The Wider Family

Everything so far followed a single thread: machines that work with language. That was a deliberate narrowing, and it leaves out a great deal.

Set a picture generator beside a language model and the contrast is instructive. One of them commits to a word and moves on, unable to revisit it. The other starts with something closer to static and refines the whole image repeatedly, every part revisable until the end. Neither approach is more advanced. They are answers to different questions about what it means to produce something.

There are systems that take in pictures and sound and text together, systems that control physical machinery, systems that predict how proteins fold. Some share almost nothing with each other in structure.

What they do share is worth stating plainly, because it survives whatever gets built next. Every one of them learns from examples rather than instructions. Every one ends up as adjustable numbers nobody can read. Every one is as good as what it was shown and no better. Understanding one deeply, which is what all of this has been, means the next thing is not a mystery. It is a variation on something now familiar.