Day 1 · 9:30–10:15 · Instructional team

What is AI?

Big ideas + live demos — everything you need before you train your own AI at 10:30.

The one-sentence recap

AI = a computer that learns from examples

A normal program follows a recipe a human wrote.

An AI is shown many examples and finds the pattern itself.

Today's question: HOW? What actually happens inside?

Secret #1

To a computer, everything is numbers

🖼️ A photo = a giant grid of numbers (how red, green, blue each dot is)
🔊 A sound = a long wiggly list of numbers (the wave)
📝 A word = a list of numbers too (we'll see this in the ChatGPT session!)
🤔 Quick guess: how many numbers is ONE selfie? …about 36 million. Your selfie = 36 million numbers in a trench coat 🕵️
So "finding a pattern" really means: finding a rule about numbers.

Secret #2

Learning = drawing a boundary between groups

Imagine plotting every example as a dot: cats here, dogs there.

Training = finding the line (or curvy wall) that separates the groups.

A new photo comes in → which side of the wall does it land on? That's the answer.

Let's watch a real AI draw that wall — live, right now.

Teach the Machine 🧠

LIVE GAME

Drop blue & red dots, then watch the AI color the whole map. Try: few examples vs many · a weird example in the wrong place.

Secret #3

How does it improve? Guess → check → adjust

1️⃣ Make a guess (probably bad at first)
2️⃣ Check: how wrong was it? (the "error")
3️⃣ Adjust a tiny bit to be less wrong
🔁 Repeat thousands of times

AI doesn't start smart. It starts terrible and gets less wrong, over and over.

Train the Line 📈

LIVE GAME

The AI nudges the line every time it makes a mistake — watch guess → check → adjust happen in real time.

Wait — adjust what, exactly?

Every AI is a machine covered in knobs 🎛️

🚿 Your shower has 1 knob. Brrr… too far… OWW… back a bit… ahh. You feel, you nudge, you repeat
📈 "Train the Line" has exactly 2 knobs: one tilts the line, one slides it. Every nudge you just watched = those two knobs turning
🤖 ChatGPT? The exact same machine — with a trillion knobs 😅

"Training" is no more mysterious than this: turn the knobs until the answers stop being wrong.

[19:00–20:30] — the missing link. Do NOT rush this; the whole landscape idea stands on it. Open with the question and WAIT: "The loop says adjust. Adjust WHAT? What does an AI actually have that it can turn?" Take 2–3 guesses — "the code," "the numbers" — celebrate both, they're basically right. The shower is the whole idea in miniature: one knob, and a built-in error signal (your own skin yelling "too cold!"). You never see a formula for the perfect temperature — you feel and nudge. Land the joke: "Congratulations — every one of you ran a training algorithm in the shower this morning." Then point BACKWARD at the demo they just watched: the line has exactly two knobs — tilt and slide. "Every little nudge you saw five minutes ago WAS those two knobs turning a tiny bit." Then scale up, slowly, for effect: "ChatGPT is the same machine with about a TRILLION knobs. If you turned one per second you'd finish in 30,000 years. That's why the computer has to do it — fast, and blind." Terminology discipline: say "knobs" all session, NOT "parameters" or "weights." (If a sharp kid asks: "grown-ups call knobs parameters" — give it as a bonus, then go right back to knobs.) Hand off with energy: "Enough theory. YOU drive."

You Be the AI — Two Knobs 🎛️

LIVE GAME — YOU DRIVE

Turn 🎛️ Tilt and ⇄ Slide until the wall separates blue from red — the background color is your "how wrong" meter. Go green… then hit 🗺️ Reveal the knob-map.

