Your robot does exactly what your code says.
ChatGPT has no code that says what to answer. So… what's going on in there?
0:30 · HOOK. "You've all programmed a robot. It does EXACTLY what your code says — nothing more.
ChatGPT is different: nobody wrote an if-statement for your question. Today: the whole trick, in 20 minutes."
Keys: ← → navigate · N notes · T timer (press now!) · B blank screen.
The whole trick
You already do this
Peanut butter and ___
After the tournament we all felt ___
ChatGPT is a machine that got scary good at this one game: finish the sentence.
1:30 (cum 2:00). Kids shout "JELLY!" — click the blank to reveal. Second one: no right answer — answers differ! Click it (shows "???").
Then click anywhere in the middle / press ↓?? — actually press R on this slide or click the second blank to show the punchline.
Punchline: "That's the WHOLE trick. Everything else today is just: how did it get so good at finish-the-sentence?"
Step 1 · Demo
Robots don't read words
0 tokens
2:00 (cum 4:00). "Shout me a LONG word." Type it. "How many chips? Vote!"
Teaching moment: robot = 1 chip, but robotics = robot + ics. It learned LEGO-brick pieces of language.
Click "Show the numbers": "to the AI, your sentence is JUST this list of numbers. It has never seen a letter."
(Honesty if asked: our chip-splitter is a mini version we built; real ones work the same way with ~100,000 pieces.)
Step 2
One token at a time, forever
→
THE BOX
→
It guesses one token… sticks it on the end… then reads its own sentence again to guess the next.
1:00 (cum 5:00). Demo auto-plays. Narrate the loop. "That dashed box is the context window —
a conveyor belt that only holds so many chips. When it's full, the oldest fall off the back.
That's why it forgets the start of really long conversations."
Step 3 · Game
Be the AI
___
2:30 (cum 7:30) ⏱ CHECKPOINT — should be at 7:30 here.
Per round: read prefix → 10 sec of shouting → Reveal → "did the class beat the AI?" Max 3 prefixes.
MUST do one spiky (Twinkle = 97%!) and one flat (breakfast — even the AI has no clue, it just picks something. Remember that for the fibs slide!).
Key line: "The box doesn't output an answer. It outputs a score for every possible token — then rolls dice."
Running late? Do only 2 prefixes (–0:45).
Step 4 · Demo
How did it get good? Practice.
It read (almost) the whole internet. Practice sentence: "twinkle twinkle little star"
WRONGNESS
98%
never trained
2:00 (cum 9:30). "It starts knowing NOTHING." Press TRAIN repeatedly — kids read the garbage out loud, cheer when it lands.
"Each press: guess the next token, peek at the real one, nudge every dial a tiny bit. A trillion times.
Sound familiar? Run the mission → watch it fail → tweak → rerun. You ARE the training loop for your robot."
CUT #1: if behind at slide 5's checkpoint, skip this demo — say it verbally (saves 1:30).
Inside the box
A billion dials. Zero facts.
1,000,000,000+
tiny dials ("parameters"), each nudged a hair during practice
📚 fact file
there is no database inside — no list of true things
It doesn't know the capital of Massachusetts.
"Boston" is just extremely probable after those words.
1:00 (cum 10:30). "When we say 'it knows' — careful. The dials make good guesses likely. Usually that LOOKS like knowing.
Hold that thought — in two slides you'll see what happens when probable and true disagree."
Demo
The randomness dial
🧊 Freezingsame question, every time:
"The best part of robotics is ___"
Write a one-line cheer for the GummyBears!
…
1:30 (cum 12:00). FREEZING: Ask 3× — identical, boring. THAT'S THE POINT. Slide to NORMAL: ask 3× — different every time!
"This is why you and your friend get different answers to the same question." SPICY: comedy.
Point at the bars: "the dial literally squashes the probabilities flat — bad guesses get a turn. It rolls weighted dice for EVERY token."
CUT #3: if behind, skip Freezing, start at Normal (–0:30).
Game · the important one
When the dice lie
Tell me about FIRST LEGO League.
Fibs caught: 0 / 3
2:30 (cum 14:30) ⏱ CHECKPOINT — should be at 14:30 after this.
Let it stream while saying "looks legit, right? Confident. Smooth." Kids will start yelling at the errors — good!
Click Fact-check mode, then click sentences kids accuse. 3 fibs planted.
THE STING: "You caught them because you LIVE FLL. It was exactly this confident about the wrong ones.
It wasn't lying — it rolled the dice from the 'Be the AI' game and they came up wrong-but-plausible. That's a hallucination.
Now: what about topics where YOU'RE not the expert?" NEVER CUT THIS SLIDE.
