Objective
Let students be two kinds of machine, back to back. First a calculator — a normal program that gives the exact same answer every time. Then an LLM — a next-word predictor that gives a different, believable answer each time. Feeling that switch is the single biggest "aha" about AI: it's why the same prompt gives different replies, why "temperature" exists, and why AI can sound sure but be wrong.
The setup
- Stand in a circle (or stay at desks — everyone just needs to hear each other). No devices.
- Pick a Scribe (you, or a student) to write answers on the board in two columns: CALCULATOR and LLM. Seeing "all the same" vs "all different" is the whole lesson.
- Decide two things in advance: a calculator rule (e.g. "double the number, then add 1") and a sentence starter for the LLM round (e.g. "On a rainy Saturday, the best thing to do is …").
The rules — read aloud
Round 1 · "You are a calculator"
- "You are all identical calculators. Your rule is: [state it — e.g. double it and add 1]."
- "When I call a number, everyone runs the rule and answers together, out loud, instantly."
- "Here's the catch: I'll call the same number more than once — and you must give the exact same answer every single time. No creativity allowed."
Round 2 · "Now you're an LLM"
- "Forget the rule. I'll say the start of a sentence. Each of you finishes it with the first believable next word that pops into your head."
- "We go around the circle — everyone says their own ending. There is no single right answer."
- "Round 2a — low 'temperature': give the most obvious, safe word. Round 2b — high 'temperature': give the most surprising word that still makes sense."
The process
- Round 1 (~3 min): call "5" → whole class shouts the result → call "5" again → same result again. Do 2–3 numbers, repeating each. Scribe writes them: identical every time. Name it: "Same input, same output, forever. That's a calculator — a normal program. We call this deterministic."
- Round 2 (~5 min): give the sentence starter. Go around; Scribe writes every different ending. Point at the board: "One prompt — how many different, totally valid answers?"
- Temperature (~3 min): run 2a (obvious words — they cluster) then 2b (wild words — they scatter). "That dial you just felt is real — it's called temperature."
- Compare (~2 min): put the two columns side by side. CALCULATOR = all identical. LLM = all different. The delta is the whole lesson.
The debrief
- Which machine would you trust to calculate your bank balance — and which to brainstorm a story? Why?
- In Round 2 nobody gave a "wrong" word — just different likely ones. So how could an LLM ever be confidently wrong? (It picks likely words, not verified-true ones — that's a hallucination.)
- Finish this: "A calculator gives the same answer every time. An LLM ______." (predicts a likely next word · varies · has a temperature dial)
WORKSHOP TIE-IN: This is what an LLM is — a next-word predictor, not a calculator. It's why the same prompt can give different answers, why the temperature setting exists (Round 2a vs 2b), and why AI can sound certain but be wrong. It sets up everything else: prompting, tools, and agents all sit on top of this one idea.
WATCH-OUT: In Round 1, insist on identical, in-unison answers — the boredom and predictability is the point. In Round 2, if a student freezes, say "there's no wrong answer, just say any word that fits" — that freedom is the contrast you want them to feel.