Part 5: Looking Inside

What's inside an AI if you open it up?

Inside an AI are dots stacked in layers, passing numbers to each other. That's called a neural network.

What one dot does is surprisingly simple. It multiplies the incoming numbers by weights and adds them all up, and if the total clears a threshold, it passes it on to the next layer.

The smarts aren't inside any one dot. As the numbers pass through layer after layer, simple math piles up into complex judgments.

Try it — What happens inside a neural network, in order

Tap the mixed-up pieces in the right order. Three sets: what one dot does, the path through the layers, and how the weights get set.

From input to output, through the layers

A neural network is numbers changing shape as they pass through one layer after another.

Numbers flow from left to right. The orange box in the middle is what makes stacking layers mean something.

1. Q1What's closest to what one dot in a neural network does?

   1. ① Multiplies by weights, adds, checks a threshold
   2. ② Memorizes whole sentences
   3. ③ Fetches answers from the internet
   4. ④ Writes down its own rules in words

2. Q2Why can a neural network make complex judgments?

   1. ① Each dot is really smart
   2. ② Simple math builds up over many layers
   3. ③ People wrote the answers in advance
   4. ④ Its screen is big

3. Q3What does a 'weight' decide?

   1. ① The size of the text
   2. ② Screen brightness
   3. ③ How strong a signal passes
   4. ④ Internet speed

One dot can only multiply and add. Stack them in layers, and out comes judgment.

Where should a word look to get its meaning?

Attention is the calculation that decides how hard each word in a sentence looks at the other words.

Attention doesn't grow without limit. Words share out a fixed total, so looking hard at one means looking less at the others.

That's how meaning gets settled. When attention piles onto the word that fits, what 'it' refers to comes into focus.

Try it — Turn the strength of attention with the handle

The higher you push the handle, the more attention piles onto one word. At 0 it looks at every word equally; at 100 it looks at just one spot. Each bar is that word's share of attention.

Four steps that set where attention goes

Attention isn't hard magic — it's a calculation that gives out scores and shares them out.

The Transformer you saw on DAY 14 runs these four steps across a whole sentence at once.

1. Q1What does attention do, in one line?

   1. ① Looks words up in a dictionary
   2. ② Decides how hard one word looks at another
   3. ③ Shortens sentences
   4. ④ Fixes spelling

2. Q2What happens when 'It was delicious' becomes 'It was full'?

   1. ① The sentence gets shorter
   2. ② Nothing changes
   3. ③ Attention moves from the cake to the dog
   4. ④ Attention spreads out evenly

3. Q3If one word gets looked at really hard, the rest...

   1. ① Vanish and come back
   2. ② All get stronger too
   3. ③ Double
   4. ④ Get smaller shares

Attention is sharing out a fixed amount of focus, and where it piles up is what the meaning becomes.

How does a computer understand the word 'apple'?

A computer doesn't know letters — it only handles numbers. So it turns each word into a bundle of several numbers and works with that.

They aren't just any numbers. It learns to give similar numbers to words that get used with similar words, so words with close meanings land in close spots.

Once meaning becomes a place, you can measure near and far with math. That's the base for tomorrow's lesson, finding similar things.

Try it — Switch words' pieces of meaning on and off

Let's pretend each word is a bundle of seven boxes of numbers. Tap a different word to see which boxes light up. The more lit boxes two words share, the closer they are.

Seen as letters vs seen as number bundles

Same word — but what you can do with it depends on how you store it.

The famous example 'king minus man plus woman lands near queen' also comes from this sense of place.

1. Q1Why turn words into bundles of numbers?

   1. ① To make the letters look nice
   2. ② To save storage space
   3. ③ Computers use numbers, and to measure closeness
   4. ④ To fix spelling

2. Q2Which words end up with similar numbers?

   1. ① Words with the same number of letters
   2. ② Short words
   3. ③ Words it learned first
   4. ④ Words used with similar words around them

3. Q3In the experiment, why is the soccer ball far from apple?

   1. ① Only one box overlaps
   2. ② Its name is longer
   3. ③ It came later
   4. ④ It has more boxes

Turn meaning into bundles of numbers, and 'similar' becomes 'close.'

Which of two sentences is closer?

Yesterday we turned words into bundles of numbers. Sentences can be turned into them the same way, and then you can measure the distance between two sentences.

Being close doesn't mean sharing words. Even with not a single word in common, sentences that mean the same thing land close together.

On the flip side, if the words are almost the same but the meaning is opposite, they end up far apart. Like 'fun' and 'not fun.'

Try it — Which one is closer to the base sentence?

Each base sentence comes with two candidates. Pick the one that's closer. You'll get scored right away, along with why.

Measuring how close they are

The longer the bar, the closer together the two sentences sit.

This is why search got smarter than it used to be — it finds the sentence you want even when you don't use the exact same words.

