Oh wow, LLMs! The bane of my existence, sucking the joy out of being a programmer!
I am, politically, very suspicious of them. I wrote Large Language Models Should Have to Obey Copyright as an expression of my political views1. The AI Non-Economy: A Rant, is, correspondingly, an expression of my economic views.
But they seem, alas, to be here to say. And of course, I’ll therefore use them, in moderation. I’m not an ideological purist. Even if I think they weren’t worth it for society to build, that isn’t the same as thinking it wasn’t worth it for me to use---the cost-benefit analysis is just different at the individual level. And to the extent they’re immoral, I don’t take on that immorality by using them. Those that built them (and allowed them to be built) are responsible for the immoral acts of building them.
Does that mean programming is dead? Not really. LLMs in Programming is my overall review of them. Does that mean that programming, and especially my fave topic, programming language design, are going away? No, as I explain in Programming Languages and Type Safety in the Era of LLMs, what’s good for humans, what’s good ergonomics for humans, what’s good guardrails for humans, is all the more important now.
I have many more thoughts, philosophical and practical, about LLMs, but I haven’t written them up yet, so you’ll just have to wait.
Footnotes
-
I do want to add that I disagree with Cory Doctorow on this as a matter of semantics. He characterizes a language model as a list of statistical facts about text, rather than a copy of a text. My friends, what is a text but a list of facts about itself? “I won’t give you a copy of the word, but I will tell you the first letter is “g”, the fourth letter is “t”, and the middle two are “f” and “i” but not in that order”? ↩