Volume 5

Volume 5 - Large Language Models

Volume 4 introduced attention and the Transformer as an architecture. This volume takes that architecture apart, piece by piece, and rebuilds each piece from scratch — how text becomes tokens, how tokens become vectors, how position gets encoded into a set that has none, how attention actually decides what to look at, and how many attention heads running in parallel become one Transformer layer. Then it follows a model through its full lifecycle: pretraining on raw text, fine-tuning it toward a task, aligning it with human preference through RLHF, and finally shipping it — inside a context window, quantized, sometimes distilled, and served at a latency and cost a real product can afford.