1. Q1What does it mean for two sentences to be 'close'?

   1. ① They share lots of words
   2. ② They were written on the same day
   3. ③ They're about the same length
   4. ④ Their spots are close as number bundles

2. Q2'This movie was not fun' and 'This movie was really fun' are...

   1. ① Same words, opposite meaning — so far apart
   2. ② Same words, so the closest of all
   3. ③ Close, because they're the same length
   4. ④ In the same spot, since both are about movies

3. Q3What if two sentences share no words at all?

   1. ① Always far apart
   2. ② Close if the meaning's same
   3. ③ Can't be calculated
   4. ④ Only half close

Being close doesn't mean sharing words. When meanings overlap, they land close together.

What happens when you ask AI something it doesn't know?

If it answers only from what it memorized, it's easy for it to make up something that sounds right. So before answering, it goes and finds related text first and keeps it open next to it.

The way it finds that text is exactly what you learned yesterday. It turns your question into a list of numbers and pulls out the pieces of text sitting closest to it.

That turns a closed-book test into an open-book one. Because it answers while reading the text it found, it can handle brand-new info, and it can tell you where each claim came from.

Try it — The steps of an open-book answer

Tap the mixed-up pieces in the right order. There are three sets: the big steps of answering, how the search works inside, and how sources get attached.

From question to source

Instead of only pulling from memory, it goes and finds text, lays it open, and answers.

The orange box is where yesterday's 'finding similar things' goes. The paper that laid out this method came out in 2020.

1. Q1Which is the right order for an 'open-book answer'?

   1. ① Write the answer first, then find text
   2. ② Get the question, find text, answer using it
   3. ③ Memorize all the text, then answer
   4. ④ Shorten the question, then answer

2. Q2How does it find the pieces of text that fit your question?

   1. ① The one that comes first alphabetically
   2. ② The longest piece
   3. ③ The piece closest to the question in numbers
   4. ④ The most recently made piece

3. Q3Which one is NOT a good thing about this method?

   1. ① Can keep up with new info
   2. ② Can show its sources
   3. ③ No retraining needed
   4. ④ Always gets it right

Answer from memory and it's easy to make stuff up, but find it, lay it open and answer, and the evidence stays.

Can you teach a fully trained AI one extra job?

You can. You take a base model that already went through its big round of training and train it a little more on just the job you need. That's called fine-tuning.

It's not learning everything again from scratch. It leaves almost all the numbers inside as they are and changes only a tiny part — a method called LoRA cut the numbers it needs to change by 10,000 times.

But it's good at some things and not others. Extra training is strong at changing tone and format, while yesterday's open-book answers are what you use to add new facts.

Try it — Same model, before and after a little extra training

Pick each of the three jobs and tap 'Before' and 'After' back and forth. You'll see what changes and what doesn't.

Raising one AI in two stages

The first two boxes are huge. The last two are tiny.

That's why thousands of models with different personalities branch off from the very same base model.

1. Q1What does extra training (fine-tuning) do?

   1. ① Rebuilds the model from scratch
   2. ② Searches the web and adds it to the answer
   3. ③ Teaches an already-trained model a bit more
   4. ④ A person fixes the answers by hand

2. Q2Why is a method like LoRA such a big deal?

   1. ① It makes every answer correct
   2. ② It cuts the model's size in half
   3. ③ It runs without the internet
   4. ④ It needs 10,000 times fewer numbers changed

3. Q3How do you get it to tell you a lunch menu that changes every week?

   1. ① Find it and attach it
   2. ② Teach it more
   3. ③ Make the model bigger
   4. ④ Write a longer question

You don't raise a grown-up model all over again, you just polish it a little, as much as you need.

Given two answers, which one would a person score higher?

You have the AI write two answers to the same question, then a person picks the better one, tens of thousands of times over. Those picks become the material for the next round of training.

It's not grading right or wrong. What gets collected is people's taste. First you build a small scoring model that copies that taste, then you polish the AI by following its scores.

It worked really well. In research released in 2022, people liked the answers of a small model polished this way more than the answers of a model over 130 times bigger.

Try it — If you were the human scorer, which one would get your point?

In six situations you get two answers each. Give the point to the one that looks better. Once you've picked them all, you'll see the standards people used.

How people's picks turn into scores

People don't write the answers. They just pick one of two.

This method is called RLHF for short. It stands for reinforcement learning from human feedback.

1. Q1In this method, what does the person do?

   1. ① Writes the right answer themselves
   2. ② Makes up the questions for it
   3. ③ Fixes spelling mistakes
   4. ④ Picks the better of two answers

2. Q2Why don't people have to score every single answer?

   1. ① The picks are used to build a scoring model
   2. ② The AI already knows the right answer
   3. ③ There's an answer key on the internet
   4. ④ There's always only one answer

3. Q3What mostly changes when you polish an AI this way?

   1. ① It knows more facts
   2. ② Its answers get more usable
   3. ③ The model gets smaller
   4. ④ It needs no internet

Nobody told it the right answer, people only told it which one was better.

