Teaching "Train Your Own Model" to Builder mode (ages 8–10)

Hook & Warm-Up

Open with the lesson's real hook line: "Time to build a REAL AI! Collect data, train a model, test it, see it predict. This is exactly what professional AI developers do!"

Ask the class: "Who's used an app where you show it a photo and it tells you what's in it — a photo app that finds your pet, a game that scans a card, a filter that only works on faces?" Let a few students share examples. Say: "All of those started exactly the way ours will today — someone showed a computer LOTS of labeled examples until it learned the pattern."

Write four words on the board: Data 📊, Train 🧠, Test 🧪, Deploy 🚀 — the same four the app shows when this lesson opens. Say: "By the end of today, you'll have done all four of these steps yourself, for real, with your own AI model that you built."

Quick prediction question before opening the app: "If I want to build an AI that can tell my two hands apart from a photo — left hand vs. right hand — what do you think I need to give it first?" Take guesses. Guide toward: photos of both, clearly labeled. Say: "Let's go build one and see if you're right."

One more warm-up before opening the app: ask the class, "Think about the last time you got really good at something after practicing — a video game level, a sport, a musical instrument. Did you get good after doing it once, or after doing it many times, noticing what worked and what didn't?" Connect it directly: "AI models get better the exact same way — not by being told the answer once, but by seeing lots of examples and gradually noticing the pattern. Let's build one and watch it happen."

Main Activity

Open the AI Playground with students in pairs (one holds the object or pose, the other operates the device) and walk through the lesson's three scenes as an actual build, not just a demonstration.

Before pairs start capturing anything, mention a privacy point worth being upfront about: "Your camera photos stay right here on this device while we work — they aren't uploaded or saved anywhere else, and they disappear once you close the tab. That's on purpose, so nothing from this activity ends up stored somewhere none of us can see."

Scene 1: Collect Data

Have each pair rename the app's two categories to something they'll sort — the default Thumbs Up/Thumbs Down works fine, or let pairs pick two classroom objects (a pencil vs. a marker, for example). Say: "Your job right now is to be a good data collector. Take pictures from different angles, different distances, and even different lighting if you can — a computer that only ever saw one angle will get confused the moment it sees a new one."

If pairs choose their own two objects rather than the defaults, spend a minute helping them pick a genuinely fair pair — two objects that are visually distinct enough to be learnable in one class period (a red marker vs. a green marker is harder for a beginner model than a marker vs. a pair of scissors). This mirrors a real decision AI teams make early: how hard a task is depends heavily on how visually different the categories actually are, not just on how good the model is.

Introduce labeling directly: "Every picture you capture gets tagged with a label — the name of its category. Without labels, the computer just has a pile of pictures with no idea what to call any of them." Then introduce balancing: "If you take 30 pictures of one thing and only 5 of the other, what do you think the computer will guess most of the time?" (Guide toward: the more common category, because that's mostly what it's seen — just like the app's quiz.) Have each pair check their own counts and even them out before moving on. Aim for at least 10-15 examples per category — more if time allows, since the app's guidance ("100+ examples per category") describes an ideal for production-quality models; a classroom demo will already show the pattern clearly with far fewer.

Push pairs to be deliberate about variety, not just quantity: have them capture a few examples holding the object close to the camera, a few farther away, a few tilted at an angle, and — if lighting allows — a few near a window and a few away from it. Ask: "If every single one of your pictures looked almost identical, would the computer actually be learning what your object looks like in general, or just memorizing that one specific photo, taken over and over?" Most pairs will correctly guess the second, which sets up the overfitting-style thinking they'll build on in later lessons.

Scene 2: Train the Model

Have pairs tap the button that trains their model. Say: "Watch what's happening — the app is comparing every new picture that comes in to all the labeled examples you gave it, and picking whichever group it looks most similar to. That's genuinely what's running behind the scenes, using a technique real engineers call transfer learning — it's borrowing an already-smart piece of vision the app comes with, instead of learning to see from zero."

