ISB AI Founders
Day 1 · 9:20

AI Foundations

ISB AI FoundersS1
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
You first

Okay, but what actually is the thing you just voted on?

▸ click to reveal after they answer
A machine that predicts the next word. That is the whole trick, and it changes everything.
How it works · live demo

Words in. Prediction out. Billions of times.

Your words
"the best snack in the ISB cafeteria is"
→
The model
billions of patterns, one guess at a time
→
Predicted next words
"...the sweet potato fries, obviously"
LIVE DEMO · no rehearsal
"The capital of France is..."
Typed live, finished live, no do over.
You do
Coming in the lab block, minutes away: you run this exact demo yourself.

For now, watch, and decide which unfinished sentence YOU will try.

Try one oddly specific to you: your street, your pet, your favorite snack.
Watch it think

One word at a time. Live.

"The best snack in the ISB cafeteria is"
It weighs every candidate, picks one, then does it again for the next word. That is the entire trick.
90 seconds live · full cut at home · optional
NOT REQUIRED TONIGHT

Want the real mechanism, not just the metaphor?

3Blue1Brown · "Transformers, the tech behind LLMs" · about 27 minutes, watch it at home.
Site walkthrough · watch it happen

Before it can predict a word, it has to chop your sentence into pieces.

tiktokenizer.vercel.app
the best snack in the isb cafeteria is
the best snack in the isb cafeteria is
8 tokens
Real, free site: tiktokenizer.vercel.app. Type any sentence, watch it split into pieces, that's what the model actually reads.
Where it learned this

It read almost everything, then closed the book.

Training happened once, on a giant pile of text: books, articles, code, forum posts, homework help sites. Then the book closed. Now it works from memory of patterns, not a live lookup.
You first

So why does it feel like it's thinking?

▸ click to reveal after they answer
Because language itself is patterned. Grammar, logic, even good arguments show up as patterns in text, so predicting the next word starts to look a lot like thinking. It isn't. It's a very convincing shadow of it.
The trick, unpacked

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?
A quick history

Old idea. Suddenly a thousand times bigger.

Predicting the next word is decades old in computer science. What changed is scale: more text, more computing power, better training, all colliding in the last few years. Same trick. About a thousand times the size.
Same idea, a few years apart

Feel the difference scale makes

A small model, 2019-ish
the dog was very running to the store yesterday for purple and also happy
Today
Here's a 3-step plan for tonight: review the periodic table trends, do five practice stoichiometry problems, then quiz yourself on the vocab list.
What it is genuinely great at

The superpowers

Language
Write, rewrite, translate, summarize, any tone
Code
From English description to working software
Thinking with you
Brainstorm, critique, roleplay, explain
Site walkthrough · it's not just Claude

Dozens of models exist. People rank them by testing them blind.

arena.ai
Two answers, no names shown. Which is better?
Model A's answer...
Model B's answer...
Leaderboard
1. Claude1450
2. GPT1440
3. Gemini1432
Real, free site: arena.ai. Rankings shift every week, nobody's model is permanently "the best."
Superpower 1 · live demo

Say it in your voice. Or anyone else's.

LIVE DEMO · no rehearsal
"Make this awkward teacher text polite and professional."
One real message, rewritten live, no do over.
You do
Paste a real message you'd actually send.

Ask for a different tone: formal, casual, funnier.

Same meaning, new voice, instantly.
Superpower 2 · live demo

Describe it in English. Get working code.

LIVE DEMO · no rehearsal
"make a button that counts how many times I click it"
Built and shipped in front of you, live. Anything can happen.
You do
Describe one tiny app idea in plain English.

No programming language required to start.

Watch it turn your words into something that runs.
Superpower 3 · live demo

A debate partner who never gets tired or offended.

LIVE DEMO · no rehearsal
"Argue why I deserve MORE screen time. Now argue the parent's side."
Both sides, live, no script, whichever one wins the room.
You do
Ask it to argue the opposite of something you believe.

Or have it poke holes in your own plan.

It will. And it will not take it personally.
Before you touch it again

Everything you just saw has a catch.

The honest part

Now for where it lies to you.

Not maliciously. Structurally. And you need to know exactly how.
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.
The technical term

This is called a hallucination.

Not lying on purpose, there's no intent behind it. It's predicting the next plausible-sounding word, and sometimes plausible is not the same thing as true.
Exhibit A · funny, live demo

Ask about something that doesn't exist. Watch it commit.

LIVE DEMO · no rehearsal
"What is National Dumpling Appreciation Day?"
Total confidence, zero truth. Watch it commit to something fake.
You do
Ask about a fake holiday or event you invent.

Watch it never say "that's not real."

Notice: total confidence, zero truth.
Exhibit B · serious

Now imagine this inside your homework.

Ask for a quote from a historical figure, a scientific statistic, or a book citation. It can invent a quote that was never said, attach a real name to it, and format it like a footnote. It looks exactly as trustworthy as the true ones.
Exhibit C · this actually happened

A real lawyer got fined $5,000 for this.

