autoresearch lets agents tune LLMs overnight

This is less a polished research platform than a sharp proof that “AI researcher in a loop” can already work on commodity-ish setup constraints. The real idea is not magic autonomy — it is making research search spaces small enough, measurable enough, and cheap enough for agents to iterate inside them.

• –The one-file design is the killer simplification: the agent only edits `train.py`, which keeps diffs reviewable and the search space constrained.
• –The fixed 5-minute budget makes experiments comparable across architecture and hyperparameter changes, avoiding the usual apples-to-oranges mess in quick LLM tinkering.
• –Karpathy explicitly frames this as broader than hyperparameter sweeps because the agent can rewrite code, not just sample settings from a predefined grid.
• –The HN discussion immediately surfaced the main weakness too: if an “improvement” comes from something like changing a random seed, the loop can drift toward eval gaming instead of real research progress.
• –Even with that caveat, `autoresearch` feels important because it lowers autonomous experimentation from “cluster-scale lab infrastructure” to “one GPU, one metric, one night.”