8 · AI I — chat & assistants

Thursday, Sep 17, 2026

Materials for this session are not published yet. They appear here before class.

Objectives

By the end of this session you can:

  • Pick the right AI tier for a task.
  • Write a context-rich prompt.
  • Critique a model’s explanation against the code it explains.
  • Name the canonical failure modes, and the fact about how models work that each one comes from.
  • Say why the deliverable of AI-assisted research is a script, not an answer.
  • Write the disclosure block.

What we cover

  • Four facts about how these models work: they predict the next word, so plausible is not the same as correct; they generate rather than look up, which is where invented references come from; they are not deterministic, so a chat transcript is not a record anyone can re-run; and their memory of a conversation is finite, so long chats quietly degrade.
  • The chat tier of AI tools: explaining unfamiliar code and errors, drafting text and READMEs, planning an analysis, asking questions about data.
  • The failure modes — each one a consequence of the four facts: confident wrongness, invented references, plausible-but-wrong recipes.
  • Context is most of the game — what you paste in determines what you get back.
  • Agentic assistants for non-code chores.
  • The standing rule for research use: ask for the script, not the answer. An answer you cannot re-run is not a result; the research artifact is always code.
  • The disclosure block, which appears on every homework from here on.

Verification habit. Never accept an answer you couldn’t check — and verify one claim per conversation, on principle.