From Zero to LLM: Machine Learning from First Principles to Large Language Models

Ask how a chatbot works and you usually get one of two answers. “It predicts the next word” is true but doesn’t explain how the prediction gets good. The other answer opens with a page of equations. This book takes the road between them.

It starts with a café trying to predict tomorrow’s iced-latte sales and builds, one idea at a time, to the models behind ChatGPT, Claude, and Gemini. Every step is there because the one before it hit a wall: attention because a single vector couldn’t hold a whole sentence, the transformer because reading one token at a time left a GPU idle, RLHF because a text predictor isn’t an assistant. Along the way you train and diagnose small networks in PyTorch, write attention and a small GPT yourself, and learn to estimate a model’s memory and compute bills.

The later parts cover what you meet at work: GPUs and training at scale, alignment and reasoning models, LoRA and QLoRA, quantization and inference, long context, multimodal models, agents, and how to tell whether a benchmark result means anything. Written for software engineers, data analysts, and technical product managers who use these models every day without having built one. You need basic Python and high school algebra; derivatives, matrices, probability, and PyTorch are taught as they come up.

Every book in the From Zero series is written for readers with no background: intuition and analogies first, math and detail only when they are needed, so a whole field becomes a path you can walk on your own.