Train you own ChatGPT in $100 using nanochat

From $100 to ChatGPT in Four Hours: Inside nanochat, the 8,000-Line Blueprint for Building Your Own Language Model

In an age where state-of-the-art language models demand billion-dollar budgets, massive data centers and armies of engineers. Andrej Karpathy has quietly dropped a bombshell that redefines what’s possible on a shoestring budget. His latest open-source project nanochat is not another gargantuan transformer with hundreds of billions of parameters. Instead, it’s a complete end-to-end implementation of a ChatGPT-like system built from scratch in a single, clean and remarkably compact codebase.

For just $100 and four hours of compute time you can train, fine-tune and deploy your very own conversational AI complete with a web UI, evaluation metrics and even optional reinforcement learning. And you don’t need a PhD or a corporate cloud account to do it. All you need is access to a single 8-GPU machine (like an 8×H100 node) and a willingness to run one command: bash speedrun.sh

This isn’t just another GitHub repo. It’s a pedagogical masterpiece, a minimalist engineering triumph and a radical act of democratization all rolled into 45 files and roughly 8,000 lines of Python.

Bridging the Last Mile Gap in AI Education