Karpathy releases autonomous research repo 82K stars

Andrej Karpathy’s `autoresearch` repository is getting another wave of attention because it packages a specific idea in a very small form: let an AI coding agent run machine-learning experiments by itself on a single GPU. GitHub showed about 82,400 stars for `karpathy/autoresearch` on May 21, 2026. The repository was published about two months ago, and its README says the setup is a simplified single-GPU implementation of Karpathy’s `nanochat` training loop. The repo’s pitch is direct: an agent edits code, runs a short training job, checks whether results improved, and keeps or discards the change. ### What exactly did Karpathy release? GitHub’s README describes `autoresearch` as “AI agents running research on single-GPU nanochat training automatically.” The core workflow is not a general-purpose research lab in the abstract; it is a compact loop around LLM training experiments, with the agent modifying code, training for five minutes, evaluating the result, and repeating that process overnight. (github.com) March 2026 is the date Karpathy attached to the repo’s introductory note, which frames the project as the starting point for autonomous AI research systems. The repository is public, shows 36 commits in the current history snapshot, and includes files such as `train.py`, `prepare.py`, `analysis.ipynb` and `program.md`. ### Why are people focusing on the Markdown file? Karpathy’s README says users are “not touching any of the Python files like you normally would as a researcher.” Instead, the human operator writes or edits `program.md`, which provides the instructions and context that guide the agent’s behavior. (github.com) That design choice matters because it moves the human role up a level. The person defines goals, constraints and research directions in Markdown, while the agent handles the repetitive loop of proposing changes, running experiments and checking whether the metric improved. (github.com) The repo calls this setting up an “autonomous research org.” ### How small is the system Karpathy is showing? The repository is built around a single-GPU training setup, not a large distributed cluster. (github.com) The README says the code is a simplified implementation of `nanochat`, and the experiment loop is deliberately short, with five-minute training runs used as the basic unit of iteration. That makes the project easier to copy and inspect than a large internal research stack. (github.com) Karpathy’s files are visible in one public repo, and the GitHub page shows a relatively small top-level structure rather than a sprawling codebase. ### Where did the 82,000-star figure come from? GitHub itself showed 82.4k stars on the `karpathy/autoresearch` page when checked on May 21. Social posts circulating the link cited roughly 82,000 stars, and that figure matches the live GitHub display at the time of review. (github.com) The star count is a measure of attention, not performance. The repository page does not claim that star growth proves the system works broadly; it presents a concrete workflow, code, and documentation that other developers can inspect or fork. (github.com) GitHub also showed about 12,000 forks on the same page. ### What does the repo say happens when you run it overnight? The README says the agent modifies the code, trains for five minutes, checks whether the result improved, and either keeps or discards the change before repeating. (github.com) By morning, the user gets a log of experiments and, in Karpathy’s words, “hopefully” a better model. The next place to watch is the GitHub repository itself. As of May 21, the public page remained live at `karpathy/autoresearch`, with the README linking to Karpathy’s own posts for added context and the latest changes visible through the commit history. (github.com)