By skyforbes 
…A breakdown of compute, code and strategy to make your own conversational AI.
Andrej Karpathy has made it possible to build your own ChatGPT-style model in just four hours — for roughly $100 in GPU time. This guide shows exactly how to do it using NanoChat, a clean, open-source project that reveals how modern LLMs really work. 💥
🚀 What You’ll Build
By the end, you’ll have a mini-ChatGPT that can:
• Tokenize text efficiently with a Rust-powered encoder
• Train on educational data to learn language patterns
• Adapt for conversation through fine-tuning
• Optionally learn from reinforcement signals
• Serve as a local chatbot with a simple web UI
Everything runs in about 4 hours on an 8×H100 GPU node (~$100 in cloud cost).
🧠 What You Need
• Python 3.10+, PyTorch (CUDA), Git, and Rust
• A Hugging Face account for datasets
• A cloud GPU provider (RunPod, LambdaLabs, Modal, or JarvisLabs)
• Some comfort with Python scripting and terminal commands