GitHub – karpathy/autoresearch: AI agents running research on single-GPU nanochat training automatically

**GitHub Spotlight: Autoresearch by Andrej Karpathy**

There’s something quietly thrilling about this idea. You go to bed… and your AI keeps working.

In karpathy/autoresearch, Andrej Karpathy shares a small but provocative experiment. The concept is simple, almost playful. Give an AI agent a compact language model training setup, let it run on a single GPU, and allow it to experiment on its own. It edits code, trains for five minutes, evaluates performance, keeps improvements, discards failures, and repeats. Over and over. Overnight.

By morning, you wake up to a research log and, ideally, a slightly better model.

If you’ve ever trained models manually, you know the rhythm. Tweak. Run. Wait. Check metrics. Repeat. It can feel like tending a garden one leaf at a time. Autoresearch shifts that dynamic. Instead of editing Python files directly, you guide the system through a simple program.md file. That file acts like lightweight “research DNA”, giving the agent context and direction.

The training setup is intentionally minimal. Single NVIDIA GPU, tested on an H100. Five-minute fixed training windows. A clean metric, validation bits per byte, so comparisons stay fair across architectural tweaks. It’s stripped down on purpose, more proof of concept than production framework.

Karpathy frames it almost like the origin story of autonomous research swarms, though this repository is firmly grounded in today’s hardware. Still, you can feel the trajectory. Start with one agent. Add more. Refine the “research org code.” Let iteration compound.

What makes this compelling isn’t scale. It’s accessibility. Frontier-style experimentation, but on a single GPU.

We’re watching research itself become programmable. Not just models, but the process of improving them. And if this is an early sketch of that future, it’s a fascinating one.