A book, its study guide, its question bank

Someone who worked it out and is showing you how: a few steps ahead of you, not lecturing at a board.

Picture the machine that writes your poem as a clerk in a record room. Somewhere inside, in a labelled file, your poem is lying finished, and the clerk’s whole job is to find it. Most of us carry some version of this picture — where else would the poem come from?

Test it. Ask for the poem twice. Two different poems come back. A clerk pulling a finished page from a file would bring the same page both times. So there is no file.

Then what is inside? Think of the man at the tea stall you cross every day after school. Thirty years at that counter, and every kind of talk has passed in front of him — results day, match day, bandh day. Has he memorised even one of those conversations? Not one. But say two sentences to him and he can tell where the third is going. Not because he has heard your third sentence before — because thirty years gave him a feel for what follows what.

A generative model holds exactly that kind of feel, learned from everything it trained on — not stored answers, but a statistical sense of what tends to follow what. Give it a prompt, and it draws one plausible reply from that sense, built fresh, word by word. Nothing is fetched from storage — every output is built at the moment you ask for it.

Has read every past paper. Names the trap out loud, with no warm-up, and talks to you as the candidate in the hall.

“Generative AI retrieves its answer from a large database.” Read that option again. It sounds right. It borrows the syllabus’s own words, and it is the picture most candidates carry into the hall — which is exactly why it is on the paper.

It is wrong, and the examiner knows how many of you will tick it. A generative model stores no answers. It learns a distribution — a statistical sense of what tends to follow what — and at generation it draws one plausible output from it.

The tell is one word: retrieves. Retrieval means the answer existed before the question was asked. Nothing existed here until you prompted it. So the rule is flat, and it will hold across every phrasing they try: when an option says the machine found the answer, strike it. The machine made the answer.

Speaks to an adult the system let down, not to a slow learner. Starts from what the reader already does well.

If you have kept away from these tools assuming there is a technical layer you missed, there isn’t one here. The picture you are probably carrying — the machine looks the answer up somewhere — is not a gap in your education. It is where nearly everyone starts, and it is plainly wrong.

Here is what actually happens. The model has read an enormous amount of text and kept none of it word for word. What it kept is a feel for how such writing tends to go — the same way you can tell a letter from your bank apart from a letter from a friend, instantly, without being able to name the rule you used.

Ask it something and it builds a reply out of that feel, one likely word after another. Nothing was stored. Nothing was found. It was made, while you waited.

A classmate, not a teacher. She says her first, wrong guess out loud, then changes her mind in the open.

I asked it for a poem about the monsoon, and it wrote me one. Then I asked again, and the poem was different.

That puzzled me for a while. If it had a poem kept somewhere, surely the same one would come back both times? So where was the first poem, before I asked for it?

I think the answer is: nowhere. It wasn’t anywhere at all. The machine hasn’t kept poems — it has kept a sense of how poems tend to go, the way I can hear that a line has one sound too many without stopping to count. When I ask, it makes one, one likely word after the next.

Which means the poem I got is not a poem it had. It is a poem it did. I am still deciding how I feel about that.

Ask it for a poem. It does not go and find one, because there is no poem to find — not in a drawer, not in a list, nowhere. What it has is a feel for how poems go. So it builds you one, word by word, while you wait. Ask twice and you get two poems. Made things vary.

A generative model learns exactly that kind of feel, built from everything it trained on — not stored answers, but a statistical sense of what tends to follow what. Give it a prompt, and it draws one plausible reply from that sense, built fresh, word by word. Nothing is fetched from storage — every output is built at the moment you ask for it.

What the model holds is a learned distribution over its training corpus — not an index into it. Generation is a sampling operation: given a prompt, it draws one plausible completion. The output is constructed, and the construction is stochastic. That is the whole reason one prompt admits many equally valid answers, and why plausibility is not truth.