AIQ
AIQ
Part of the Learning from Mistakes lesson guide. Teaching a different grade? 🌈 Explorer (5–7) · 💻 Hacker (11–14) · ⚡ Architect (15–18)
Ask: "How did you learn to ride a bike? Did someone hand you a manual that explained exactly how to balance?" Let a few students answer — most will say a parent or sibling held the seat, they wobbled, they fell, and they tried again.
"How did you learn to ride a bike? Not from a textbook — you tried, fell, tried again! Reinforcement learning works the same way!"
Write "reinforcement learning" on the board and say: "That's a big phrase for something you already understand from riding a bike. Today we're going to see how AI uses that exact same strategy — trying, failing, and trying again — to get really good at things, sometimes better than any human." Open the lesson on the app and preview the four hook icons: 🎮 Try, ❌ Fail, ✅ Succeed, 🏆 Master.
Add one more thread before opening the app: "In the last lesson, 'Learning from Examples,' the AI was shown a bunch of examples that were already labeled correct or incorrect — like flashcards with the answers on the back. Today's AI gets no flashcards at all. Nobody tells it the right move. It only finds out afterward whether what it tried worked out or not. That's a much harder problem — see if you can guess why as we go." This one framing question does a lot of work: it sets up the entire lesson as harder and different from the one before it, rather than a repeat with new words.
Work through the app's three scenes as a class, reading each fact aloud and pausing to check understanding before moving on.
Game AI: "An AI can play a game thousands of times. Each time, it tries slightly different moves and watches what scores points." Reward: "A good move earns a reward — like points. A bad move earns a penalty. Over time, the AI leans toward whatever earned it rewards before." Repeat: "It's common for an AI to play the SAME game a million times before it masters it — every single game teaches it something." Improve: "At the very start, it's basically guessing randomly. After thousands of rounds, it starts using strategies no human ever taught it." Ask: "Why do you think it needs so many tries, instead of just a few?" (Because early on, almost every action is random guessing — it takes a lot of repetition before patterns of what works start to emerge.)
Pause here for a quick check: have students turn to a partner and, in their own words, explain the difference between what an AI in "Learning from Examples" gets (labeled examples with the answer already given) and what an AI in "Learning from Mistakes" gets (only a reward or penalty after it acts, with no answer given up front). This is the single most important distinction in the whole lesson, and it's worth confirming out loud before moving on to the famous examples.
AlphaGo: "Google's AlphaGo learned the board game Go by playing millions of games against itself — and it beat the reigning world champion." Robot Walking: "Robots learn to walk by falling down thousands of times inside a computer simulation, then transferring what they learned to a real robot body." Self-Driving: "Self-driving AI earns rewards for smooth, safe driving and penalties for mistakes, so over time it learns safer routes and habits." Ask: "What do you think counts as a 'mistake' for a self-driving car? Is it only crashes, or could something smaller count too?" (Smaller things count too — braking too hard, following too closely, or a rough turn can all earn a small penalty, well before anything resembling a crash.) AlphaStar: "DeepMind's AlphaStar learned the video game StarCraft II by playing against itself, and reached a skill level higher than 99.5% of ranked human players." Ask the class: "What do AlphaGo and AlphaStar have in common?" (Both learned almost entirely by playing against copies of themselves, rather than being taught by humans.) If your class knows StarCraft II or a similar real-time strategy game, point out that it's a much messier problem than Go: there's no fixed board, players act at the same time instead of taking turns, and you can't always see what your opponent is doing — which makes AlphaStar's result even more impressive than it might first sound.
Worth adding one accurate nuance here rather than oversimplifying: AlphaGo's very first version actually got a head start by studying a large number of recorded human games before it began playing against itself. A later version, called AlphaGo Zero, skipped that step entirely and learned purely from self-play, starting from completely random moves — and ended up even stronger. That's a good detail to mention if a student asks "so did it copy humans or not?" — the honest answer is "at first, a little, but the strongest version didn't need to."
Introduce the four vocabulary words the app uses, one at a time, tying each to the examples already discussed: Agent — "the learner itself; in our examples, that's AlphaGo or the walking robot." Environment — "whatever the agent is acting inside: the Go board, the road, the physics simulation." Reward Signal — "the +1 for good or −1 for bad that the environment hands back after every single action." Policy — "after enough practice, the agent builds a policy: a strategy that says 'in this situation, here's my best move.'" Have students say the loop back to you in order: agent takes an action in the environment, gets a reward, and slowly builds a policy.
Draw the loop as a simple diagram on the board — a box labeled "Agent" with an arrow labeled "action" pointing to a box labeled "Environment," and a second arrow labeled "reward" pointing back — and walk through it once more using the self-driving car: the car (agent) chooses to brake smoothly (action) inside traffic (environment), and gets a small positive reward for a comfortable stop. Ask students to fill in the same diagram in their notebooks using AlphaGo instead: what's the agent, what's the action, what's the environment, and what's the reward? (Agent: AlphaGo. Action: placing a stone. Environment: the Go board and the opposing player. Reward: winning or losing the game.)
Close by reading the app's summary together: "Reinforcement learning is learn-by-doing! AI tries actions, gets rewards or penalties, and builds a winning strategy."
Close with: "Every time you try something new, fail, and try again — a video game level, a sport, a math problem — you're doing your own version of reinforcement learning. The difference is AI can do it millions of times in a row without getting tired or frustrated."
Extension activity — Design a Reward (15 minutes): In pairs, have students invent a simple game an AI could learn (for example: a maze, a simple obstacle course, or a jeepney trying to reach a destination without hitting traffic). Have them write down one action the AI could take, one thing that should earn a reward, and one thing that should earn a penalty. Have a few pairs share their game, and ask the class: "Could the AI 'cheat' and get the reward without actually doing what you wanted?" This plants an early seed for the idea that a poorly designed reward can backfire — a theme that comes back in later lessons.
To stretch the activity further, have each pair swap their reward design with another pair, and challenge that pair to find a "loophole" — some way an imaginary AI could rack up rewards without actually solving the maze, obstacle course, or jeepney route. This turns the exercise into a small game of "break the reward," which is a real and important part of how AI engineers actually test reward designs before training a real system on them, and gives every student a hands-on feel for why designing a good reward is much harder than it first sounds.
Close the whole session by connecting back to the four hook icons from the start: "Try, Fail, Succeed, Master — that's the loop for the AI in today's lesson, and it's honestly the loop for getting good at almost anything. The difference between you and the AI isn't the loop itself. It's that the AI can run through that loop millions of times without ever getting discouraged, while you have to keep choosing to try again yourself."