Hard ideas, drawn one step at a time.
Short narrated animations on how AI works, calculus and statistics. Every lesson is one step from the ideas it rests on, so you can always step back and never fall behind.
Where to next
Nearest ideas first, counted in minutes from where you are. Choose one to make it your destination.
- What LLMs do1 minStart here
- Functions2 minStart here
- The matrix math2 minStart here
- Probability2 minStart here
- EmbeddingsWhat LLMs do3 min1 lesson away
- QuantizationThe matrix math4 min1 lesson away
- Why attention?Embeddings4 min2 lessons away
- Counting outcomesProbability4 min1 lesson away
- Sine and cosineFunctions4 min1 lesson away
- SlopeFunctions5 min1 lesson away
- Neural networksThe matrix math5 min1 lesson away
- TokenizationEmbeddings5 min2 lessons away
Pick where you want to get to
You get every lesson it rests on, in order. Skip the ones you already know.
How ChatGPT works
8 stops14 min
- What a language model actually does1:28How language models work NEXT
- Tokens and embeddings1:27How language models work
- The problem attention solves1:19How language models work
- The only math you need2:07How language models work
- Neural networks2:27How language models work
- Attention: words talking to each other2:15How language models work
- Multi-head attention + the MLP1:50How language models work
- The whole Transformer, assembled1:31How language models work DESTINATION
Courses
Each course is a line of short lessons, in the order they build on each other.
-
How language models work
20 stops34 min
From next-word prediction to the full Transformer — tokens, embeddings, attention, the MLP, training — then what turns it into a chat assistant.
Open the courseWhat LLMs doInterpretability
-
Frontier AI
10 stops18 min
The engineering behind today's models: flash attention, long context, vision, LoRA, quantization, distillation, retrieval, agents and diffusion.
Open the courseFlash attentionDiffusion
-
Calculus
12 stops27 min
Functions, slope, and sine and cosine; then limits, the derivative, the integral and the fundamental theorem — ending at the gradient, the idea that trains neural networks.
Open the courseFunctionsThe gradient
-
Statistics
10 stops21 min
Probability, counting, distributions, expected value and spread, Bayes' rule, the law of large numbers, the normal curve and sampling.
Open the courseProbabilitySampling