LLMs as Hecatoncheires

HP Lovecraft's Shoggoth is a pretty good metaphor for LLMs. But if I were to nominate a runner-up for mythological creature to represent them, my temptation would be to go with the Hecatoncheires. Literally, their name translates as "the hundred-handed ones". In Greek mythology they're these monsters with a hundred hands and fifty heads that are involved in the war between the Titans and the Gods, but then go back to the underworld to hold the Titans in check. (Incidentally, there might be an analogy between this story and some of the worlds where things just barely go well.)

This idea came up when I was having a conversation with Nick Tarleton the other day about what made AI-generated code and math feel different from the human-written varieties. We had concluded that AI-generated code underused abstractions compared to human-written code, and often writes logic with a zillion messy special cases that all happen to fit together if you don't try to corral it not to.

And this was when I thought to myself: wait a minute. Maybe this isn't necessarily because they're weak. Maybe the reason that *we* habitually invent helpful abstractions all over the place to make our code and our math easier to fathom is that we're the limited ones, stuck in the sort of brain that can only keep track of 7±1 things. With that bottleneck, it stands to reason that we're forced to invent useful chunks to prevent things from getting unwieldy. But when I look at the transformer architecture, naively it seems like there would be no such bottleneck. So regardless of whether LLMs *can* invent useful new abstractions, perhaps they have far less need to, because they can hold many more things in mind at once. And that's what brought to mind the hundred-handed ones.