Part 3: What It Can Do

When AI squeezes 20 pages into three lines, what does it throw out?

Summarizing isn't chopping text off from the start. It's choosing what to keep and what to throw out, and AI goes through the same steps people do.

AI scans the whole text in one go, then weighs 'how important is this?' for each sentence. Then it picks only the high scorers and rewrites them in its own words.

That's also why you can predict where summaries go wrong: words that weren't in the original sneak in. Always check a summary you get against the original at least once.

Try it — Put the steps of a summary in order

Tap the mixed-up pieces in order. It's almost the same order you'd go through to shrink the reading for a graded class project.

The four steps of a summary

These are the steps AI goes through inside when it shrinks a long text.

Because the last box rewrites things in its own words, phrases that weren't in the original can sneak in.

1. Q1What's the first thing AI does when it summarizes?

   1. ① Scans the whole text once
   2. ② Cuts sentences from the start
   3. ③ Reads only the title
   4. ④ Finds similar texts on the internet

2. Q2When you ask for a summary, what changes the result the most?

   1. ① Writing the request nicely
   2. ② Saying how many lines and who it's for
   3. ③ Asking over and over
   4. ④ Using only short sentences

3. Q3What must you do after you get a summary?

   1. ① Turn it in right away
   2. ② Measure its length
   3. ③ Check it against original
   4. ④ Change the font

Summarizing isn't the skill of shrinking — it's the skill of choosing what to throw out.

Why is 'You're taking my hair' a mistranslation?

Translating isn't swapping words one at a time. It's grabbing the meaning of the whole sentence first, then building that meaning again in another language.

Old translators swapped in words and short chunks like a dictionary. Today's AI, a lot like yesterday's summarizing, lets the words in a sentence look at each other to grab the meaning first.

In 2016, Google first put this method into a real translation service (Chinese to English came first). That's when clunky word-for-word translations dropped sharply.

Try it — Word-swap translation vs. meaning translation

We translated the same Spanish sentences into English two ways. Switch back and forth between the two, and you'll see where it falls apart.

Three steps of translating the meaning

Notice there's no box for swapping words at all.

Back before this middle box existed, translation always fell apart at idioms and missing subjects.

1. Q1Why is it wrong to swap 'Me estás tomando el pelo' word by word?

   1. ① The number of words is different
   2. ② The grammar is too hard
   3. ③ An idiom carries one meaning as a whole
   4. ④ Because 'pelo' is a Spanish word

2. Q2What does AI do when it translates '¿Ya comiste?' into English?

   1. ① Copies the words over as they are
   2. ② Writes out how it sounds
   3. ③ Cuts the sentence shorter
   4. ④ Works out the missing subject 'you' and adds it

3. Q3Why is it better to translate a long text all at once?

   1. ① Nearby lines become evidence
   2. ② It has more letters
   3. ③ It translates faster
   4. ④ It fixes the spelling

Swap the words, and you get a sentence. Carry the meaning, and you get understood.

Writing homework: how much do you hand over, and where do you take over?

AI is most useful not as a machine that writes for you, but as a tool that helps you work over what you wrote.

The rule is surprisingly simple. Your thoughts, your experiences and fact-checking are yours to do; polishing, shortening and listing ideas you can hand over.

Sentences AI wrote still go out with your name on them. So only keep sentences you can read and nod along to — that's the final checkpoint.

Try it — This step: hand it over, or do it myself?

We split a writing assignment into six pieces. For each one, pick one of the two. Each pick gets graded right away.

OK to hand over vs. mine to do

The line gets drawn at 'who's the only one who can do this?'

Hand over just the left side, and your writing time drops a lot. Hand over the right side too, and there's nothing of you left.

1. Q1Which part of a book report should you never hand over?

   1. ① Fixing spelling
   2. ② Listing title ideas
   3. ③ Making sentences shorter
   4. ④ What I thought about it

2. Q2AI gave you twenty title ideas. What next?

   1. ① I pick one myself
   2. ② Use the first one
   3. ③ Get twenty more
   4. ④ Glue them all together

3. Q3What should you do with numbers for a presentation?

   1. ① Use the AI's answer as is
   2. ② Check the original source
   3. ③ Round them roughly
   4. ④ Leave the numbers out

You can hand over the fixing, but hand over the choosing too, and nothing's left.

