Part 1: What AI Is

How does AI talk?

AI doesn't write a whole sentence in one go. All it does is pick just the next piece.

It adds the piece it picked, then picks the next one — repeat that hundreds of times and you get the long answer you see.

So AI isn't fetching an answer from somewhere. Every time, it freshly picks the most likely next piece.

Try it — the next-word picking machine

Let's play AI. You'll see the candidate words for the blank and the chance of each one. Press the button and one of them gets drawn and added to the sentence.

In one picture

This is everything AI does in one step. The rest is just this, repeated.

Go around this loop a few hundred times and you get a paragraph. That's why AI is closer to “picking, over and over” than to “thinking.”

1. Q1What does AI do in a single step?

   1. ① It makes the whole sentence in one go
   2. ② It picks the next piece and adds it on
   3. ③ It finds a similar sentence on the internet
   4. ④ It pulls out one of the answers it saved earlier

2. Q2In the experiment above, what happens if you turn “Boldness” down to 0?

   1. ① You get a different sentence every time
   2. ② You get the same sentence no matter how many times you press
   3. ③ The AI gets more accurate
   4. ④ It gets faster

3. Q3“AI searches for the answer somewhere and then tells you.” True or false?

   1. ① True
   2. ② False

AI is a machine that picks the next piece over and over, really fast.

What's a token, the 'piece' AI talks in?

The “next piece” from yesterday has a name in the AI world: a token.

A token is not a letter and not a word — it's a piece cut to be easy for AI to compute with, and its size changes from case to case.

So the same sentence gets a different count in tokens, and time and cost go up with that count.

Try it — the sentence-splitting machine

Pick a sentence and cut it up three ways. Each way gives a different number of pieces.

In one picture

Same sentence, but the unit people count (words) and the unit AI counts (tokens) are different.

The orange boxes are pieces that got split off. The common word ('My') stays whole; slang and less common words ('lowkey', 'devoured') get cut finer. (Splits from the tokenizer GPT-4 uses.)

1. Q1What is a “token”?

   1. ① Always exactly one letter
   2. ② Always exactly one word
   3. ③ A piece AI cuts text into to make computing easier (its size varies)
   4. ④ The whole word list people set up in advance

2. Q2What happens for the AI when the number of tokens goes up?

   1. ① It gets faster
   2. ② There's more to compute, and it takes longer
   3. ③ No effect at all
   4. ④ It always gets smarter

3. Q3Different AIs can split the same sentence into tokens differently. True or false?

   1. ① True
   2. ② False

To AI, a sentence isn't letters or words — it's a line of pieces called tokens.

Why does AI answer differently every time?

When AI picks the next piece, it gives each candidate a probability, then does a lucky draw using those odds.

Temperature is the knob that decides how flat that probability table gets — turn it down and only #1 gets drawn; turn it up and lower candidates get drawn often too.

So with a low temperature the answers come out pretty much the same, and with a high one they get more varied, but the risk of going off the rails grows too.

Try it — turn the temperature knob

So far you only looked at the candidates. Today you change the probability table itself. Move the knob and the bars really change length.

In one picture

Same candidates, same ranking — change only the temperature and the bars look this different.

On the left, almost only #1 gets drawn; on the right, even #5 gets drawn fairly often. The candidate order is exactly the same in all three.

1. Q1What happens if you turn the temperature down very close to 0?

   1. ① The ranking of the candidates flips
   2. ② Almost always only the #1 candidate gets drawn
   3. ③ There are more candidates
   4. ④ The AI gets smarter

2. Q2When you raise the temperature, what grows along with it?

   1. ① Answers get more varied, but so does the risk of going off the rails
   2. ② Computing gets faster
   3. ③ Accuracy always goes up
   4. ④ The number of tokens goes down

3. Q3Set the temperature to 0, and the same question always gets exactly the same answer. True or false?

   1. ① True
   2. ② False

Temperature isn't a knob that makes AI smarter — it's a knob for how bold it gets.

