Definition
Plain language
A small open-source language-model training pipeline used as a benchmark for AI research agents.
As stated in the literature
Andrej Karpathy's compact GPT training pipeline, used as an agentic-optimization target where agents modify the training script to lower validation bits-per-byte.
Why it matters: Its small, self-contained pipeline makes it a clean target for testing whether agents can genuinely improve a model's training.
For example, an agent is told to edit this compact training script so the resulting model predicts text a little more accurately.
Heard on the show
“The benchmark is something Andrej Karpathy released called nanochat — a small language model training pipeline.”Episode 095 — Seven Wins to Zero: How Organizing AI Agents Like a Lab Changes the Search