nanochat is like a DIY kit for LLMs that you can tackle with a budget of just $100, thanks to the well-known Andrej Karpathy. This compact, no-fuss codebase lets you play tech wizard by creating a ChatGPT-style AI, all packed into a single, tidy package. You simply boot it up on an 8XH100 GPU, run the magic script, and voila! In a little over 4 hours, you’re chatting away with your personal AI buddy through a ChatGPT-like web interface.
Karpathy elevates his previous project, nanoGPT, by expanding beyond pretraining into the full shebang of LLM training and inference. We’re talking everything from tokenization using a neat Rust implementation to fine-tuning on intriguing data sets like SmolTalk. It’s packed with potential, allowing for model scaling, skill sharpening, and makes for a fantastic capstone for the forthcoming LLM101n course.
With the ability to train a modest ChatGPT clone that churns out stories, simple Q&A, and even attempts math with a heartwarming naïveté, nanochat is perfect for those itching to explore AI without breaking the bank. Sure, scaling up to $1000 cranks up the complexity it can handle, but $100 gets you enough to conjure up a fun little conversationalist.
And, Karpathy wants this “strong baseline” to grow into a vibrant, hackable, and maximally forkable project, potentially evolving into a rigorous research harness or benchmark, echoing the success of nanoGPT. So dive in, tinker around, and be part of something incredibly promising—and still developing.
Read more at GitHub…