Geek Out Time: Can AI Do Research by Itself? I Tried AutoResearch on Singapore Hawker Food

Andrej Karpathy’s autoresearch ( https://github.com/karpathy/autoresearch ) hit 62k GitHub stars in weeks. The idea is genuinely interesting: give an AI agent a real LLM training setup, let it run experiments autonomously overnight, wake up to a log of 100 experiments and hopefully a better model. The agent edits the training code, trains for 5 minutes, checks if the metric improved, keeps or reverts, and repeats — indefinitely, while you sleep. Over the weekend, I spent a Saturday afternoon trying to run it, but in a setup that is almost the opposite of what it was designed for. Instead of H100 GPUs and internet-scale data, I used Google Colab’s free tier (T4), swapped out Claude Code for DeepSeek because it was cheap and readily available, and trained a tiny GPT model on a Singapore hawker food dataset.

This was not meant to be a serious benchmark. It was more of a curiosity-driven experiment to see how far this autoresearch idea holds up when you strip it down to something small, cheap, and slightly improvised. What came out of it was less about model improvement and more about understanding when this pattern actually works, and why it sometimes does not.

This is an excerpt — the full article continues on Medium.