What does it mean for AI to 'learn' something?

For AI, learning isn't understanding the way it is for people — it's a process of nudging numbers (weights) a little at a time so they get closer to the right answer.

The method is simple: hide part of a sentence, have it guess the next word, then tweak the weights slightly by however much it was off.

Do this on internet text and repeat it hundreds of billions of times, and you get a big bundle of numbers with grammar and common sense melted right in.

Try it — put the training cycle in order

Put the steps of 'learning by repeating' in order, for three situations.

In one picture

Repeat four steps hundreds of billions of times, and random guesses slowly turn natural.

The orange box (④ adjusting the weights) is the key step that repeats every time. As these repeats pile up, the predictions start to sound natural.

1. Q1What exactly does it mean when AI 'learns'?

   1. ① It comes to understand meaning, like a person
   2. ② It nudges numbers (weights) a little at a time to get closer to the right answer
   3. ③ It saves the whole internet
   4. ④ It memorizes a fixed table of rules

2. Q2Training means making big fixes so it gets the answer exactly right in one go.

   1. ① True
   2. ② False

3. Q3What happens if training runs just once?

   1. ① It gets smart right away
   2. ② Almost nothing changes
   3. ③ It actually gets worse
   4. ④ Training is finished

Training isn't a 'lightbulb moment' — it's 'fixing a little,' over and over.

Just how big is this number?

Parameters are how many weight numbers keep getting adjusted during training. The more there are, the bigger the container for patterns.

Today's big models have tens of billions to 1 trillion parameters. That's still short of the number of synapses in a human brain (about 100 trillion).

More parameters doesn't automatically mean smarter — what it learned, and how, has to come along with it.

Try it — the number-sense slider

Move the slider to guess how many digits it has (10 to what power?), then see how close you are to the real value.

In one picture

Put the three numbers side by side and you can feel how different their digit counts are.

The bar lengths aren't exact proportions — they're drawn to give you a feel for the difference in digits.

1. Q1The more parameters, the smarter the AI — always.

   1. ① True
   2. ② False

2. Q2What exactly is a parameter?

   1. ① The number of sentences the model memorized
   2. ② How many weight numbers keep getting adjusted during training
   3. ③ The number of servers
   4. ④ The number of people using it at the same time

3. Q3As far as we know, the estimated number of synapses in a human brain is bigger than the parameter count of any AI model released so far.

   1. ① True
   2. ② False

A bigger number doesn't always mean smarter.

Why does AI calmly make up facts?

AI isn't looking up facts — it just picks the most likely next words, so it can produce wrong sentences with total confidence.

This is called hallucination — it's not an intent to lie, but a limit of a design that builds answers that sound likely.

It's especially likely to get things wrong when you ask about facts that are missing, rare or very new in its training data — always check somewhere else.

Try it — probability slider

Pick a question and see how the odds spread out. For a question with no real answer, #1 changes every time you roll.

In one picture

Even with no right answer, AI picks one believable-sounding answer.

Even when a question has no real answer, AI puts together something believable and answers confidently.

1. Q1What's the root cause of hallucination?

   1. ① The AI means to lie
   2. ② The AI is built to pick words that sound likely, by probability
   3. ③ Server errors
   4. ④ The user asked the question wrong

2. Q2Hallucination happens more often when you ask about things that are missing or rare in the training data.

   1. ① True
   2. ② False

3. Q3If AI answers in a confident tone, what it says is always true.

   1. ① True
   2. ② False

Saying it confidently doesn't make it true.

Why does AI forget what you just said?

AI doesn't store a chat the way your brain does. Every time, it rereads the whole conversation so far from the top, then answers.

But how much it can read at once (its context length) has a limit, so when a chat gets long, the early part gets pushed out.

So for anything really important, bringing it up again later in the chat helps the AI and you.

Try it — context window tracker

See for yourself how old messages get pushed out as a chat goes on.

In one picture

The window has a fixed size. When a new message comes in, the oldest one gets pushed out.