2023, New York. A lawyer used ChatGPT to help write a real court filing. It invented six court cases that never happened, complete with fake quotes and case numbers. Nobody checked. The judge did. The lawyer got sanctioned, in front of a packed courtroom.
Source: Mata v. Avianca, Inc. (S.D.N.Y. 2023)
Same confident voice, different truth

You cannot hear the difference

Made up
The Great Panda Uprising of 1987 ended when local farmers negotiated a truce with the bamboo councils.
Real
World War Two ended in 1945, with Japan's surrender in September following the atomic bombings in August.
Hold both at once

Astonishingly capable. Confidently wrong.

Your job is not to trust it or distrust it. Your job is to DIRECT it and VERIFY it.
So what do you do about it?

You get precise. Every single time.

The fix for a bad AI answer is almost never a smarter AI. It's a better question.
The skill

Bad answers usually mean bad prompts.

The model can only work with what you give it. Four parts turn a wish into an instruction.
Prompt anatomy

Four layers, stacked

Context
"Grade 8, never taken biology" changes the whole answer.
↓
Instruction
"Cut to 400 words, keep the thesis" beats "help me with this."
↓
Format
"As five flashcards." Don't ask, get a wall of text.
↓
Iterate
"Too long, cut it in half." The first answer is a draft.
↓
A precise, usable answer
You first

Which of these four do people skip the most?

▸ click to reveal after they answer
Format. Everyone states the task. Almost nobody says what the output should look like, then complains it's too long.
Same task, two prompts · studying

Feel the difference

The lazy one
help me with my science homework
The strong one
I am a Grade 9 student with a physics test on Newton's laws Friday. Make me a 10-question practice quiz, mixed difficulty, answers hidden at the end. Ask me one question at a time and grade my answers.
Same task, two prompts · planning

Feel the difference

The lazy one
help me plan my week
The strong one
I'm in Grade 10 with three tests next week: biology Tuesday, math Wednesday, history Friday, plus soccer practice Monday, Wednesday and Friday 4 to 6pm. Build me an hour-by-hour study schedule for the next 5 days that fits around practice, includes 15-minute breaks, and puts the hardest subject first each day.
Same task, two prompts · creative

Feel the difference

The lazy one
write me a poem
The strong one
Write a 12-line poem in the voice of a nervous new student on their first day at ISB. Free verse, no rhyming, include one concrete sensory detail about the school hallway.
Your turn

Time to feel the gap yourself.

Exercise 1 · why this mattersFOLLOW ALONG · laptops open

You won't believe it until you feel it.

Everyone in this room just watched three comparisons on a screen. That's not the same as feeling the gap yourself, on your own task, in your own hands. You're about to run both prompts, back to back, right now.
Exercise 1 · debriefFOLLOW ALONG · laptops open

What to listen for

When pairs share, listen for the exact moment the answer got useful, usually right after context or format got added, not instruction. Nobody's lazy prompt was "wrong." It was just vague.
Exercise 2 · why this mattersFOLLOW ALONG · laptops open

This prompt, you keep.

Exercise 1 was practice. This one is real: you're building a prompt that does actual recurring work for you, something you'll genuinely use again this month, not just today.
Exercise 2 · debriefFOLLOW ALONG · laptops open

What to listen for

The best prompts shared here are usually boring and useful: a homework schedule generator, a flashcard maker, a WeChat-message-to-parents translator. Boring and reused beats clever and used once.
Act 2 · The lab

Lids up. Everything you just watched, you now do.

Mission 0 FOLLOW ALONG · laptops open

Get in. Before anything else.

3:00
1Open a browser. Go to claude.ai
2Log in with the account card on your table.
3See the chat box? Hand up. We do not start until every hand is up.
3:00
Every screen in. Then we play.
Exercise 1 · do it nowFOLLOW ALONG · laptops open

The prompt battle

5:00 on the clock
1Pick any real task from your life (study, plan, write, decide).
2Run the lazy one-line prompt. Screenshot the answer.
3Rebuild it with all four parts. Run it again. Already good at this? Harder target: make Claude quiz you Socratically, one question at a time, adapting to your wrong answers.
4Compare: what changed?
5:00
Then: post your before/after to CLASS CHANNEL
Exercise 2 · do it nowFOLLOW ALONG · laptops open

One prompt that works for YOU

5:00 on the clock
1Build one prompt that does real, recurring work for you personally.
2Test it. Refine it once.
3Save it somewhere you will find it again.
5:00
Then: share the best with CLASS CHANNEL
Copy this prompt · run it now

Stuck? Let Claude teach you

You are helping me learn to prompt well. I will give you a task. First give me the mediocre one-line prompt most people would write, then the strong version with context, constraints and output format, then explain the difference in two sentences. Task: [X].
Checkpoint

Four parts. One real prompt. You already did it.

Context, instruction, format, iterate. Name them cold, and you have a tool that works for every session this week.
“Make something people want.” Today you proved it.
One more thing, live

Let us make a song about this room. Right now.

Sixty seconds, one prompt, a real song. By Friday you will make one of these for your own company. Same skill: telling the AI exactly what you want.
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 1-hour deep dive from a well-known AI researcher. Heavier than today's class, in a good way.
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
Where this goes

Today it answered you. By Thursday it builds for you.

Same skill, bigger stakes: everything this week runs on precise instructions.