Running autoresearch Hands-On — Overnight Experiments on a Single GPU
From environment setup to agent execution and overnight results analysis. Tuning guide for smaller GPUs and practical tips.
Running autoresearch Hands-On — Overnight Experiments on a Single GPU
In Part 1, we looked at how Karpathy's autoresearch is structured. Here's the three-line summary:
- A single
train.pycontains the GPT model + optimizer + training loop. - An AI agent (Claude Code, etc.) modifies this file, trains for 5 minutes, and keeps the change if val_bpb improves — otherwise discards it.
program.mddefines the agent's behavior rules. Humans only edit this markdown file.
In Part 2, we'll set up the environment, launch the agent, and analyze the results from an overnight run.
Environment Setup — Getting Started
Requirements
| Item | Minimum | Recommended |
|---|---|---|
| GPU | NVIDIA GPU (CUDA support) | H100 80GB |
| Python | 3.10+ | 3.12 |
| Package Manager | uv | uv |
| Agent | Claude Code or Codex | Claude Code |
You don't need an H100. It runs on 4090, A100, 3090, and more. The difference is how many tokens get processed within the fixed 5-minute budget. We'll cover GPU-specific tuning later.
Related Posts
Paper of the Week #2 — The Score Is in the Abstract, the Bill Is in Table 2
I rebuilt HoH's Planner→Developer→QA loop (arXiv 2609.01481) around Claude Code on 8 hidden-test tasks: the score gap stayed inside rerun noise while tokens tripled, 58k vs 177k. HoH's own Table 2 reports 3.25x. Plus the matched-loss control promised in issue #1, graded.
CLAUDE.md Templates for Five Project Types — Copy, Paste, and the Rules That Actually Change Agent Behavior
Five CLAUDE.md/AGENTS.md templates — Next.js, Python ML, monorepo, data pipeline, research notebook — plus which lines agents actually obey and how to use a rules file to govern side effects.
Mobile Claude Code: three approaches, and what actually works
Three ways to reach Claude Code from your phone — tmux + SSH, /remote-control, and server-based agents. The real fix isn't "mobile support" but decoupling compute from your device.