Only messages 4–6 are inside the window. Messages 1–3 have been pushed out and don't count for this answer.

1. Q1Why does AI ‘forget’ the start of a chat?

   1. ① Its feelings got hurt
   2. ② It can only read so much at once (its context)
   3. ③ It ignores you on purpose
   4. ④ It's testing you

2. Q2Even in a really long chat, AI always remembers exactly what you said first.

   1. ① True
   2. ② False

3. Q3If you repeat important info partway through a chat, AI uses it better.

   1. ① True
   2. ② False

If it matters, say it again.

Why does changing one question change the whole answer?

A prompt is the whole text you give the AI — change the text, and the odds for the next word change completely.

A short, vague prompt gets a vague answer. A prompt that spells out the role, conditions and format gets a much more useful one.

Writing good prompts isn't a trick to fool the AI. It's just saying more clearly what you want.

Try it — sharpen the prompt

Every condition chip you tap makes the prompt more specific — and the result gets better too.

In one picture

Same AI, different prompt, different result.

The more you spell out audience, format and length, the more useful the result.

1. Q1What's the key to a good prompt?

   1. ① A secret spell that tricks the AI
   2. ② Saying clearly what result you want
   3. ③ Keeping it as short as possible
   4. ④ Using lots of big words

2. Q2Change the prompt, and the odds for the AI's next word change too.

   1. ① True
   2. ② False

3. Q3No matter how specific your prompt is, the result barely changes.

   1. ① True
   2. ② False

Don't change the AI — change the question.

Is a bigger model always better?

A big model is one with lots of parameters, which you met on DAY 05 — it knows a lot and holds up well on hard problems it's never seen.

But big models are slow and expensive, and they won't even fit on your phone. They only run on servers somewhere across the internet.

So sometimes what a big model learned gets poured into a small model. The job decides the right size.

Try it — which one gets this job?

Here are six situations. Pick the big model or the small one for each. Every pick is marked right away, with the reason.

In one picture

They're not ranked by skill — they just live in different places.

A small model is fast because it's close by. A big model holds up on hard stuff because it knows a lot.

1. Q1When a model is ‘big,’ what does it mostly have a lot of?

   1. ① Letters shown on the screen
   2. ② Parameters (numbers tuned during training)
   3. ③ People using it
   4. ④ Time it takes to answer

2. Q2What does a small model clearly do better than a big one?

   1. ① Solves hard new problems better
   2. ② Knows more facts
   3. ③ It's fast, cheap and runs offline on your device
   4. ④ Never says anything wrong

3. Q3There's a way to pour what a big model learned into a small model.

   1. ① True
   2. ② False

Big isn't what's good — the right size for the job is.

Could this AI fit on my laptop?

With a closed model, the core stays inside the company's servers. You can only talk to it through an app or some other window — you can't get the model itself.

With an open model, you can download that whole core — the pile of parameter numbers from DAY 05, called the weights — and run it on your own computer.

But ‘open’ isn't all-or-nothing. Most models open the weights but not the training data, so how open a model is depends on the piece.

Try it — split ‘open’ into four pieces

We've split one model into four pieces. Tap the three ways of releasing it below and watch which pieces open and which stay locked.

In one picture

You ask questions the same way, but the model lives in a different place.

You only meet a closed model at a window. An open model comes onto your machine as a file.

1. Q1When a model is ‘open,’ what can you mainly get?

   1. ① The password to the company's servers
   2. ② The numbers tuned in training — the weights
   3. ③ The AI's chat history with people
   4. ④ A share of the company's money

2. Q2What can you do when you use a closed model?

   1. ① Download the model file onto your laptop
   2. ② Look at the data it was trained on
   3. ③ Ask and get answers through an app or API
   4. ④ Change the numbers inside the model yourself

3. Q3If a model's weights are released, its training data is usually released too.

   1. ① True
   2. ② False

Open doesn't mean free — it means you can get the core in your own hands.