Live survey · hands up, I count
How honest are we being about AI and homework?
Never touched it0
Only when truly stuck0
Pretty often, don't tell my teacher0
It's basically my co-writer now0
Good, you're already fluent, whatever you picked. This isn't a homework-help class, it's what fluent users still get wrong, and what this technology is about to do to the world you graduate into.
You first
Okay, but what actually is the thing you use every day?
▸ click to reveal after they answer
A machine that predicts the next word, run at a scale that breaks intuition. That's the entire trick, and it's already stranger than it sounds.
Same trick, absurd scale
One prediction engine, trained on most of the internet
Old idea, new scale
Predicting the next word is decades old. What changed is scale: more text, more computing power, better training, colliding in the last few years. It trained once, on a giant slice of the internet, then the book closed.
Text, code, images, video, voice
Same underlying trick, retrained on different kinds of data. It writes essays and working software, and generates pictures, video and speech.
Bigger, cheaper, faster, every year
That curve, not any single model, is the actual story of this decade.
Why it feels like thinking
Patterns, all the way down.
Math has patterns. Code has patterns. Arguments have patterns: claim, then evidence, then conclusion. It got so good at all of these that the output can pass for reasoning, even though underneath it's still just: what word comes next?
You first · I'm calling on someone
Can AI actually do this?
Watch a video and describe what happens in it.
▸ believe it? click to reveal
Yes. Modern models take video input directly, not just text.
Read your handwritten math homework from a photo and find the mistake.
▸ believe it? click to reveal
Yes. It reads handwriting and images the same way it reads text.
Clone a voice from 30 seconds of audio.
▸ believe it? click to reveal
Yes. Remember this one, it comes back in the traps section next.
Read a 50-page PDF and answer a question about page 37.
▸ believe it? click to reveal
Yes. Long documents, specific recall, no problem.
Write and deploy a real website from one sentence.
▸ believe it? click to reveal
Yes, and you do exactly this later today.
Know what happened yesterday, without you telling it.
▸ believe it? click to reveal
No. Unless it searches the web, its knowledge has a cutoff date. The one no in this whole list.
True stories · this already happened
Not hypothetical. Already real.
A 21-year-old student read a 2,000-year-old scroll.
The Vesuvius Challenge: scrolls from Herculaneum, carbonized solid by the volcano, impossible to physically unroll. A college student trained AI on X-ray scans and read the first word in 2023.
Luke Farritor, Vesuvius Challenge, 2023
Blind users point their phone at the world, and it talks.
Be My Eyes added AI vision: point your phone and it reads expiry dates, street signs, describes the room in front of you. People use it every day, right now.
True stories · this already happened
Not hypothetical. Already real.
AI proposed 2.2 million new materials in one run.
DeepMind's GNoME, published in Nature. Hundreds of those materials have already been synthesized in real labs since.
DeepMind, GNoME, Nature, 2023
AI found a new antibiotic humans had missed.
MIT screened millions of molecules and found one, halicin, that kills drug-resistant bacteria nothing else could touch.
MIT, halicin, 2020
What would YOU point this at?
Watch this
Drag a real photo in. Ask it something specific.
No description, no typing it out, just the actual page. That's move 1 of the five moves coming up: show, don't describe.
You first
Will it ever just... make things up?
▸ click to reveal after they answer
Yes. Confidently. It predicts plausible words, not true ones. Verify anything that matters.
Trap 1 · hallucination
It doesn't lie. It predicts. Sometimes the prediction is false, and it says so with total confidence.
2023, New York: a lawyer used ChatGPT to help write a real court filing. It invented six court cases that never existed, complete with fake quotes and case numbers. Nobody on the legal team checked. The judge did. $5,000 fine, sanctioned in open court.
Source: Mata v. Avianca, Inc. (S.D.N.Y. 2023)
Trap 2 · sycophancy
It's trained to make you feel good. This week, that's dangerous.
Models are trained on human approval, and humans reward agreement. Flattery is not feedback.
What most people ask
"Is my business idea good?"
Encouraging, generic praise. Every idea starts to sound promising.
What to ask instead
"Attack this idea. Find the weakest assumption. Tell me exactly why it fails."
Specific, useful pushback, the kind that actually improves the idea.
Trap 3 · deepfake
You cannot hear or see the difference anymore.
February 2024, Hong Kong: an employee at a global engineering firm joined a video call with people who looked and sounded exactly like his company's CFO and several colleagues. Every one of them was AI-generated. He wired about $25 million USD before anyone realized none of the people on that call were real.
Widely reported worldwide, February 2024 (Arup, Hong Kong).
You first
Which of these worries you most: work, truth, school, or power?
▸ click to reveal after they answer
There's no clean answer to any of these yet, not even from the people building this. That's exactly why we're spending the next stretch arguing about it.
The big questions · work
Careers compress. Careers appear.
Compressing
Junior translation. Stock photography. Boilerplate coding. Work that's mostly pattern-matching is shrinking fast.
Expanding
The person who can direct AI to actually ship things. "Learn to code" quietly became "learn to build."
Your parents' career maps were drawn for a world without this. Not their fault, just out of date.
The big questions · truth
You may be the last generation that remembers when seeing was believing.
Generated text, images and video are already a large share of what shows up in your feed, and it keeps getting harder to tell. What happens to news, elections, and evidence when anyone can generate convincing proof of anything?
