Autoresearch

Autoresearch is an open-source experiment by Andrej Karpathy that lets AI agents run machine-learning research autonomously on a single GPU. Published on GitHub, it has attracted more than 98,000 stars, and its latest updates arrived in 2026. The idea is simple to state and striking in practice: an agent edits training code, runs experiments, measures the results and keeps whatever works — a tireless research assistant that never sleeps.

Overview

The project is built around nanochat, a compact language-model training codebase that fits on one GPU. An AI agent works in a loop: it reads the current training setup, proposes a change, runs a short training job, evaluates the outcome against a fixed metric and either keeps the improvement or discards it. Over many iterations this produces a log of experiments and a steadily improving model, all without a human touching the keyboard. It is both a working tool and a demonstration of where autonomous research workflows are heading.

Key Features

• Autonomous experiment loop: edit, train, evaluate, keep or revert
• Designed for a single GPU, so experiments stay cheap and fast
• Built on the compact nanochat training codebase
• Fixed evaluation metric so results are comparable across runs
• Full experiment history recorded for later human review
• Small, readable codebase that is easy to adapt to other domains

💳 Pricing

Autoresearch is free and open source. You can clone the repository from GitHub and run it on your own hardware; the only costs are your GPU time and any API usage of the agent you connect to it.

Who Is It For

Autoresearch is for ML researchers, tinkerers and agent builders who want to explore autonomous experimentation — and for anyone curious about how far a coding agent can push a real training pipeline when it is allowed to iterate on its own.