How did graphics cards made for gaming become the heart of AI?

A graphics card (GPU) is a part built to do a huge number of easy calculations all at once. To color in the millions of dots on a game screen at the same time, that was the only way.

It became AI's heart because AI training is exactly that kind of math. A card widely used for training packs about 80 billion tiny switches inside.

That one card does about 1,000 trillion calculations a second. Even that isn't enough, so Llama 3 was made by tying 16,000 cards together into one giant cluster and running them for 54 days.

Try it — Getting a feel for graphics cards by order of magnitude

Use the slider to guess roughly what power of 10 it is, then tap to check the answer. The goal isn't to nail it, it's to feel how many zeros you're dealing with.

A few smart people vs thousands of simple ones

These two parts are good at completely different kinds of work.

The moment the Transformer you saw on DAY 14 made 'calculating all at once' possible, this part was a perfect fit.

1. Q1Why are graphics cards such a good fit for AI math?

   1. ① They do tons of easy calculations all at once
   2. ② They solve hard problems all on their own
   3. ③ They have fast internet
   4. ④ They have huge storage space

2. Q2About how many tiny switches are in one card used for AI training?

   1. ① About 8,000
   2. ② About 80 billion
   3. ③ About 800,000
   4. ④ About 80 quadrillion

3. Q3How many cards were tied together to train Llama 3?

   1. ① 16 cards
   2. ② 160 cards
   3. ③ 16,000 cards
   4. ④ 1,600 cards

AI didn't pick graphics cards. The shape of the math was the same all along.

How much electricity does it take to ask AI one question?

One question takes less than you'd think. In measurements released in 2025, a typical text question to Gemini used 0.24 watt-hours (the median) — about the same as running a TV for 9 seconds.

The problem is that those questions pile up nonstop. So instead of one question, you have to look at the electricity of the whole data center, the building where AI lives.

The world's data centers used 415 terawatt-hours of electricity in 2024, about 1.5% of all the electricity in the world. The International Energy Agency expects that to more than double to 945 terawatt-hours by 2030.

Try it — Lining up electricity in the same unit

All three use the same unit, the watt-hour (Wh). Use the slider to guess roughly what power of 10 it is, then tap to check the answer. You'll see how far apart the orders of magnitude are.

How big a slice of all electricity do data centers take?

The unit is terawatt-hours. 1 terawatt-hour is 1 trillion watt-hours.

The International Energy Agency explained that 945 terawatt-hours is about as much electricity as the whole of Japan uses in a year.

1. Q1How much electricity does one text question to AI take?

   1. ① About as much as an air conditioner all day
   2. ② About as much as a TV on for 9 hours
   3. ③ About as much as a TV on for 9 seconds
   4. ④ No electricity at all

2. Q2In 2024, what share of all electricity did the world's data centers use?

   1. ① About 0.001%
   2. ② About 50%
   3. ③ About 15%
   4. ④ About 1.5%

3. Q3What does the International Energy Agency expect for 2030?

   1. ① It more than doubles
   2. ② About the same as now
   3. ③ It drops by half
   4. ④ Nobody has any idea

One question is 9 seconds of TV. Add them all up and it's a whole country.

Can your phone run AI all on its own?

Some of it, yes. Phones these days have a small model with about 3 billion numbers inside, so even with no internet they can summarize, polish your writing and sort your notifications.

But big jobs are too much for it. Long writing or hard questions get sent to the building you saw yesterday, and the answer comes back. The phone model and the server model differ in size by tens to hundreds of times.

So phones today split the work into pieces. Whatever it can do, it finishes inside — it's faster, uses less electricity, and your text doesn't leave the phone.

Try it — Does this job finish on the phone, or go out?

Tap the buttons to switch situations. Green pieces finish on the phone, orange pieces leave the phone, and faded ones don't work at all.

Pieces on the phone vs pieces on a server

It's the same AI, but what you get and what you give up depends on where it runs.

The on-phone model released in 2024 had about 3 billion numbers and produced about 30 tokens a second on an iPhone.

1. Q1What's true about an AI model that runs on your phone?

   1. ① It's much bigger than server models
   2. ② It can't handle photos at all
   3. ③ It uses more electricity
   4. ④ It's small and runs without internet

2. Q2What happens when a hard question gets sent to a server?

   1. ① It does better, but your text leaves the phone
   2. ② Your phone battery lasts longer
   3. ③ You don't need internet
   4. ④ The answer is always wrong

3. Q3How do phones actually work today?

   1. ① All on the phone
   2. ② Split into pieces
   3. ③ All on a server
   4. ④ They do nothing

Your phone doesn't send it all out or keep it all in. It holds on to just the pieces it can handle.