Karpathy drops 630-line AI that trains itself

Andrej Karpathy's autoresearch project automates AI model training with just 630 lines of code, letting AI agents modify and optimize LLMs autonomously.

A few days ago, a new project on GitHub exploded in popularity.

Andrej Karpathy, former OpenAI researcher and ex-Tesla Director of AI, released a new project called autoresearch. In just about a week, it skyrocketed to 35.6k stars.

What made this project create such a huge sensation wasn’t that it used some mind-blowing new technology. Rather, it showed us a glimpse of the future: AI is no longer just writing code — it can now take over the core of AI algorithm research itself.

Give an AI a real LLM training environment, let humans guide it through prompts, and let the AI modify Python code, tune parameters, change architectures, and run model training experiments around the clock.

This is no longer “using AI to write code.” This is true Automated AI Research.

This is what Karpathy calls the “post-AGI feeling”.