The Practice Room/AI Fluency/What AI actually is

AI Fluency • episode 1 • 62 s

What AI actually is

People assume three things about AI. That it understands the question. That it knows what is true. And that it learns from them.

Next walkthroughWhat a pattern actually is

Three things people assume

That it understands the question. That it knows what is true. That it learns from them. None of those is how it works, and every one of them leads somewhere unhelpful.

It is a prediction machine. Someone who has done a thousand crosswords can fill in a clue they have never seen before. Not because they know the answer, but because they know the shape of it.

What one answer actually is

It works one small piece of a word at a time, and each piece is a guess about what fits next. A page of writing is roughly six hundred and fifty guesses in a row.

Nothing in that process checks the world. It checks what usually follows what. That is the single fact underneath almost every surprising thing these tools do.

What each assumption costs you

Assuming it understands leads you to write short and vague, because you expect it to work out what you meant. It does not work out what you meant. It predicts from what you typed.

Assuming it knows what is true leads you to skip checking. A confident sentence and a correct sentence look identical coming out of a prediction machine, because confidence is part of the shape it learned.

Assuming it learns from you leads you to correct it once and expect it to stick. It does not stick on its own. That is why the memory files in the other tracks exist.

It is not thinking

It is extremely well read, and it is finishing your sentence. Both halves of that matter. Extremely well read is why the answers are often good. Finishing your sentence is why they are sometimes confidently wrong.

Check yourself

What is it doing when it answers you?

Predicting what comes next, one small piece of a word at a time.

Roughly how many guesses is a page of writing?

About six hundred and fifty.

You correct a mistake. Does it learn from that?

Not on its own. That is why tools have memory files you write yourself.

Why does a wrong answer sound as certain as a right one?

Confidence is part of the shape it learned to produce. Nothing in the process is checking the world.

Words used here

Token • Why it makes things up • Context window

Go deeper

DeepLearning.AI runs short free courses on the machinery underneath, if the crossword version has left you wanting the real one.

What she says

People assume three things about AI. That it understands the question. That it knows what is true. And that it learns from them.

None of those is how it works.

It is a prediction machine. Someone who has done a thousand crosswords can fill in a clue they have never seen before. Not because they know the answer. Because they know the shape of it.

That is what this does, one small piece of a word at a time. A page of writing is about six hundred and fifty guesses in a row.

It is not thinking. It is extremely well read, and it is finishing your sentence.

That is your AI tip for today.

No permission required.

Filed under Level: Start Here

Comments

No comments yet. Ask what did not make sense.

Comments are read before they appear.

No schedule, no filler

Get the next entry