Artificial Intelligence, one class at a time
A growing series of interactive classes on modern AI. Each class pairs a hands-on visualization aid with the slides, so you learn the math by playing with it — not just staring at it.
Visualizations first
Every class ships with an interactive visualization. Drag sliders, flip switches, step gradient descent — see the math move.
From first principles
No hand-waving. We build from one line of calculus up to transformers and the GPU memory wall, so each idea sits on the one before it.
Patience scales
Concepts compound. What feels impossible at slide 30 is obvious by slide 80. The fundamentals keep paying dividends.
Classes 01–03 — available now
latest · 2026-05-01Intro to Neural Networks — From a Simple Neuron to the GPU Memory Wall
The whole deep-learning stack, built from one line of calculus. Derivatives and the chain rule, the 1943 McCulloch–Pitts neuron, a single linear neuron fit live with sliders, squared error and MSE, backprop derived stage by stage, gradient descent you can step by hand, PyTorch's autograd, and the GPU memory wall that makes all of it expensive.
Information Theory — Entropy, Cross-Entropy, and How Models Learn From Surprise
How to measure just how wrong your predictions are — from intuition to the math. Surprise and Shannon entropy, cross-entropy as the loss every modern model is actually trained on, KL divergence as the distance between distributions, maximum-likelihood estimation, and how sampling pulls real text out of a probability vector.
Applied Information Theory — Distillation, Scaling Laws, Speculative Decoding, and L²M
Where the math from Class 02 actually lives in modern systems. Distillation as KL between teacher and student (Hinton 2015, all the way back to Buciluǎ 2006), Chinchilla and the compute-optimal scaling laws, speculative decoding as a draft-verify trick that respects the target distribution, and L²M as the information-theoretic limit on long-context memory.