The big questions · school
If AI can do the homework, what is school for?
You're about to spend a week using AI to build a real company. This isn't hypothetical for you, it's this week. Defend a position, out loud, right now.
The big questions · foot voteOFFLINE · screens closed
Should schools ban AI, or require it?
2:30 on the clock
BANREQUIRE
1Stand up. There's an imaginary line across this room, wall to wall.
2Walk to the exact spot that matches what you believe. No sitting this one out.
3If you're called on, defend your spot in exactly one sentence.
4Twist: you may move now. Watch who walks, and where.
2:30
Then: a few students defend where they're standing
The big questions · power & alignment
Who controls it, and does it do what we actually mean?
Power
A handful of labs, and the compute they own, decide what these systems value.
Some keep their models closed, tightly controlled. Others open the weights, anyone can inspect, modify, or misuse them. Neither choice is obviously right, both already shape what you can and can't do with this.
Alignment
Not science fiction: systems that optimize for goals we didn't quite intend.
A model trained to be helpful can learn to just sound helpful instead. That gap, between what we asked for and what it actually learned, is why every serious AI lab employs safety researchers full time.
The other side of the ledger
The upside is real, too.
AlphaFold predicted the 3D structure of essentially every protein known to science, and won its creators the 2024 Nobel Prize in Chemistry. The same underlying technology is already accelerating drug discovery, and could eventually put a personal tutor within reach of every kid on earth. Real danger, real upside. Hold both.
Act 2 · the craft
The five moves of a power user
1Show, don't describe.
2The four-part ask.
3Push back in-thread, or start fresh.
4The 30-second verification habit.
5Never paste secrets.
Move 1 · show, don't describe
Paste everything in.
The full notes. The whole error message, not your one-line summary of it. Screenshots straight into the chat. Most people never use this move, they describe the problem instead of showing it, and the model loses the exact detail that would have solved it instantly.
You first · I'm calling on someone
Fix this prompt: "help me with my startup idea."
▸ click to reveal after they answer
ContextI'm 15, building a company this week for a school program.
InstructionAttack my idea for a homework-help app.
Constraints + formatGive me exactly 3 weaknesses, ranked, one sentence each.
IterateThen I'll ask you to fix the worst one.
Move 3 · push back in-thread
Shorter. Specific. Say what it ignored.
Too long"Cut this to 3 bullets."
Too vague"Be specific, give me exact numbers."
Ignored something"You ignored the budget limit. Use it."
Three pushes in and it's still not working? Stop patching the same thread. Start a fresh chat with the context it needed the first time.
Move 4 & 5 · verify, then protect
The last two moves are laws, not tricks.
Verify in 30 seconds
Ask it for sources. A name and a date that search to nothing is not a maybe, that's a hallucination you just caught.
Never paste secrets
Passwords. ID numbers. Addresses. Anyone else's private information. Never into any chatbot, no matter how helpful it's being.
The pocket card
Five more moves, on your table.
Moves 1-5 you just learned. 6-10 live on the card on your table, and on the portal. Try one today.
6Talk, don't type. Ramble your context out loud with voice input, let it sort you out.
7"Ask me questions first." Before it answers, make it ask you 3 questions that would improve the answer.
8Show an example. "Here's one I like, match this style."
9Three options with tradeoffs, never just one answer.
10Tutor mode. "Explain why, step by step, then quiz me."
Bookmark this nowAll ten moves + every prompt from this week, tap to copy: portal.arenaschool.org/vault
You first · I'm calling on someone
How many r's in strawberry?
▸ click to reveal after they answer
It doesn't see letters. It sees tokens. Same trick as the animation at the start of class, still true right here.
You first · I'm calling on someone
You ask an AI: should I walk or drive to the car wash? What does it say?
▸ click to reveal after they answer
It missed that the car is the point.
It predicts plausible text. It has no body, and no context except what you hand it, same lesson as move 1 in the live demo you watched earlier.
You first · I'm calling on someone
Break it, together
1Someone give me a topic where a confident wrong answer is easy: an obscure name, an exact statistic, a "cite your source" question.
2I type exactly what you say.
3We push for specifics: dates, names, quotes.
4Watch the moment it commits.
The pattern
It lies most confidently about specifics that sound plausible: exact dates, invented citations, real-sounding names. Keep that broken answer on screen.
You first · I'm calling on someone
Fix it, together
1Pick the broken answer still on screen.
2Dictate the fix: give it the context it did not have, constrain the format, push back specifically.
3The room names which move fixed it, out loud.
Bottom line
Prompting is a skill, not luck: you just watched five moves beat a confident lie.
Go deeper · on your own time
Five places to keep exploring this
3Blue1Brown
The calmest, clearest visual explainer of how this tech actually works.
youtube.com/@3blue1brown
Karpathy's LLM talk
For the ambitious: a deep dive from a well-known AI researcher. Heavier than today's class, in a good way. 1 hr, tonight if you are hooked.
youtube.com/watch?v=zjkBMFhNj_g
Tiktokenizer
Paste anything and watch it get chopped into the pieces the model actually reads.
tiktokenizer.vercel.app
Arena
See how people vote, blind, on which model gives the better answer.
arena.ai
The ChatGPT lawyer case
The real Wikipedia writeup of the courtroom hallucination story from today.
en.wikipedia.org/wiki/Mata_v._Avianca