I Made AI Make AI (LLM)

A content creator uses Claude and Cursor to generate a complete codebase for a tiny GPT-style language model (~13.7M parameters) built with PyTorch. The project covers tokenizer training, data preparation, pre-training on a small markdown dataset, and supervised fine-tuning (SFT) on a consumer RTX 5070 GPU. Despite starting with gibberish outputs, iterative training and fine-tuning eventually produced coherent responses. The video documents the full workflow including debugging CUDA issues and Python version compatibility problems.