Post

i think the scummiest part of LLM codegen is that you never know if you're getting an amalgamation of some common idea it's been trained on a million copies of, or just getting One Guy's passion project with the identifiers changed a bit, a thing i have overheard happening more than once

you'd get this same sort of problem generating art or writing, of course, but a number of programmers do not seem to understand that the thing they do is a creative pursuit, as if the existence of a test suite makes it a hard science

so we land instead on this industry-wide authorship laundering, encouraged by having the output come from a natural language interface with a cutesy name that makes it seem like a Guy rather than a fuzzy database with all the attribution filed off

and this isn't about copyright or whatever, just about common human decency between people who Make Things. artists of all stripes seem to understand this pretty intuitively, that you give nods to those you borrow from, because... because that's why we do any of this, to share and inspire

and in fact the theft is arguably worse with programming — you can't share artwork without sharing its pixels, and you can't share writing without sharing its words, but you CAN share a program without sharing its code and in many cases that is the default thing that happens

so codegen LLMs /must/ mostly be trained on code that someone actively decided to share, that its authors wished to contribute to the commons! this industry was just recently fretting over how to support the open source projects it relies on, too. now it exploits them