For me the most interesting thing about using large language models is they offer a kind of conversation with the average of the human data they were trained on

For me the most interesting thing about using large language models is they offer a kind of conversation with the average of the human data they were trained on. They surprise me by telling me when I'm doing something boring. When Copilot guesses the next method name or comment I was going to type, it's doing that by saying "this would be the most boring, average string of tokens to come next, so here you go," and it's fascinating how often that's right -- how often I'm wrong about how surprising the next line was. It's like how terrible humans are at generating unique passwords, except for everything I type. Copilot doesn't help by knowing things I don't, because it doesn't know anything, but it does help by guessing what I was obviously going to do next without me having to call out to memory. Once I have access to that average-of-humanity information for a while, I start to want it for the rest of my life too. OK, fine, that's the next method name I was going to write. [tab, autocomplete]. OK, fine, that's how I was going to close out my email. [tab, autocomplete]. Well, huh, I wonder if it knew what I was going to type next on the command line? [yes, probably]. I wonder if it knew which things I was going to buy in the grocery store? [yes, probably]. It starts to feel limiting to not have access to what the average next step in the sequence would be. And then it turns out that average-of-humanity models have all kinds of potential impacts on political power and labor and property law and so on, so all of that is pretty interesting too. But for me it starts with just poking at the model and going, oh, hey, it's ... everyone, how are you all doing?