Not because they're smarter than us. Because they don't have to decide.
A car can go 200 kmh
Nobody says the car is a great athlete. You turn the key, press the pedal, the engine does the rest. It's just a machine doing the thing machines are good at.
A computer can snap · clap · stomp at 800 BPM
The metronome you just raced doesn't think, doesn't get tired, doesn't second-guess. It's following a recipe — one beat at a time, forever, in exactly the right order.
When every step is written down ahead of time, a computer will beat us every time. That's what computers have always been good at: mechanical, rule-based work. No decisions required.
But what happens when the steps aren't written down? When someone asks a fuzzy question, or the situation keeps changing?
05
Meet ELIZA — the world's first chatbot
Type something and she'll talk back. ELIZA turns 60 in 2026 🎉
YOU:
Built in 1966 at MIT. The whole program is just a few hundred lines of code. See the original →
06
Why people fall for chatbots
When something talks back, our brains assume it understands us — even when it's just matching patterns. A few examples:
"AI" becomes embarrassing. Scientists hide their work under new names like machine learning and data mining.
Same story, twice in a row: people get excited → money rushes in → AI doesn't work as well as promised → money runs away.
09a brief history of vibes
Boom · bust · boom · bust · boom · 🚀
Seventy years of AI hype, plotted.
Two winters happened. Now we're in the biggest boom yet — and this time, the AI actually works. So: are we about to crash again? Or is this one different? What do you think? 🤔
10Puzzle
Connect all 9 dots with 4 straight lines
Don't lift your pen. Don't retrace a line. Can you do it in four?
Rules
Four straight lines only.
One continuous path — each new line starts where the last one ended.
Nothing in the rules said the lines had to stay inside the 3×3 grid.
The trick
Start at the bottom-left dot.
① straight up, past the top row.
② diagonal down through the middle, past the right edge.
③ left across the bottom row, back to the start.
④ diagonal up through the middle, past the top-right.
The dashed square is the box nobody told you was there.
This is where the phrase "think outside the box" comes from. The solution was always allowed — you just couldn't see it if you assumed walls that weren't in the rules.
Yes — 1,760 PS3s, wired together. That's what you just saw.
1,760
PlayStation 3s, stacked in Rome, New York
#33
One of the 35 fastest supercomputers in the world
$2M
Total cost — about 10× cheaper than a "real" supercomputer
The PS3 had a special chip that was great at doing lots of math at once. The Air Force used their giant PlayStation pile for processing radar images, AI research, and high-def video. Then Sony pushed an update that took away the Linux hack. The Air Force never bought another PS3.
A classic outside-the-box move: the cheapest super-fast computer chips on Earth were built for video games. The Air Force was the first to swoop in and use them for real work.
The task: look at a picture and name what's in it.
What happened: their AI got it wrong only 15% of the time. The best other team got it wrong 26% of the time. Not close. Almost overnight, every AI lab in the world started using gaming cards.
That's the moment AI took off for real. Every smart chatbot, photo filter, and self-driving car you've heard of traces back to this bedroom project.
Eight scientists at Google publish a paper. The whole AI world changes.
The paper introduced the Transformer — a totally new way for computers to read and understand language. Every famous AI today — ChatGPT, Claude, Gemini — is built on this idea. It's probably the most important computer science paper of the last 20 years — it's the engine that drives every chat app you use.
And the funniest part? Almost nobody noticed at first. The paper was pitched as a way to translate languages a bit better and way cheaper to train — only 3.5 days on 8 GPUs. It took about a year for everyone to realize the same idea could power any kind of AI.
Before
AI read sentences one word at a time, like a slow reader.
It was bad at remembering things from earlier in the sentence.
Training a big one took forever.
After
Every word "looks at" every other word at the same time.
It can connect ideas across long paragraphs easily.
Training is way faster (especially on GPUs).
16Warm-up
One-Word Story
A group tells a story together — one word at a time.
How to play
Everyone stands in a circle.
Person 1 says one word to start the story.
Person 2 adds one word. Then Person 3. And so on around the circle.
No planning ahead. Say the first word that fits what came before.
Keep going until someone lands on a natural ending — say "The end."
Example
"Once" → "upon" → "a" → "dragon" → "ate" → "my" → "homework" → "and" → "burped." → "The end."
Each word only makes sense because of the words before it. You're choosing the most likely next word, given the context.
Before the AI can think about a sentence, every word gets turned into a list of numbers. Try it!
Each list of numbers is like a spot on a giant map of meanings. Words with similar meanings end up near each other on the map. The AI does all its thinking by moving these number-lists around.
Try typing unbelievable — see how it gets chopped up. Try The and the — they get different numbers (capital matters!). The real AI tokenizers below are way smarter than this one.
