Post #2702975

2026-04-03 08:24 UTC

@agersant@mastodon.gamedev.place @Forbearance@mastodon.xyz What agersant is saying about licenses is a pretty huge piece of this. To answer your question directly ("are they always plagiarizing, or only sometimes?"), the short answer is yes, because fundamentally a machine learning algorithm can only output as much as it is input. But there's a layer of abstraction with LLMs, where it isn't just raw blocks of words being processed, it's relations between words.

Replies (1)

@protowlf@mastodon.gamedev.place 2026-04-03 08:35

@agersant@mastodon.gamedev.place @Forbearance@mastodon.xyz This is a simplification, but IMO it hits the fundamental point: when a code LLM "invents" something new, it only appears that way because the final words didn't happen to line up with the relations between words it did necessarily plagiarize to produce its output. But that begs another question: if an LLM is "only" scraping my code to change weights in a model, is that still stealing? I mean, it didn't take my code blocks, it just took the relations between words.

Open ##2702976