparameters, and training logic for LLM configuration discovery.
• Fixed-Time High-Efficiency Iteration: Each training run is strictly 5 minutes, with `val_bpb` as the core metric for rapid evaluation and iteration.
• Human-Programmed Guidance: Researchers guide AI agents by programming instructions within `program.md` files.
• Single-GPU Environment Optimization: Designed for a single NVIDIA GPU, efficiently discovers optimal model configurations tailored to the specific platform.
• Accelerated Frontier Research: Significantly shortens experiment cycles, enabling automated, unsupervised overnight experiments to drive LLM breakthroughs.