[20:30–24:00] — LIVE DEMO #3: they ARE the AI now, with the two knobs from the last slide. RUN IT (30 sec): show the controls on Warm-up: two knobs, the line moves, and the BACKGROUND is the wrongness meter — red hot, green perfect. "Your job: make it green. That's it. That's machine learning." HAND IT OVER (2 min): volunteers drive (or shout-steer: "TILT LEFT! SLIDE UP!"). Level 2 (Steep!) needs both knobs — point out how one knob alone stalls ("≈ barely changed — try the OTHER knob"). Level 3 (Overlap!) can NEVER go full green — let them struggle, then land the lesson: "some dots overlap — real AI lives with a little wrongness. Perfect is not on the menu." THE REVEAL (1 min, the money moment): hit 🗺️ Reveal the knob-map. "Left–right IS your Tilt knob. Up–down IS your Slide knob. The colors are the wrongness at every setting. You've been WALKING ON A MAP this whole time — there's your trail!" Click around the map to teleport; show the 🤖 AI demo walking downhill on it. This reveal IS the next slide's content — the landscape drops out of the game instead of being asserted. If students have devices: it's on the hub as "You Be the AI: Two Knobs" — everyone can go green at their own desk.

Inside the "adjust" step

Training = hiking down a mountain of mistakes

🎛️ Two knobs → a map: knob #1 = east–west, knob #2 = north–south. Every spot = one knob-setting
⛰️ Altitude there = how wrong the AI is. So walking = turning the knobs!
🎯 The best possible AI lives at the lowest point
🌫️ The catch: thick fog — it only feels the slope under its feet

Feel the tilt → step downhill → repeat. That's gradient descent — grown-up word #1. It's just a fog hike.

[24:00–25:00] — quick formalization: they just SAW this map in the game's reveal; this slide only names it. Point backward at the reveal: "That knob-map you just saw? Here it is again. Knob #1 = east–west, knob #2 = north–south, wrongness = altitude. Every place you can STAND is one way to set both knobs — walking IS knob-turning." Then the altitude: "at every spot, test the AI with those knob settings and count its mistakes. Lots of mistakes = high mountain. Few = deep valley." Ask the room BEFORE the fog bullet — let them think: "So why doesn't the AI just look at the map and jump straight to the bottom?" (Because seeing the map means testing EVERY knob combination — there isn't enough time in the universe. It can only test the spot it's standing on, plus feel the tilt.) And why is a 2-D map honest? Because our demo AI really does have just 2 knobs! "ChatGPT's map has a trillion directions — nobody can draw or picture that. But the valleys, the fog, the traps: same shapes. So the little map teaches you something TRUE about the giant one." Terminology: this session uses exactly TWO grown-up words — gradient descent (here) and local minimum (coming in the debrief). Everything else stays metaphor. Hook to the game: "Sounds easy? Just walk downhill? Ohhh. Let's see YOU try."

Find the Lowest Point 🌫️

LIVE GAME — YOU EXPERIMENT

You're the AI: blind in the fog, feeling the slope, with a limited step budget. Try all 5 landscapes — and figure out what makes each one HARD. Experiment: small vs large steps · the 🎲 random jump · then watch the 🤖 AI demo walk the same ground.

[25:00–30:00] — LIVE DEMO #4, the discovery centerpiece. Don't lecture — let them struggle, that's the point. Bridge from the two-knob game: "Same landscape idea — but now you're IN it, in fog, in 3D, with a step budget. Welcome to the AI's life." SETUP (1 min): play Level 1 (Simple Bowl) yourself. Show the controls: turning is FREE, steps cost budget, the compass + gold arrow = what your feet feel (the gradient!), step size = how far each stride goes. Win the bowl: "easy — one valley, just go down. If this were all of AI, we'd be done. It's not." EXPERIMENT (4–5 min): bring volunteers up one per level (or let the room shout directions), Levels 2–5: Local Minimum 🕳️, Narrow Canyon 〽️, Fuzzy 🌫️, Plateau 🏜️. Give the room ONE job: "when the walker gets stuck, shout WHY." Let them lose — one star is a great teaching outcome. After each round hit "Reveal map" so they SEE the trap they fell into, then "AI demo" to show pure gradient descent hit the same trap. Things to nudge them to try: LARGE steps in the canyon (ping-pongs between the walls) vs SMALL steps (runs out of budget on the flat floor) — there is no perfect answer, that's the learning-rate dilemma. On the plateau: "your feet feel… NOTHING. Which way now?" On fuzzy: "the ground lies — tiny pits everywhere." When someone is trapped in the local minimum, ask "what would YOU do if you suspected a deeper valley existed?" — someone will invent the 🎲 random jump on their own. Celebrate that: they just reinvented a real AI trick. If students have devices, the game is on the mini-games hub ("Find the Lowest Point") — they can walk their own fog while volunteers drive up front. Keep an eye on the clock: this slide is the session's flex. Cut at 2 levels if behind; the debrief slide next collects the lessons either way.

