nanochat: a full-stack ChatGPT clone you can read in a weekend
Text2026-01-07
Source: github.com
Andrej Karpathy's minimal, hackable training-and-serving stack is a fast way to understand how modern LLMs are actually built.
If the rest of this section is about results, nanochat is about understanding. It's a small, readable codebase that takes you from tokeniser to training loop to a served chat UI — the whole pipeline, stripped of the production complexity that usually hides how it works.
It pairs well with a rented GPU for an afternoon. The libraries underneath are on the programming page; the models it imitates are on text & LLMs.