cadence_learn
start_with_ai / level 01 of 10 ▶ 8 min

what this thing actually is.

Before you learn to use AI well, it helps enormously to know what it's doing under the surface. It's simpler than people make it sound — and once you see it, everything else about AI starts making sense.

after this level you'll be able to
  • Explain in one sentence what an AI model actually does
  • Say why it sometimes states wrong things with total confidence
  • Spot the difference between what it knows and what it's inventing
the whole idea

it predicts what comes next

the one sentence

An AI model reads what's in front of it and predicts what should come next — one small piece at a time — based on patterns it learned from an enormous amount of writing.

That's genuinely it. When you type a question, the model isn't looking up an answer in a filing cabinet. It's doing something much closer to a very, very well-read person finishing your sentence — then finishing the next one, and the next, until it has produced a whole reply.

Here's the same idea as a picture. Given a few words, the model works out which piece is most likely to follow, picks one, adds it, and repeats:

YOU TYPED The cat sat on the WHAT MIGHT COME NEXT mat 61% floor 14% sofa 9% roof 3% …and thousands more, each less likely pick one, add it, do it all again Repeat a few hundred times and you have a paragraph.
Every answer is built one piece at a time. There is no moment where the model "decides" the whole reply and then writes it down. It commits to each piece as it goes.

Modern models are far more sophisticated than this cartoon — they weigh everything you've said, they've been trained hard to be helpful and accurate, and the best of them reason through problems in genuinely impressive ways. But the engine underneath is still "what should come next, given everything so far." Hold onto that and the rest of this course will feel obvious rather than magical.

the consequence that matters most

why it can be wrong and sure at the same time

Here's the thing almost nobody explains to beginners, and it's the single most useful thing in this level.

The model is producing text that looks like a good answer. When it knows the answer, "looks like a good answer" and "is a good answer" are the same thing. When it doesn't, they come apart — and the model can't feel the difference from the inside.

TWO ANSWERS, AS THE MODEL SEES THEM ACTUALLY TRUE "The Eiffel Tower is in Paris and opened to the public in 1889." Confident. Specific. Fluent. COMPLETELY INVENTED "The Eiffel Tower was restored in 1962 by the architect L. Ferrand." Confident. Specific. Fluent. Same shape. Same certainty. Only one is real — and the model can't tell which.
This has a name: a hallucination. It isn't lying, because lying needs knowing the truth first. It's filling a gap with something that fits the pattern.
the habit this should give you

Confidence in an AI's tone tells you nothing about whether it's right. Names, dates, numbers, statistics, quotes, legal or medical specifics, and anything you'd be embarrassed to get wrong — check those against a real source, every time. Everything else, use your judgment.

the other half of the picture

what it knows, and what it has no idea about

The model learned from a huge amount of publicly available writing, up to a certain date. That's a lot of the world — and none of your world.

ALREADY IN THERE How language works General knowledge & history How to write, code, summarize Patterns from millions of examples NO IDEA — UNLESS YOU GIVE IT Anything about you Your notes, files, messages What happened recently Your last conversation with it You close that gap by handing it your material That's Level 3 — and it's where AI stops being a toy.
The right-hand column is the whole game. Most people who find AI underwhelming are only ever asking about the left-hand column, where it's competing with a search engine.

One more thing worth knowing early: unless you're using a feature that remembers you, each new conversation starts blank. It isn't ignoring you or forgetting on purpose — the slate is genuinely clean. That's why the same question can get a great answer one day and a mediocre one the next: the difference was what you gave it, not its mood.

do it now

catch it being wrong — on purpose

Reading about this does very little. Watching it happen once, in front of you, does a lot. This takes about four minutes and it's the most valuable four minutes of the whole course.

your turn 4 minutes
  1. Open any AI chat. ChatGPT, Claude, Gemini, Copilot — all fine, all free to start. It genuinely doesn't matter which one you pick today.
  2. Ask it about something you personally know inside out — and pick something obscure. Not "tell me about London." Something like your specific school, a small local business, a niche detail of a hobby, a minor character from a book you love, the rules of a game you play.
    try something in this shapeTell me about [the very specific thing you know well]. Include a few concrete details and dates.
  3. Read it as an expert, not as a student. Somewhere in there — maybe not everything, but somewhere — you will almost certainly find a detail that is plainly wrong, or a confident-sounding fact that was never true. Find it.
  4. Now push on it. This is the part that sticks.
    paste this nextHow confident are you in that answer? Which parts might you have got wrong?
    Notice what happens. It will often either defend the wrong detail smoothly, or fold and apologize for things that were actually correct. Either way you've just watched it do the thing this level is about: produce what a good answer looks like, without a reliable sense of whether it is one.
if nothing went wrong

Then you picked something too well-known — the model has seen it thousands of times. Go smaller and more local. The gaps live in the corners, and finding the corners is the skill.

clearing the air

four things people believe that aren't true

what people think

"It's a search engine with better manners."

what's actually true

A search engine finds pages someone wrote. A model generates new text from learned patterns. Some can now search the web as well — but that's a tool bolted on, not what the model is.

what people think

"It knows when it doesn't know."

what's actually true

This is the dangerous one. It has no reliable internal signal for "I'm making this up." That job is yours, and it's the main skill you're here to build.

what people think

"You need to be technical to use it."

what's actually true

The opposite. The people who get the most out of it are usually the ones who explain things well and know their own subject — not the ones who can code.

what people think

"It's either brilliant or useless."

what's actually true

It's a fast, tireless, occasionally wrong assistant. The value comes from knowing which jobs to hand it — which is exactly what the next nine levels cover.

level complete

three things to carry into level 02

01

It predicts, it doesn't look up

Every answer is built one piece at a time from learned patterns — not retrieved from a store of facts.

02

Sure ≠ right

Tone carries no information about accuracy. Verify anything that matters, and you'll never be caught out.

03

Your job is judging

You're not asking an oracle. You're directing a fast, capable assistant — and deciding what's good.

why this level came first

Nearly every bad experience people have with AI traces back to not knowing this. They trusted something they shouldn't have, or gave up because it "made things up," or never realised the real power was in handing it their own material. You now know all three. That already puts you ahead of most people using it daily.

Stuck? Peter reads these personally and replies to your email.
Ask Peter →