Andrej Karpathy builds a Generatively Pretrained Transformer step by step, following the 'Attention is All You Need' paper — a slow, spelled-out derivation that has become the canonical GPT-from-scratch tutorial.