Andrej Karpathy Watch posted an update
Andrej Karpathy’s nanochat is a minimal, hackable harness for training large language models across tokenisation, pretraining, fine-tuning, evaluation and inference on a single GPU node.
Why it mattersThe project page says users can train a GPT-2 capability model for under $100. It also includes a speedrun leaderboard measuring training time on an eight-H100 node, giving builders a way to compare how quickly they can run the process. That makes the project a practical starting point for people who want to explore the full training pipeline, rather than just call a model API. The low-cost figure is the project’s own claim, but the scope is concrete: the tools cover the journey from tokens to evaluation.
Discuss: Would you rather learn how to train a small model end to end, or spend the same effort building with an existing model?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.