If code comes out when you describe it, why doesn't it end there?

It's true that if you describe it in words, code comes out. But the real half of coding happens after the code comes out — running it, reading the errors, fixing it.

With AI-written code, nobody knows if it's right until it runs. What decides isn't whether it looks believable, but whether running it gives the result you wanted.

So a red error message isn't you getting scolded — it's a note telling you what went wrong. Showing the whole thing to AI is the fastest way to fix it.

Try it — The order it takes to get code running

Tap the mixed-up pieces in order. Whether you're building a game or organizing a homework spreadsheet, this order is the same.

Coding isn't one line, it's a loop

Using AI doesn't make this loop go away. It just goes around faster.

From the last box, you go back to running it. Almost nothing is done in one loop.

1. Q1AI wrote you some code. Who decides whether it's right?

   1. ① The result you get when you run it
   2. ② If AI says it confidently, it's right
   3. ③ If the code is long, it's right
   4. ④ No errors means it's definitely right

2. Q2When an error pops up, how should you ask AI about it?

   1. ① Just write 'it doesn't work'
   2. ② Show it the whole error message
   3. ③ Delete the code and ask again
   4. ④ Pass on a summary of the error

3. Q3What's the trick for building something big?

   1. ① Ask for it all at once
   2. ② Write long code
   3. ③ Small pieces, one at a time
   4. ④ Ignore the errors

Getting the code is just the start; coding begins the moment you run it.

The more detail you give, the more the picture looks like your idea?

An image AI doesn't go find a picture somewhere. It draws by clearing away noise, bit by bit, from a screen full of dots.

Your description is the dial that sets which way that clearing goes. So anything you don't pin down in words, AI decides on its own — the colors, the time of day, the mood.

But detailed isn't the same as greedy. Put in too many things at once, and some of them actually get left out. Fixing as you go is faster than cramming it all in at once.

Try it — How does the result change as your description gets more detailed?

Turn the dial to make the description more detailed. Imagine you got ten pictures back — this shows roughly how many of each kind you'd get.

Drawing by clearing away the noise

It doesn't start from a blank page — it starts from a screen full of dots.

For anything you didn't mention, AI picks the direction for you. That's why even the same description gives a slightly different picture every time.

1. Q1How does an image AI make a picture?

   1. ① It finds a similar picture on the internet
   2. ② It clears noise from a screen full of dots
   3. ③ It traces lines a person drew
   4. ④ It cuts and pastes photos

2. Q2Why do you get so many way-off pictures when you only type 'cat'?

   1. ① AI doesn't know what cats are
   2. ② The text is too short
   3. ③ AI decides whatever you didn't specify
   4. ④ There aren't enough cat pictures

3. Q3What if you put in twelve things?

   1. ① All go in exactly
   2. ② Colors get darker
   3. ③ The picture gets bigger
   4. ④ Some easily get dropped

Whatever you don't say, AI decides for you, so start by saying what you want to decide.

How many dots are in one second of video?

A video isn't really a "moving picture." It's lots of still pictures shown one after another, fast — usually about twenty-four to thirty of them every second.

So a video AI has dozens of times more work than the picture AI you met yesterday. And it can't draw each picture on its own — each one has to connect to the one before it, or it won't look natural to us.

In numbers, one second means more than 60 million dots made from scratch and lined up with each other. That's why making videos is so slow and heavy.

Try it — How many digits? Guessing one second of video

Move the slider to guess how many digits the number has. Then compare it with the real value, and you'll feel why video is so heavy.

One picture vs one second of video

Same screen size — look how different the job gets.

If the cup is on the left in picture 2 and on the right in picture 3, the video instantly looks weird.

1. Q1Why is making a video so much heavier than making a picture?

   1. ① It has more colors
   2. ② The screen is bigger
   3. ③ It needs sound added
   4. ④ It needs dozens of pictures that all match

2. Q2About how many pictures go into one second of video?

   1. ① About 30
   2. ② 1
   3. ③ About 300
   4. ④ About 3,000

3. Q3What do you call one picture in a video?

   1. ① Pixel
   2. ② Frame
   3. ③ Resolution
   4. ④ Subtitle

One second of video is thirty pictures, and those thirty have to connect.

How does your phone write down what you say?

