Andrej Karpathy Releases Open-Source AutoResearch Tool for Autonomous AI Experiments
- Andrej Karpathy released AutoResearch, a 630-line Python tool that enables AI agents to run autonomous machine learning experiments on a single GPU.[3]
- The tool modifies code in a train.py file, trains models for five minutes per run, evaluates using validation bits-per-byte metrics, and iterates without human intervention.[1][3]
- The GitHub repository gained over 8,000 stars within days of its March 2026 release.[1]
- AutoResearch supports around 100 experiments overnight on commodity hardware.[1]
Andrej Karpathy, former director of AI at Tesla and co-founder of OpenAI, has open-sourced AutoResearch, a compact Python framework that allows AI agents to independently conduct machine learning experiments on a single GPU. Released on GitHub in March 2026, the tool automates code modification, short training runs, evaluation, and iteration, completing numerous experiments without human oversight.[1][3]
Tool Functionality
AutoResearch operates as a self-directed loop where an AI agent modifies a single training file, train.py, trains a model for five minutes, assesses validation bits-per-byte (val_bpb) metrics to check for improvements, and repeats the process. The agent continues indefinitely until manually stopped, enabling up to 100 runs overnight on standard hardware. This minimalist 630-line implementation focuses on LLM pretraining tasks like nanochat training.[1][3][5]
Release and Reception
Karpathy published the framework on GitHub in March 2026, where it quickly amassed 8,000 stars. Demonstration videos show the agent handling real LLM training setups pointed at instruction files, highlighting its autonomous nature. The project builds on Karpathy's background in AI research from OpenAI and Tesla's Autopilot program.[1][2][4]