Minigame In the News Discuss AI Theory Task Delegation Task
Back
What Is Generative AI? /what-is-generative-ai--ai-theory

Platform Awareness

1 / 14

You use this every day.
You have no idea how it works.

Teacher note

Source: 3a-transcript. Open on the tension, not a definition. Let the discomfort of 'I use it and can't explain it' sit for a beat before you resolve it.

Discussion drills

  1. Diagnose 1

    Claude is asked: "Who wrote the first proof of the Riemann Hypothesis?" It produces a confident two-sentence answer: a mathematician's name, a year, and a brief description of the method. The Riemann Hypothesis has not been proved.

    Walk through what happened at the token level. Why did the model produce an answer at all? Why is the answer confident? Where exactly did the generation depart from truth?

  2. Predict 2

    You type into Claude the first 40 words of a technical manual from 1987 about a niche industrial solvent extraction process. The manual is obscure enough that it almost certainly was not in training data.

    Predict three specific properties the continuation will have. Then predict two specific things it will lack. Explain each prediction using the mechanism.

  3. Construct 3

    Write two prompts on the same broad topic — one that keeps Claude on a well-worn path, one that forces it onto a thin path. Then write a one-sentence prediction for each about what the output will look like. What specifically did you change between the two prompts to move the path?

  4. Compare 4

    "Claude got it wrong because it doesn't really understand — it just pattern-matches." Using only the mechanism of next-token prediction, evaluate this claim. What is it getting right? What is it missing? Does "pattern-matching" actually describe what is happening?

  5. Falsify 5

    Someone argues: "If you give Claude more context — more information in your prompt — it cannot hallucinate about that information, because it will just use what you gave it." Make the strongest possible attack on this claim.

  6. Judge 6

    Rank these three tasks by how reliably next-token prediction alone can carry them — most to least reliable — and justify each position: (a) Translating a legal document from French to English. (b) Summarising a scientific paper Claude has never seen. (c) Answering "What time does the nearest Costco close on Sundays?"

Apply this in the project

Every claim Claude makes in today's brief is a prediction — some from well-worn paths, some from thin ones. Your job is to notice which is which.

Seoul AI 2026