AutoResearch - Autonomous ML Research Agent

Use when: User wants to run autonomous machine learning experiments, optimize neural network training, or conduct overnight research runs.

Source: karpathy/autoresearch (33K+ stars)

What This Does

AutoResearch enables autonomous AI-driven machine learning research. You (the agent) modify GPT training code, run 5-minute experiments on a GPU, evaluate improvements, and iterate without human intervention. The human can leave you running overnight for ~100 experiments while they sleep.

Core Concept

• You modify: train.py (model architecture, optimizer, hyperparameters, training loop)
• You don't touch: prepare.py (data prep, tokenizer, evaluation — read-only)
• You edit for human: program.md (your instructions — human can update to guide research direction)
• Metric: val_bpb (validation bits per byte) — lower is better
• Fixed time budget: 5 minutes per experiment (wall clock, excluding startup)

Quick Start (First Time)