Nanochat Lets You Build Your Own Hackable LLM
Nanochat offers an accessible way to create your own customizable large language model, emphasizing user modification and experimentation.
Script: GPT-4o mini Voice: OpenAI TTS
Transcript
Host A Welcome back to the show! Today, we’re diving into Nanochat, a new project by Andrej Karpathy that lets you build your own hackable LLM.
Host B That's right! It’s designed to be minimal and accessible, encapsulated in a single script. Anyone can create a simple ChatGPT clone!
Host A And what's even cooler? It only costs about $100 for the computational work and takes roughly 4 hours on an NVIDIA GPU.
Host B Exactly! This gives you a micro-model with 1.9 billion parameters trained on 38 billion tokens. It can write stories and answer questions!
Host A But, of course, it’s not going to compete with the top commercial models right away. Scaling it up to $1,000 yields much better results.
Host B Right! The $1,000 version can tackle simple math problems and coding tasks. That’s a significant leap in capability!
Host A And with Karpathy’s track record, there’s a ton of potential for modification and experimentation. His past work shows just how creative you can get.
Host B Absolutely! This is an exciting time for anyone interested in LLMs. So, if you’re curious, check out Nanochat and start building your own model!