Download
autoresearch is an experimental framework that enables AI agents to autonomously conduct machine learning research by iteratively modifying and training models. Created by Andrej Karpathy, the project allows an agent to edit the model training code, run short experiments, evaluate results, and repeat the process without human intervention. Each experiment runs for a fixed five-minute training window, enabling rapid iteration and consistent comparison across architectural or hyperparameter changes. The system centers on a simple workflow where the agent modifies a single training file while human researchers guide the process through a program.md instruction file. Designed to run on a single GPU, it keeps the research loop minimal and self-contained to make autonomous experimentation practical. Over time, the agent logs experiments, evaluates improvements, and gradually evolves the model through automated trial-and-error.
Features
- Autonomous research loop where AI agents modify code, train models, and evaluate results.
- Fixed 5-minute experiment cycles enabling rapid iteration and fair comparisons.
- Single editable training file allowing agents to modify architecture, optimizer, and hyperparameters.
- Lightweight setup running on a single GPU with minimal dependencies.
- Human-guided research instructions defined through a programmable program.md file.
- Automated experiment logging that tracks progress and improvements over time.
Project Samples