Pro tip
Vague memory vs. open book
🌫 The dials
a fuzzy memory of everything it ever read… somewhere… probably…
= vague memory of the whole internet
📖 What you paste in
your actual code, your actual essay, the actual rules PDF
= an open book right in front of it
Want better help? Give it the open book.
Paste your real robot code — don't make it guess from memory.
1:00 (cum 15:30). Most practical slide for them. "Asking 'why does my robot turn wrong?' = asking its fuzzy memory.
Pasting YOUR CODE and asking = open-book test. Way better answers."
CUT #4: if behind, merge this into slide 11 verbally (–1:00).
Step 5
From parrot to assistant
Raw model, fresh from reading the internet
"What's a good robot name?" What's a good team name? What's a good mascot? 47 Questions Every Coach Should Ask. Click here to subscribe—
After assistant school
"What's a good robot name? Here are 3 ideas: Sir Buildsalot, Gearbear, Brick Lightyear. Want sillier or more serious?"
Humans wrote tens of thousands of example conversations to teach it how to be helpful. It learned the style of helping.
1:00 (cum 16:30). "Raw, it just CONTINUES text — your question might get answered, or get 3 more questions, because
that's what internet text looks like. So humans wrote example dialogues: question → helpful answer. Same finish-the-sentence game,
new textbook. But who decides what counts as 'helpful'? → next slide, that's YOU."
Game · final boss
You are the trainer
Its personality right now
😊 Helpful
😤 Bossy
🥹 Suck-up
2:00 (cum 18:30). Max 3 rounds, keep it moving. Class votes by show of hands per card; click 👍 under the winner
(👎 under the loser if you want drama). Meter shifts live. Round 3 is the TRAP — the suck-up answer is fun to upvote.
If they fall for it: final answer is a useless compliment-fest → "Congratulations, you trained it to flatter you instead of help you.
Real AI labs fight this exact problem — it's called gaming the reward."
If they DON'T fall for it: override one vote yourself "as a chaos experiment." NEVER CUT THIS SLIDE.
The actual point
Drive your own robot
CURIOSITY
Ask it why, not just what. Make it explain until YOU get it.
JUDGMENT
You fact-check — it can't. Confident ≠ correct. You proved that today.
ORIGINALITY
Your ideas in, its polish out. Never the reverse.
You'd never let your robot drive itself off the table. Same deal. It's a fast, confident teammate that sometimes makes things up.
You're the team captain.
1:00 (cum 19:30). Not preachy — frame as DRIVING. "It's an amazing teammate: tireless, fast, knows a fuzzy version of everything.
And you now know exactly what it is: a finish-the-sentence machine rolling dice. Use it like a power tool, not an oracle."
Trivia time
Did the ideas stick?
~2:30, game-show pace — this runs INTO Q&A time by design; it IS the recap.
Read the question, kids shout A/B/C (or point: A=left arm, B=both, C=right), click the crowd's answer.
Wrong picks shake red and stay visible — let them regroup and pick again. The green "why" line is the recap — read it aloud.
Hard stop at 20:00? Skip straight to the big-questions slide and do ONE of them instead.
No right answers
Chew on these
?
If ChatGPT writes your essay… whose ideas are in it?
?
It read the whole internet — including the wrong and mean parts. What should it do with those?
?
Would you let it referee an FLL match? What job would you never give it?
?
When it says "I think…" — is anything actually thinking? How would you tell?
Flexible (this is the Q&A springboard). Pick ONE question, 30 seconds think-pair-share, then take answers.
There are no right answers here — the goal is that they leave still arguing about it in the hallway.
Nice closer before the last slide: "You now know more about how this thing works than most adults. Use that."
Go build something.
Thoughtful, capable, responsible —
working with AI without giving up curiosity, judgment, or originality.
gummybearsrobotics.org
Questions?
0:30 (cum 20:00). DONE — Q&A spills past 20:00 by design.
Good seed questions if silent: "What's something you'd NEVER use ChatGPT for?" / "How would you catch a hallucination in homework?"
How ChatGPT Actually Works — take-home card
The whole trick: it's a finish-the-sentence machine. Text → number chips ("tokens") →
it scores every possible next token → rolls weighted dice → repeats, one token at a time.
Why it makes things up
There's no fact file inside — just a billion dials that make likely text. When "probable" and "true"
disagree, you get a hallucination: confident, smooth, wrong.
3 questions to ask any AI answer
How would I check this if a stranger said it?
Am I the expert here — and if not, who is?
Did I give it the open book (my real code/text), or make it guess from fuzzy memory?
The deal
CURIOSITY ask why, not just what ·
JUDGMENT you fact-check, it can't ·
ORIGINALITY your ideas in, its polish out