NemoClaw Knowledge Wiki

Karpathy's AutoResearch: An AI Agent for Independent LLM Program Improvement

3 min read

Karpathy’s AutoResearch: An AI Agent for Independent LLM Program Improvement

Generated: 2026-04-26 · API: Gemini 2.5 Flash · Modes: Summary

Karpathy’s AutoResearch: An AI Agent for Independent LLM Program Improvement

Clip title: I Used Karpathy’s Autoresearch to Train an LLM! Author / channel: Thu Vu URL: https://www.youtube.com/watch?v=XXR0zZ0_16M

Summary

This video explores the burgeoning field of AI-generated code, noting that over 41% of global code is projected to be AI-generated by 2026. The main topic revolves around “AutoResearch,” an open-source project by Andrej Karpathy (former Tesla Director of AI and OpenAI co-founder), which enables an AI coding agent to autonomously improve a program in a loop. This initiative signifies a significant shift in human-AI collaboration, moving beyond “vibe coding” (human prompts, AI writes, human reviews) and “agentic engineering” (human orchestrates agents) to “independent research,” where humans act more as “research advisors” by setting direction and letting the AI run independently.

The core of AutoResearch’s design is built upon three distinct files: prepare.py, train.py, and program.md. The prepare.py file handles immutable aspects like data preparation and defining the evaluation metric. The train.py file serves as the agent’s sandbox, containing the training code that the AI agent is allowed to modify and iterate upon. Crucially, program.md contains the human-written instructions, research directions, principles, and constraints, guiding the AI’s experiments. The iterative process, dubbed a “ratchet loop,” involves the AI proposing hypotheses by modifying train.py, committing changes, running training for a fixed budget (e.g., 5 minutes), and evaluating the results against the defined metric. If an improvement is observed, the commit is kept; otherwise, it’s reverted, ensuring that only progress is maintained.

For an AI agent to be “self-improving” in this context, three conditions must be met: a clear, automatically measurable metric (ideally a single number), a single file for the agent to edit, and a time-boxed experimental loop. The video demonstrates this by attempting to train a tiny language model on folklore and mythology datasets using Mistral Vibe, an open-source command-line coding assistant that supports agentic capabilities, including an “auto-approve” (dangerous) mode. After setting up the environment, installing dependencies, and preparing the data, the AI agent is given instructions via program.md to begin its autonomous research.

Over several hours, the agent ran multiple experiments, making various modifications to the train.py file in an attempt to optimize the validation bits-per-byte (val_bpb) metric. Out of eleven experiments, six resulted in kept improvements, visibly decreasing the metric in a step-like graph. While the generated story samples still lacked perfect coherence, the final iterations showed notable improvements in grammar and sentence structure compared to the baseline. The key takeaway emphasizes that while AI agents can effectively handle the execution and iterative aspects of research, the human role in defining worthwhile research directions, understanding the problem, and providing the ultimate judgment for what constitutes “better” remains paramount. This ability to frame problems for autonomous agents is presented as a vital skill for the future.

Graph View