✅ Quiz 16 — AI & Machine Learning
Answer in your head or on paper, then open each fold. If you get one wrong, that’s the quiz working — reread that section over coffee.
1. Fill in the arrows. Traditional programming: ___ in, ___ out. Machine learning: ___ in, ___ out.
Show answer
Traditional programming: **rules** in, **answers** out (humans write the rules). Machine learning: **examples** in, **rules** out (the machine finds the patterns itself). The spam filter learned from two million labeled emails instead of 10,000 hand-written if-rules.2. Why did hand-written rules lose the spam war?
Show answer
Spammers adapt faster than humans can write rules. Every new rule ("block 'FREE MONEY'") got dodged ("FR3E M0NEY"), and rule piles started catching real email too. A model trained on millions of examples finds subtler patterns and can be retrained as spam evolves.3. Training vs. using: which one costs millions, and what does the exam analogy say about chatting with an AI?
Show answer
Training costs millions — months of computing to tune the model. Using it (inference) is fast and cheap-ish. The exam analogy: when you chat with an AI, it's *taking the exam*, not studying — no new learning happens during your conversation, and the model stays frozen until its makers retrain it. That's also why it doesn't know last week's news.4. What is a model made of, and what’s honestly strange about the knobs?
Show answer
A model is millions or billions of adjustable numbers — "knobs" on a giant mixing desk — tuned automatically during training until answers stop being wrong. The strange part: no human set them, and often nobody — including the creators — can say what any single knob means. That's why AI can be hard to explain and audit ("black box").5. In an image-recognizing neural network, what do the layers do — and is it a brain?
Show answer
Each layer spots patterns in the previous layer's output, stacking simple into sophisticated: edges → shapes → parts (eyes, wheels) → whole things (faces, cars). It's *loosely inspired* by neurons but it is not a brain — it's arithmetic in layers, a paper airplane to the brain's falcon.6. A large language model has one core trick. What is it, and why does that humble trick produce such impressive results?
Show answer
Predicting the next word. Done at colossal scale — trained on a mountain of human text with billions of knobs — predicting the next word *well* forced the model to absorb grammar, facts, styles, and the shapes of arguments. Stack enough good next-word predictions and you get essays, code, and poetry.7. Define “hallucination,” and explain why an LLM’s hallucinations arrive with no warning signs.
Show answer
A hallucination is output that is fluent, confident, and false — an invented paper, a wrong date, a fake court case. There's no warning because the model's only compass is "what sounds plausible next" — truth isn't part of the mechanism — so falsehoods wear the exact same confident tone as facts. Like the friend who read the whole library but never says "I'm not sure."8. Your colleague pastes an LLM’s answer — including two statistics and a citation — straight into a client report. What’s the rule they broke, and what should they do instead?
Show answer
Never trust an LLM for facts you can't verify. Statistics, names, dates, and citations are exactly where hallucinations hide. The fix: use the LLM for the draft (its strength), then verify every checkable claim against a real source before it ships — brilliant intern, unread work, never straight to the client.9. The hiring-model story: what happened, and what’s the general moral?
Show answer
A company trained a hiring model on ten years of its own decisions — which skewed male — so the model learned to downgrade CVs mentioning women's activities and colleges. It was scrapped. The moral: garbage in, gospel out — a model trained on a biased history will faithfully repeat that history, with a machine's false authority.10. Sort these into “exists today” vs. “does not exist today”: (a) software that spots tumors in scans, (b) AGI — AI as flexible as a human across all tasks, (c) photo search that finds “dog”, (d) an AI that knows when it’s wrong.
Show answer
Exists today: (a) and (c) — narrow pattern-recognition is AI's home turf. Does not exist today: (b) AGI is hypothetical; and (d) — today's models have no reliable sense of their own wrongness, which is exactly why hallucinations come out confident.11. Name three pieces of everyday machine learning you probably used this week without calling it “AI”.
Show answer
Any three of: face unlock, photo search finding "dog", maps predicting arrival time, autocorrect/autocomplete, spam filtering, and every recommendation feed ("the algorithm™") — music, video, shopping. ML has been in your pocket for years; only the chatbots made it famous.12. 🗣️ Explain it to a friend. Your uncle says: “I read that ChatGPT knows everything — and my neighbor says it lies constantly. Which is it?” Give him a two-minute answer using at least one analogy from this module. Cover: what an LLM actually does, why it’s both brilliant and unreliable, and one rule for using it safely.