this EX-OPENAI RESEARCHER just released it...

TL;DR

Andrej Karpathy released an open-source AI agent that autonomously runs machine learning experiments overnight to improve its own training code.

Key Points

• 1.What it is: "AI Auto Researcher" — a free, open-source tool on GitHub that runs on a home computer, using an LLM agent to autonomously edit training code, run 5-minute experiments, keep improvements, discard failures, and repeat.
• 2.How it works: You write plain-English instructions in a `program.md` file; the agent modifies `train.py`, tests changes (architecture, hyperparameters, optimizer, batch size), logs results, and loops — no human touching the code.
• 3.Real results: Over 2 days it ran ~700 experiments, found 20 stacking improvements, and reduced GPT-2 training time from 2.02 hours to 1.8 hours — an 11% speedup — on Karpathy's already well-tuned NanoGPT project.
• 4.Karpathy's reaction: He called it "wild," saying the agent performed the full daily workflow he's done for two decades — reading results, forming hypotheses, testing, iterating — entirely autonomously.
• 5.Scalability signal: All improvements found in small models transferred to larger models, suggesting the findings aren't just toy results but potentially broadly applicable.
• 6.The bigger vision: Karpathy is exploring a distributed, crowdsourced version where thousands of people run agents simultaneously, all contributing to one shared codebase — a global open-source recursive self-improvement network.
• 7.Why it matters: Unlike lab-internal AI research automation (Google AlphaEvolve, Sakana AI), this is publicly accessible, meaning the "intelligence explosion" trigger may come from a global open community rather than a single secretive lab.