AI doesn't hear your words in one go — it chops them into tiny pieces. It writes the sound down as numbers 16,000 times a second, then sweeps over them with a 25-millisecond window.

Each piece on its own is just a "pitch pattern." What you actually said only gets decided by looking at the pieces around it — like the video you saw yesterday, one slot alone doesn't tell you.

Speaking is the same road in reverse. It turns text into a sound pattern, then follows that pattern to stamp out air vibrations more than 16,000 times a second and send them out of the speaker.

Try it — How "Can you write it down" becomes text — and then sound again

Tap the buttons to change steps. You'll see speech get chopped into pieces, the pieces join into words, and the words turn back into vibrations.

Speech → text → speech, one full loop

Writing speech down and speaking out loud travel the same road in opposite directions.

This is why speech-to-text apps work best in a quiet room — the pattern is only clear when each piece has little noise mixed in.

1. Q1What's the first thing AI does when it turns speech into text?

   1. ① Chops it into very short pieces
   2. ② Understands the whole sentence at once
   3. ③ Looks the words up in a dictionary
   4. ④ Looks at the speaker's face

2. Q2When two words sound the same, like "write" and "right," how does AI choose?

   1. ① It picks whichever was louder
   2. ② It checks which one fits the words around it
   3. ③ It always picks the shorter word
   4. ④ It asks a person every time

3. Q3When text is turned into a voice, what comes out at the end?

   1. ① A photo
   2. ② A list of words
   3. ③ Air vibration values
   4. ④ A search term

A voice is vibration chopped into tiny pieces, and text is the answer you get by matching each piece with its neighbors.

Search box vs asking AI: what actually changes?

Search doesn't give you an answer — it gives you a list of places to go. It tells you where the answer might be, and going in, reading and picking is up to you.

AI takes what it has already read and builds you one chunk of sentences. There's only one thing to read, so it's fast, but it doesn't say where those words came from.

So there's just one fork in the road — can you see the source? Even when AI adds links, you only know if a link really says that by clicking it.

Try it — Same question: search box vs AI

Pick a situation and switch between "In a search box" and "Asking AI." The bar shows how clearly you can see where the words came from.

A list of places vs one chunk of answer

Ask the same question, and the two tools hand back answers in different shapes.

It also means that for any sentence going into your homework, you'll have to click through to the original at least once.

1. Q1What's the biggest difference between search and an AI answer?

   1. ① How fast they are
   2. ② Whether you can see where the words came from
   3. ③ How many words they use
   4. ④ Whether they work on a phone

2. Q2What goes wrong when AI pulls several sources together for you?

   1. ① The text gets too short
   2. ② It can never add links
   3. ③ Sources mix, so you can't trace each line
   4. ④ It only uses old sources

3. Q3AI put a link next to a sentence. So…

   1. ① It's all correct
   2. ② The link is decoration
   3. ③ It always beats search
   4. ④ Click it and check

Search tells you where to go, AI tells you what it is — and checking stays in your hands.

Research for a graded class project: how much can you hand to AI?

Researching with AI gets you started faster. When you don't know what to search for, it quickly comes up with search terms and ideas for an outline.

But what speeds up is the front part — finding and sorting. It shortens long texts and gathers scattered stuff, but while it shortens, words that weren't in the original can slip in.

So numbers, years, people's names and quotes need you to click through to the original yourself and check. In 2023, a lawyer in the US was fined after handing a court fake cases that AI had made up.

Try it — Researching "Why did the stream in our town get so dirty?"

Pick a research step and switch between "On your own" and "With AI." The bar shows how much you'll have to check again yourself afterward.

Your share to check, piece by piece

Split research into pieces, and you can see which parts you can hand off and which you can't.

The scale is drawn by feel. It's not an exact measurement — it shows where your own hands start being needed.

1. Q1What speeds up the most when you research with AI?

   1. ① Checking whether the numbers are right
   2. ② Answering questions when you present
   3. ③ The start: search terms and angles
   4. ④ Copying quotes exactly

2. Q2You want to use a sentence from AI's summary in your homework. What do you do?

   1. ① Paste it in as is
   2. ② Ask a friend
   3. ③ Make it longer
   4. ④ Find that point again in the original

3. Q3Where should you check numbers and years?

   1. ① Where first published
   2. ② AI's answer
   3. ③ A friend's slides
   4. ④ Any random blog

You can share the finding with AI, but the checking is yours all the way.

