The kinds of models, beyond chat

Your photo app finds the beach without ever being told what a beach is. Your phone unlocks at your face and turns a voice note into text. A feed picks what you see next, and a chatbot drafts the letter you have been putting off. Behaviours this different look like different inventions. Under every one of them runs the machine you have already met: weighted sums stacked in layers, trained by computed blame. That page ended on the recipe refusing to change with size. It refuses to change with job, too. Whatever makes a photo sorter a different kind of thing from a chatbot, it is not the machinery.

It is the examples. Training turns a pile of solved cases into behaviour, and every pile you have met shared a shape worth naming: each was a pile of pairs. A message beside its junk verdict. A photo beside its tag. Something on the left, and the right answer for it on the right. A finished model is a machine for turning left-column things into right-column things, and that one fact organises the whole zoo. You never need to see inside a model to say what kind it is, which is lucky, since nobody can read the inside anyway. You watch what goes in and what comes out.

Judging, and carrying across

The oldest families boil something large down to a verdict. Image in, label out: the photo sorter, the face unlock, the system that flags a tumour on a scan. Mail in, junk-or-not out. These are recognisers, classifiers in the field's own word, and their signature is that the output is a choice from a menu fixed before training began. The menu is also a wall. A sorter trained without a sunset label cannot find you a sunset, however orange the sky.

Transcribers and translators are gentler with the input: sequence in, sequence out. Speech goes in and written words come out. A sentence enters in one language and leaves in another. Nothing is judged and nothing is boiled away. The content crosses a border and changes its clothes. Their piles hold recordings beside transcripts, sentences beside their translations, and the output is assembled to fit the input rather than picked from a menu.

The pile that writes itself

The most consequential family is the one the label has nearly slid off. History in, ranking out: what you and millions of people like you watched, tapped, skipped, bought and abandoned goes in, and out comes an ordering of what to show next. Every feed is one of these, and so are the shop's suggestions, the queued songs, the video that plays next. They are called recommenders, and what sets them apart is where their pile comes from.

A recogniser's photographs had to be tagged by people paid to tag them. A transcriber's recordings needed every word typed out by hand. A recommender's pile writes itself: every tap is a solved case saying more of this, every skip one saying less, supplied free, in the ordinary run of using the product, by everyone at once. No other family gets its examples at that price. That is why recommenders arrived everywhere first and drew so little notice doing it, why fresh versions are retrained from the swelling pile as routine, and why the word AI slid off the family years ago, the way the label always slides off software that works. The software that chooses what billions of people will look at next is the software the word no longer points at.

One further family does quiet work underneath several of the others and needs only a sentence here: an embedder takes anything at all — a word, a photo, a song — and puts out coordinates, a position on a learned map where similar things sit near each other, the trick called embeddings that serves inside search and recommendation alike and that how machines represent meaning is entirely about.

The same pile, read backwards

Then the family everyone now means by the word. A generator takes a request typed in words, called a prompt, and puts out a thing that did not exist before: a paragraph, a picture, a voice, a video clip. Small in, large out, the recognisers' arrow reversed. And the reversal goes further than size, because nothing about a pile of pairs says which column is the question. Pictures paired with descriptions, read in one direction, train a machine that names what it sees. Read in the other direction, the same pairs train a machine that paints what it is told. What look like different species of AI are one trick pointed opposite ways.

The backward direction is harder, and the training loop you know says why. A guessed label can be scored against the one right answer stored in the pile. A guessed picture cannot, because a description has countless faithful pictures, and punishing every difference from the single stored example punishes good guesses along with bad. Machines that name what they see won their famous exam in 2012. A machine that paints what it is told became something anyone could download and run in 2022. Same pairs, opposite direction, ten years between them.

The chatbot belongs to this family, and the frame holds even there. A language model's pile is text itself, every stretch of it paired with the word that came next, and a page of its own traces what reading that pile forward builds.

Walls that were never load-bearing

The newest development is not another family. It is the walls between families coming down. By the first layer everything is numbers anyway: the photo a grid of brightness values, the sound and the sentence encodings of their own. So nothing stops one model being trained on several kinds of pair at once, and a model trained that way is called multimodal. Show one a photo of your fridge and ask it, in words, what to cook. On this page's map it is a recogniser and a generator wearing one skin. The families were never different machines, only different piles, and piles can be mixed.

Reading a product from outside

What you actually use is a stack, and the stack usually holds a small coalition of families. Ask a voice assistant for tomorrow's weather and your question travels a relay: a transcriber turns your speech into text, a generator writes the reply, and a third model reads the reply aloud in a manufactured voice. No single link in the relay is the AI. The product is the coalition, plus plumbing.

An earlier page gave you a question that always has an answer: what did it learn, and from what? You now hold the sharper version. What was paired with what? Products you have never seen will keep arriving, wearing labels that tell you nothing. The pairs will place every one of them.