Course book
How Modern AI Models Work
Three illustrated chapters, one for each video course. The videos build working code; the chapters isolate the ideas and work each key calculation with small numbers you can check by hand.
- Chapter 1Four Decades of Removing ObstaclesFrom Boltzmann machines and backpropagation to AlexNet, dropout, distillation, layer normalization, capsules and the forward-forward algorithm.
- Chapter 2Deciding Without GeneratingThe prefill and decode phases, reading a decision from one pass, attention masks for encoders and decoders, and reusing the key-value cache.
- Chapter 3Learning From RewardThe agent and environment loop, discounted returns, automatic differentiation worked by hand, stable softmax and the policy gradient.