Stuck on a problem: get just the answer, or the steps?

When you ask AI about schoolwork, first decide whether you want just the answer or the steps. Ask the same question either way, and what comes back is completely different.

There's one rule for choosing — will I have to do this on my own later? If you'll have to do it yourself, get the steps. If you just need to know whether you got it right, get the answer.

Bonus: making AI go step by step helps AI solve better too. A 2022 study found that adding one line — "Let's think step by step" — made AI get a lot more math problems right.

Try it — In this situation: answer or steps?

Here are six situations. For each one, pick whether it's better to get just the answer or have the steps explained. You'll be scored as soon as you pick.

Getting the answer vs getting the steps

It's not good vs bad — they're for different moments.

If you got the steps, go one step further — solve it again without looking, and then it's really yours.

1. Q1What decides whether you get the answer or the steps?

   1. ① Whether I'll have to do this alone later
   2. ② How long the problem is
   3. ③ Which AI it is
   4. ④ How much time is left

2. Q2You've gotten the same type wrong three times. What should you ask for?

   1. ① Just the right answer, fast
   2. ② To walk through it line by line
   3. ③ A different problem
   4. ④ Just grading

3. Q3What should you do after getting the steps explained?

   1. ① Move to the next problem
   2. ② Copy it and hand it in
   3. ③ Solve it again, no peeking
   4. ④ Send it to a friend

Problems you'll meet again: get the steps. Problems you won't: get the answer.

Turning 100 survey sheets into a graph: what comes first?

Don't draw scattered info straight into a graph. Stick to this order: gather → table → fix → graph.

Tables have a rule: one entry per row, one value per cell. Once things are lined up like this, people and AI can both calculate right away.

What AI is good at is lining things up and calculating. Whether units got mixed, there are blanks, or a weird value snuck in — you have to look for that yourself.

Try it — From 100 survey sheets to a graph

Tap the mixed-up pieces in the right order. If you skip a step, you'll find out where things break.

Four boxes from info to graph

Skip a box and the final picture wobbles.

Line your info up in a table before handing it all to AI, and the answers get much more accurate.

1. Q1What's the right order for sorting out data?

   1. ① Graph → table → gather → fix
   2. ② Gather → table → fix → graph
   3. ③ Table → graph → gather → fix
   4. ④ Gather → graph → table → fix

2. Q2What rule should you follow when making a table?

   1. ① Just leave blanks empty
   2. ② Fit as much as you can in one cell
   3. ③ One entry per row, one value per cell
   4. ④ Give columns different names in each row

3. Q3Which of these is hard to hand to AI?

   1. ① Lining it up in a table
   2. ② Adding up totals
   3. ③ Drawing the graph
   4. ④ Noticing mixed-up units

The graph is just the last box. It's won or lost in making and fixing the table.

Presentation prep: which parts are AI's, and which are yours?

Preparing a presentation isn't one big lump. It splits into pieces like narrowing the topic, gathering sources, making slides and practicing.

The pieces AI is good for are rough drafts and busywork — listing outline ideas, shortening long sources, sketching a first draft of the slides.

The pieces you have to do exist because you're the one standing up there: deciding what to say, checking numbers, practicing, and answering questions.

Try it — Splitting up a 5-minute talk on "Recycling in our class"

Tap the buttons to split up the pieces. Watch which pieces light up and which go dim, and you'll see the rule.

How much you can hand off, piece by piece

Even in the same presentation prep, how much you can hand off depends on the piece.

The scale is drawn by feel. Don't memorize the numbers — look at where the top and bottom split.

1. Q1Which piece is good to hand to AI?

   1. ① Practicing the talk
   2. ② Answering questions
   3. ③ Listing outline ideas
   4. ④ Speaking in front of friends

2. Q2Why is it risky to hand over the whole presentation?

   1. ① The text gets smaller
   2. ② It takes longer
   3. ③ The slides get slow
   4. ④ You can't explain sentences you never read

3. Q3What must you do after you get a draft?

   1. ① Check numbers and sources
   2. ② Print it right away
   3. ③ Make the font bigger
   4. ④ Pick a background color

Hand off the pieces you can fix, and do the pieces with your name on them yourself.

Why are some answers boring and others off-the-wall?

