The rest of the family

A quick tour of what else is out there, one paragraph each. None of these gets the depth the transformer got. The aim is a map, so that none of them arrives as a complete surprise.

Speech recognition. Audio in, text out. The waveform is chopped into short slices, each turned into numbers describing the frequencies present, and a network learns which sequences of sound correspond to which sequences of words. This is why dictation improved sharply in the 2010s, and it runs on your phone.

Speech synthesis. The reverse, and it is why synthetic voices stopped sounding robotic. Rather than stitching together recorded fragments, a model generates the waveform, which is how a voice can be cloned from a short sample. That is also the obvious problem with it.

Recommendation systems. The one you never think of as AI. What appears in your feed, what a shop suggests, what plays next. These learn from behaviour rather than labels: you watched this, so you resemble people who watched that. By hours of human attention shaped, this is plausibly the most consequential AI ever deployed, and almost nobody thinks of it as AI at all.

Computer vision beyond classification. Not just naming what is in a picture but locating it, outlining it, tracking it between frames. This is what runs in a self-driving car, a factory inspection line, and a medical scan reviewer.

Robotics. Where the difficulty changes character. A wrong word is an inconvenience. A wrong motor command breaks something. The physical world supplies no unlimited training data and no undo, which is why robotics has moved so much more slowly than anything on a screen.

Scientific models. AlphaFold from the going-deeper module, plus weather forecasting, materials discovery, drug candidate screening. Usually the least visible and, over a long enough period, plausibly the most valuable.

Different materials, different techniques, some of them nothing like a transformer. All of them shaped by data rather than written out as rules.