How LLMs work, in six short lessons.

From the moment your text is cut into tokens to the moment a model writes a confident wrong answer. Each step asks you to predict first, then shows what real tokenizers and models do. About 50 minutes in all, free, with no sign-up.

6 lessons32 steps26 interview questions

  1. 1TokensBefore a language model sees your sentence, a tokenizer cuts it into pieces and swaps each piece for a number.About 8 minutes · 5 steps
  2. 2EmbeddingsInside the model, every token becomes a long list of numbers, an embedding.About 8 minutes · 5 steps
  3. 3AttentionA word means different things in different company.About 9 minutes · 6 steps
  4. 4Picking the next tokenA language model doesn’t write a sentence; it scores every token it knows as the next one.About 8 minutes · 5 steps
  5. 5Context and the KV cacheEverything a model works with has to fit in its context window, and every token in that window takes memory while it writes.About 9 minutes · 6 steps
  6. 6HallucinationA model can write a confident, well-formed answer that is simply wrong.About 8 minutes · 5 steps

Start with lesson 1

Built from real models.

  • Predict, then see

    Every step starts with a question to guess. Guessing first, then checking, makes an idea stick far better than reading it.

  • Real numbers

    Token counts from the tokenizers themselves, next-token chances and answers from an open model run at full precision, and model sizes from their own config files. Nothing typed in by hand.

  • Ready for interviews

    Each lesson ends with the interview questions it prepares you for, their answers in our bank, and practice with Prism.