Whether AI gives you a boring answer or an off-the-wall one comes down to this: when it picks the next word, does it grab only the most likely one, or less likely ones too?

There's a real knob that widens and narrows that range, and it's called "temperature". Turn it down and you only get safe words; turn it up and rarely used words start popping out.

So when you're gathering ideas, the trick is to turn the knob up, generate lots, and pick for yourself. The higher you go, the more nonsense you get, too.

Try it — What happens to ideas when you turn the knob up?

At 0, the knob only grabs the most likely word; at 100, it grabs unlikely words too. Turn it and watch where the bars lean. (A real AI's "temperature" is usually between 0 and 2. Here we stretched it to 0–100 so you can get a feel for it.)

Knob turned down vs turned up

Same question, but one knob changes the whole character of the answer.

Even if the chatbot doesn't show you a knob, saying "Give me twenty more, and make them weirder" has a similar effect.

1. Q1What has to change for AI to come up with off-the-wall ideas?

   1. ① Leave the AI on for longer
   2. ② Speed up the internet
   3. ③ Write a shorter question
   4. ④ It picks from a wider range of next words

2. Q2What else goes up when you turn the knob up?

   1. ① Answers that make no sense
   2. ② Accuracy
   3. ③ Calculation speed
   4. ④ The length of the answer

3. Q3What's the trick to using AI in a brainstorm?

   1. ① Trust the first answer
   2. ② Generate lots, then pick
   3. ③ Ask only once
   4. ④ Set the knob to 0

Turn the knob up and your ideas get wider, but the picking is still up to you.

You've got a big job: where do you even start?

When you feel stuck, don't ask AI for the answer — ask it to break the job down. Planning isn't about knowing the right answer; it's about deciding what to do first.

There's a set shape for breaking things down: goal → split → order → check. Fill in those four boxes and you've got a plan.

AI research uses the same trick. When AI was made to break a hard problem into small ones and solve them in turn, it did much better than when it was asked all at once.

Try it — Tap the four boxes in order

Tap the mixed-up pieces in plan order. Swap even one box, and you'll see right away why the plan falls apart.

The four boxes of a plan

Fill these four boxes and any job becomes a plan — leave one out, and that's exactly where things go wrong.

Instead of handing AI the whole thing, have it do these four boxes one at a time — then you can see which box went wrong.

1. Q1What should you decide first when you make a plan?

   1. ① The to-do list
   2. ② The goal you have to reach
   3. ③ The deadline
   4. ④ Which friend to work with

2. Q2Which is a better request when you have AI make a plan?

   1. ① Just handle all of it
   2. ② Just tell me the answer
   3. ③ Give goal and limits, ask it to break it down
   4. ④ Wait without saying anything

3. Q3What does the "check" box do in a plan?

   1. ① Adds more work
   2. ② Erases the goal
   3. ③ Pushes back the deadline
   4. ④ Catches problems early

Planning isn't knowing the right answer — it's breaking a big lump into small pieces.

If it can finish a job on its own, how much can you hand it?

There's really only one rule for what you can hand off and what you can't — can you undo it if it goes wrong?

An AI that you give just a goal, and that then searches, opens files and tries again when it fails, all on its own, is called an agent. That's where it parts ways with a chatbot that answers once and is done.

But agents often miss things people find easy. So even when you hand a job off, the trick is to leave a last step where a person looks it over.

Try it — Can you hand this whole job off?

Here are six situations. Think about just one thing — if it goes wrong, can you undo it or not?

Easy for people, hard for AI — the first tests in 2023

These are the numbers from the first time anyone properly measured how well AI can "do it all on its own."

These are the 2023 GAIA and WebArena numbers. AI has gotten a lot better since, but that still doesn't mean handing things off is free.

1. Q1How is an agent different from a regular chatbot?

   1. ① It talks more nicely
   2. ② It works without the internet
   3. ③ It uses tools on its own to take many steps
   4. ④ It can draw pictures too

2. Q2What decides whether you can hand a job off?

   1. ① How long the job is
   2. ② How many words it has
   3. ③ How fast the AI is
   4. ④ Whether you can undo it if it goes wrong

3. Q3What about sending money from an account?

   1. ① A person checks it
   2. ② Hand it all over
   3. ③ Have AI do it twice
   4. ④ Fix it later

If you can undo it, hand it off; if you can't, look at it yourself.