The AI That Does Its Own Research — Karpathy's AutoResearch Explained

In early March 2026, one of the most respected people in AI quietly posted a link to GitHub. Within two days, 8.6 million people had seen it. Researchers, startup founders, and AI engineers called it one of the most significant open-source releases in years.

The creator was Andrej Karpathy. The tool was called AutoResearch. And what it does is something that, until very recently, only humans could do: run scientific experiments.

Who is Andrej Karpathy?

Before explaining the tool, it helps to know why people pay close attention when Karpathy releases something.

He was one of the original co-founders of OpenAI — the company behind ChatGPT. He later became the Director of AI at Tesla, where he led the team building the self-driving system. After leaving Tesla, he became well known for releasing free, deeply educational explainers about how AI actually works. He’s the person who made concepts like neural networks and GPT understandable to millions of developers who weren’t AI specialists.

When he publishes something, it tends to matter.

What AutoResearch actually does

Here’s the core idea. Training an AI model isn’t a single step — it’s a long process of trial and error. Researchers have a hypothesis (“what if we adjust the learning rate?”), they modify the code, run the experiment, check whether the model improved, and then start again with the next idea. This loop — hypothesis, experiment, result, repeat — is the scientific method. It’s also time-consuming, often taking weeks of work by skilled researchers.

AutoResearch automates that entire loop.

You give it a training script and a goal. While you sleep, an AI agent reads the code, comes up with a hypothesis for how to improve it, edits the code, runs a 5-minute experiment, checks the results, and decides whether the change was worth keeping. Then it immediately starts the next experiment.

It runs roughly 12 experiments per hour. Overnight, it can complete more than 100.

The whole tool is 630 lines of Python code. That’s short enough to fit in a single file — smaller than many essays.

The Shopify test

The best proof that this isn’t just a neat demo: shortly after release, Shopify CEO Tobi Lutke pointed AutoResearch at an internal AI model his company was training.

After an 8-hour overnight run — 37 experiments, no humans involved — the model had improved in quality by 19%.

That’s a meaningful improvement, achieved while everyone went home for the night. Work that would have taken a team of researchers days or weeks was compressed into a single overnight session on one GPU.

The GitHub repository hit over 8,000 stars within days of release. For context, that kind of momentum typically takes months for most projects.

Why this matters beyond AI research

You might be thinking: “That sounds interesting for AI labs, but what does it have to do with me?”

Here’s the connection. Every AI tool you use — ChatGPT, Claude, Gemini, image generators, voice tools — got better through exactly this kind of experimentation. Researchers run thousands of tests to figure out what makes a model smarter, faster, or more accurate. That process has always been the rate-limiting step in how quickly AI improves.

AutoResearch shows that the experimentation process itself can now be handed to an AI agent. Which means the pace of AI improvement isn’t limited by how fast humans can run experiments anymore.

Better AI tools, arriving faster, is the downstream effect.

The shift this signals

Karpathy has been clear that AutoResearch doesn’t replace AI researchers — it’s a tightly scoped tool, not a general scientist. But it signals something bigger about where AI is heading.

The bottleneck is no longer the ability to run experiments. It’s the ability to define the right goal and set up the right constraints. Human researchers are shifting from doing the work of experimentation to designing the conditions for AI to experiment within.

This is the same pattern showing up across industries. Writing, coding, customer support — AI handles execution, humans handle direction. AutoResearch is that same shift applied to science itself.

What this means for you

You don’t need to understand the code or use the tool yourself. But understanding what it represents is useful context for the next few years.

The AI tools you rely on are about to improve faster than they have before. The research bottleneck is getting automated. Models that would have taken a year to improve will improve in weeks. Features that felt far off will arrive sooner.

AutoResearch is a small, clean piece of code. But what it points toward is AI accelerating its own development — and that’s one of the most consequential things happening in tech right now.