Text in, one token out, repeat. The loop is the magic.
Every reply you've ever gotten from ChatGPT works exactly like this — guessing one word at a time, over and over. A short answer like "Hello, how can I help you today?" takes about 10 – 15 trips through the loop.
19
How "attention" works
The AI is reading "The cat sat on the ___" and trying to guess the next word.
1 Look around
For each earlier word, the AI asks: "how useful are you for guessing what comes next?"
20the answer…
It picks "mat" ✨
It mixes meaning from the words that mattered most — and picks the most likely one.
2 Guess the word
Paying close attention to "sat" (something to sit on) and "cat" (cats sit on things), the AI ranks every possible next word and picks the one with the highest score: "mat". Then it adds "mat" to the sentence and starts the whole thing over for the next word.
21
So when does it stop?
The same way One-Word Story ended — with a special "The end."
The trick
The AI keeps picking the next word in a loop. One of the "words" it's allowed to pick is a special invisible one that just means "I'm done."
When the AI picks that word, the loop stops and the answer is shown to you. That's it — no timer, no word count, no human telling it to stop.
What it's called
AI researchers call it a stop token (or an end-of-sequence token). In GPT-style models it's literally written <|endoftext|>.
If you've written code, you've seen EOF — "end of file." Same idea: a special marker that means "nothing more after this."
You never see it. The chat app hides it and stops streaming new text.
Every sentence the AI gives you ends because it chose "The end." as the most likely next word. Same move you made in the warm-up — just with a fancier name.
22
See the AI thinking, live
Two amazing websites that show you what's happening inside an AI brain.
Google invented the Transformer. A startup called OpenAI bet everything on making it huge.
OpenAI's idea was simple: take that 2017 paper, and make it way bigger. Feed it the entire internet. Train it on thousands of GPUs for months. See what happens.
2018GPT-1 — 117 million little knobs to tune. Reads books to learn how language works. Pretty good!
2019GPT-2 — 13× bigger. Now it can write whole paragraphs that make sense. OpenAI says it's "too dangerous to release."
2020GPT-3 — another 117× bigger (175 billion knobs!). Now you can just tell it what to do in plain English and it figures it out.
2022InstructGPT — humans teach the AI to be helpful instead of just rambling. This becomes ChatGPT.
The big surprise: making it bigger just kept working. No fancy new ideas needed — just more computers and more data. Bigger AI = smarter AI.
What about today? Just making AI bigger doesn't help as much anymore — we're running out of internet to feed it. So now AI companies are teaching AI to "think longer" before answering. That hack is working too.
Want to track how big AI is getting? epoch.ai/trends keeps a running scoreboard.
24November 30, 2022
ChatGPT goes viral
OpenAI wraps a chat interface around a model they'd had for about two years — and suddenly millions of people try it.
1M
people sign up in 5 days
100M
in two months — the fastest-growing app in history
The Transformer is just the engine. ChatGPT — and every AI you use — is also made of tons of other code: a chat window, an account system, safety filters, memory of past conversations, a way to handle millions of people at once. The AI brain is one piece. Everything around it is what makes it a product — something regular people can actually open up and use.
GPT-3 was released in June 2020. ChatGPT launched November 30, 2022, powered by GPT-3.5.
25
How an AI assistant thinks
It's a four-step loop. The AI does it over and over until your task is done.
That's the secret. By itself, AI just talks. Give it a loop and some tools, and suddenly it can actually do stuff — like build this whole slideshow.
Loops inside loops. Each of those four steps is the Transformer running — and inside the Transformer, it's predicting one word at a time (the loop we saw earlier). So when an AI assistant fixes a bug for you, it's running thousands of tiny prediction loops to do one big task.
26Activity
Let's do that loop — with a marker
Three volunteers at the front. The same four-step loop, but humans play the parts.
Roles
① Requester — names one thing to draw. e.g. "a busy street."
② Planner — breaks it down, checks with the Requester, then directs.
③ Drawer — holds the marker. Only draws what the Planner says.
The loop
Requester asks.
Planner lists up to 5 parts + colors.
Planner asks: "Requester, look good?" → tweak if not.
Planner directs the Drawer, one step at a time, until done.
That's prompting. The Planner is the prompt. The Drawer is the AI — fast, literal, bad at guessing what you meant.
27
Pick your AI assistant
There are lots of them. Today we're using Claude Code.
Watch what happens. Claude will read the game's code, plan a change, write the change, and tell you what it did. You play the game. You ask for more. You're in charge.
Bonus: ask it to add something weird. Flying cats. Disco lights. Whatever you can imagine.
