AutoResearch Guide
Unofficial fan-made guide
AutoResearch GitHub Guide

AutoResearch by Andrej Karpathy

AutoResearch is an open-source project that lets AI agents run repeated ML training experiments inside a small, self-contained codebase. This page is an unofficial guide that explains what the project is, how it works, and where to start.

What Is AutoResearch?

AutoResearch is a lightweight autonomous research loop for language model training. Instead of manually tuning everything yourself, you let an AI coding agent test changes, run short training jobs, compare outcomes, and keep iterating.

If you searched for autoresearch github, karpathy autoresearch, or autoresearch ai, this is the project most people currently mean: karpathy/autoresearch.

How AutoResearch Works

The repository is intentionally small. The core workflow revolves around three files and a tight experiment loop.

1. Prepare the environment prepare.py handles data preparation, tokenizer setup, and runtime utilities.
2. Let the agent modify training logic train.py is the main file the agent is expected to change while exploring model and optimization ideas.
3. Guide the research loop program.md acts like an instruction sheet for the agent, shaping how it should reason about experiments.
4. Run short experiments and compare outcomes The project uses a fixed-time training budget so each run stays comparable.

Quick Facts

Author Andrej Karpathy
Official Repo karpathy/autoresearch
Main Language Python
License MIT
Primary Use Case Autonomous LLM training experiments
Core Metric val_bpb

Requirements and Limitations

The original repository is built around a single NVIDIA GPU workflow, and the README notes it was tested on H100-class hardware. That does not mean smaller setups are impossible, but it does mean you should expect tradeoffs if you are running on a Mac, consumer GPU, or lower-compute machine.

  • smaller datasets are often more practical on limited hardware
  • lower sequence lengths can reduce memory pressure
  • smaller vocabularies and lower model depth may be necessary
  • forks may be more useful than the default setup on non-NVIDIA environments

Forks and Smaller-Hardware Options

One reason this project is interesting for searchers is that the ecosystem is already branching out. The original repository points to notable forks for MacOS, Windows, and AMD.

If your real question is not "What is AutoResearch?" but "Can I run something like this on my machine?", you will probably want to explore forks and reduced-compute adaptations instead of using the defaults unchanged.

Useful Links

Official GitHub Repo Source code, issues, pull requests, and updates

Visit GitHub

AutoRA Project A different project with a similar name

View AutoRA

Supporting Guides Follow-up articles for hardware and setup questions

Read the GitHub guide

FAQ

What is AutoResearch?

AutoResearch is an open-source project that lets AI agents run repeated ML training experiments inside a compact codebase.

Where is the AutoResearch GitHub repository?

The repository most people are looking for is github.com/karpathy/autoresearch.

Can AutoResearch run on Mac or smaller hardware?

Not with the original defaults in every case. Many people look to forks and lower-compute adaptations for Mac, Windows, AMD, and smaller GPUs.

Is AutoResearch the same as AutoRA?

No. They are different projects.

Is this an official site?

No. This is an unofficial fan-made guide and is not affiliated with Andrej Karpathy or the project maintainers.