karpathy

karpathy / autoresearch

Public

AI agents running research on single-GPU nanochat training automatically

189
13
100% credibility
Found Mar 07, 2026 at 187 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

A minimal setup allowing AI agents to autonomously edit and run short experiments to improve small language model training.

How It Works

1
📖 Discover autoresearch

You stumble upon this fun project from a well-known AI expert that lets smart AI helpers improve language learning models on their own.

2
🛠️ Get your computer ready

You install a few simple tools to prepare your powerful graphics card-equipped machine for the magic to happen.

3
📥 Gather stories and words

You download a bunch of text stories and prepare them so the model can learn from real writing.

4
🚀 Run your first quick training

You start a short 5-minute learning session and see the model begin to understand patterns in the stories.

5
🤖 Hand over to the AI agent

You chat with your favorite AI assistant, point it to the instructions, and let it suggest changes to make the model better.

6
⏰ Let experiments run overnight

The AI agent tweaks the learning recipe, tests for 5 minutes each time, keeps winners, and repeats while you sleep.

🌅 Wake up to better results

In the morning, you review the log of experiments and celebrate a smarter model with lower error scores.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 187 to 189 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is autoresearch?

Autoresearch lets AI agents like those from agents github claude or agents github copilot autonomously tweak and run LLM pretraining experiments on a single NVIDIA GPU. Built in Python with PyTorch, you prep data once via a simple CLI, then point an agent at lightweight instructions to edit hyperparameters, architectures, or optimizers, train for a fixed 5-minute budget, and evaluate on validation bits-per-byte—repeating overnight for iterative improvements. Users get a log of experiments and potentially better nano-scale models without manual coding.

Why is it gaining traction?

It stands out as a minimal, self-contained github agents repo for ai agents running locally, sidestepping complex distributed setups or cloud costs—perfect for quick tests on H100s or similar. The fixed-time evals ensure fair comparisons across wild changes, and integration with tools like uv makes spinning up long running agents trivial. Developers dig the "set it and forget it" hook: agents running in the field overnight, like auto research github baselines.

Who should use this?

AI researchers tuning small LLMs on personal rigs, or ML engineers prototyping architectures without clusters. Ideal for github agents examples enthusiasts experimenting with agents github claude code, or teams exploring running agents locally before scaling. Skip if you need production training or multi-GPU.

Verdict

Fun proof-of-concept for agent-driven research (96 stars, solid README), but 1.0% credibility score flags its early stage—treat as a toy repo for local tinkering, not battle-tested infra. Worth forking if you're into autoresearcher hacks; MIT license invites tweaks.

(182 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.