What they all share

Six different families, several of which work nothing like each other. So what makes them one subject rather than six?

Not the architecture. A diffusion model and a transformer are built differently. Not the material. Waveforms, pixels, protein sequences, and clicks have nothing in common.

What they share is the move this journey started with.

In every one of these cases, somebody faced a problem where the rules could not be written down. Nobody can state the procedure for recognising a cat, or for which frequencies make a p sound, or for how a chain of amino acids will fold. People had tried, for decades, with real ingenuity, and the results were brittle in exactly the way the long winter showed.

So the approach changed. Instead of writing the rules, supply examples, or outcomes, or feedback, and let a training process shape a very large number of adjustable numbers until the behaviour comes out right.

That is the whole family resemblance. Not what they are made of. What replaced the rulebook.

And it explains the shape of what happened, which otherwise looks like a suspicious number of unrelated breakthroughs arriving at once. They were not unrelated. One approach became practical, once the data and hardware and methods were there, and it was then pointed at every problem that had been stuck for the same reason.

Which is also why the boundary of the word AI keeps moving. Every technique on the last slide was called AI while it was difficult. Some of them are now called speech recognition, or search ranking, or forecasting, because they work and have become ordinary. That is the usual fate of things in this field. Succeed completely enough, and you stop counting.