Note for accuracy: this app trains almost instantly rather than showing a slow climbing accuracy number, because of that shortcut — worth saying out loud so students aren't confused, especially if they've seen slower "training" demos elsewhere. Ask: "If the app is comparing new photos to your examples, what happens if your examples were all blurry or all from one angle?" (Guide toward: the comparisons get less reliable, since there's less to compare against.)

Give pairs a concrete way to picture "borrowing" a piece of vision: "Imagine you already knew how to sort a huge pile of shapes by edges, corners, and colors, from years of practice. Now someone hands you two new small piles and just asks, 'sort these the same way, but into THESE two groups instead.' You wouldn't need to relearn what edges and corners are — you'd just apply that skill to a new, smaller sorting job. That's exactly what our app is doing: it already knows how to notice visual patterns in general, and it's just applying that skill to your specific two categories."

Scene 3: Test & Improve

Have pairs swap devices with a neighboring pair and test each other's models with a brand-new example neither pair used in training. Say: "This is the only fair test — testing with a picture you already trained on doesn't prove anything, because the computer basically already 'knows the answer' for that exact photo."

When a model gets something wrong, walk pairs through failure analysis out loud: "Don't just shrug — ask why. Was the lighting different? Was the angle new? Was that category under-represented?" Have pairs add 2-3 more targeted examples addressing the likely cause, retrain, and retest. Close the activity by having each pair report their model's accuracy before and after that one round of fixing.

Circulate while pairs test and retest, and listen for a common shortcut worth catching: students sometimes want to test with a photo they already used for training, because it's convenient and "always works." If you spot this, stop the pair and ask: "Does that prove your model learned the pattern, or does it just prove the model remembers this one exact photo?" Have them explain the difference in their own words before moving on — this distinction matters enough that it's worth pausing the whole class briefly if several pairs make the same mistake.

Discussion

• What was the hardest part of getting your model to work well — was it collecting data, training, or testing? Why?
• Why does testing with a picture the computer never saw during training matter so much? What would testing with training photos actually prove?
• If your model kept mixing up two categories, what would you try changing first, and why?
• The app's most important step, according to its own quiz, is having good, diverse, labeled data — not a fast computer. Does that match what you noticed while building your own model?
• Can you think of a real app or tool that probably had to go through this exact same collect-train-test-fix cycle before it worked well?
• Your model borrowed an already-trained piece of vision instead of learning to see from scratch. Why do you think that shortcut makes it possible to build something like this in a single class period, instead of needing days or weeks?
• If two different pairs trained a thumbs-up/thumbs-down model but ended up with different accuracy, what's more likely to explain the difference — the app itself, or how each pair collected their examples? What evidence from today's activity supports your answer?

Quiz Walkthrough

Wrap-Up & Extension

Close with: "Today you did the exact same four steps — collect, train, test, deploy — that professional AI developers do every day. The only difference is scale: they might use millions of photos and huge computers; you used a few dozen photos and a browser tab. The process is genuinely the same."

Take one more minute to connect it back to the "deploy" step specifically, since it's easy to skip past: "Right now, your model only exists on this one device, in this one browser tab. If a real company wanted to use a model like yours in an actual app, 'deploying' it would mean putting it somewhere every user's phone or computer could reach it — that's the step we're not doing today, but it's worth knowing it exists."

Extension activity — Three-Category Challenge: Have pairs add a third category to their model (for example, adding "peace sign" alongside thumbs up/down, or a third classroom object) and predict, before training, whether accuracy will go up or down with an extra category to tell apart. After training and testing, have pairs report what actually happened and why they think so — most groups will notice that more categories usually means more chances for confusion unless each one gets enough clear, varied examples, reinforcing the balance-and-quantity lesson from the main activity.

If time allows, push it further with a class-wide leaderboard: have every pair test their finished three-category model against the same set of five surprise photos you show on the board or projector (ones you took yourself, not shared with any pair beforehand), and tally how many each model gets right. Ask afterward what the highest-scoring pairs seemed to have in common — most groups will land on "they collected more varied, better-labeled, well-balanced examples," tying the whole activity back to the lesson's central point about what actually makes a model reliable.