From Zero to Agents: A Foundational AI/ML Course, Built From First Principles

Volume 1 built the vocabulary — vectors, matrices, gradients, and a neural network assembled from a single perceptron up. Volume 2 spends that vocabulary on the machinery that actually makes modern language models work.

This volume covers three modules, each derived, implemented, and verified with the same no-shortcuts standard as Volume 1. Sequence Models and Attention confronts the vanishing-gradient problem across time, fixes it with LSTM gating, then rebuilds attention from real learned projections into a complete, working, generative transformer trained from scratch. Fine-Tuning and Adaptation takes a pretrained model and adapts it efficiently with LoRA — deriving the low-rank update, choosing and allocating its rank with a principled method, measuring quantization's real accuracy cost, and testing a common claim about catastrophic forgetting directly rather than assuming it. Agents arrives, finally, at the course's own title: the ReAct loop, multi-tool selection, automatic failure detection, agent memory colliding with an embedding problem from Volume 1, and a working multi-agent system built as a direct extension of a single agent's own loop.

Every result in this book was run, not asserted — including the places where the first fix only partially worked, and the book says so. Two genuine, unresolved research directions are flagged along the way rather than papered over.

Volume 2 closes with a conclusion tying both volumes together: what was built, what the throughline actually was, and what's still genuinely open.