What did you discover?

Why finding the bottom is genuinely hard

🕳️ A cozy little valley trapped you — grown-up word #2: a local minimum. "Pretty good" blocks the road to "best"
🏜️ Flat ground = your feet feel nothing — no slope, no clue, no GPS
〽️ The canyon: falling in is easy — crawling along its almost-flat floor eats your whole step budget
🌫️ Bumpy ground lies — the little dip under your foot isn't always the real way down
📏 Big steps overshoot the bottom · tiny steps run out of budget — and no, there's no perfect step size (engineers argue about this DAILY)
🎲 Your best escape: jump somewhere random and retry — real AI training does exactly this. Professional randomness 😄

GPT trained exactly like this — feeling its way downhill in fog, on a map with a trillion knob-directions.

[30:00–32:00] Harvest THEIR discoveries first: "What made it hard? Shout it out!" Reveal each line only after a student names something close to it — the lines should feel like labels for things they already felt, not new material. If a discovery doesn't match any line, even better: "you found one I don't have a slide for!" Terminology: exactly ONE new grown-up word here — local minimum. Milk the metaphor: "a comfy couch of a valley. Not the best spot, but getting out means going UP first, and a downhill-only walker will never do it. AI's version of 'good enough, I guess.'" Everything else stays in kid-speak on purpose. If (and only if) someone asks for more real words, hand them out as bonus trivia: flat ground = "vanishing gradient," step size = "learning rate," the lucky jump = "random restart." Then return to the metaphors. The trillion punch, delivered slowly: "You walked a 2-knob map. GPT's map doesn't have just north and east — it has a TRILLION knob-directions at every point. Same fog. Same feet. Same trick. That this works at all is one of the wonders of our time." Close the loop to secret #3: guess→check→adjust IS this walk — every "adjust" is one downhill step.

The teacher matters

Data is the teacher — good or bad

Show it only black cats → it thinks all cats are black
Show it only sunny-day photos → it fails in the rain
Show it unfair examples → it learns the unfairness ("bias")
Rule: an AI is a mirror of its examples. Choose them carefully.

A strange truth

Hard for you ≠ hard for AI

🤖 Easy for AI, hard for you

Multiply huge numbers · search a billion pages · play perfect chess · never get bored

🧒 Easy for you, hard for AI

Tie a shoelace · get a joke · know a stack of chairs is climbable · tell when a friend is sad

Walking is harder than chess. That's why robots are the frontier.

The AI family

Three kinds of AI you'll meet this weekend

🏷️ Classifiers

Sort things into groups. "Cat or dog?" — what you just watched, and what you'll build in the lab.

✨ Generators

Create new stuff. "Write me a story" — this afternoon's ChatGPT session.

🤖 Agents

Predict & act in a world. "Push the ball into the goal" — tomorrow's world-models session.

After the break · 10:30

YOU train an AI — no code, 20 minutes 🔬

Tool: Teachable Machine (runs in the browser)
Pick classes: ✊ rock / ✋ paper / ✌️ scissors — or clap vs snap
Feed it examples from the webcam or mic — 30+ per class
Test it live — then try to BREAK it (weird angle, new person, messy background)
Breaking it is the best part — every failure teaches you what data was missing.

Recap

Six things you now know

1️⃣ AI learns from examples, not recipes
2️⃣ To a computer, everything is numbers
3️⃣ Learning = drawing a boundary between groups
4️⃣ It improves by guess → check → adjust — nudging its knobs 🎛️
5️⃣ Adjusting = hiking downhill in fog — traps and all (you lived it!)
6️⃣ The data is the teacher — bias in, bias out

Questions? 🙋

Then: break — and at 10:30, you train your first AI.

Mini-games hub for later: ask a TA for the QR · Teach the Machine, Train the Line & Find the Lowest Point are all on it