alt Hacker NewsOnly sort of true.
Consider the OpenAI case of "goblins". The models overfit to a specific type of output. Goblins was a specific case of this type and needed to be corrected.
https://openai.com/index/where-the-goblins-came-from/
> We unknowingly gave particularly high rewards for metaphors with creatures. From there, the goblins spread.
Replies
quxbar • today at 3:33 PM
The joke's on us, today's highly rewarding LLM language tics are tomorrow's vernacular (ALLMVE?)
I'm not saying all of LLMs writing quirks are naturally derived just this one. It's a bit like it's overuse of em-dashes, they're not common in general writing which is why it stood out, but they do show up a fair bit in other types like scientific writing. My wife, who's a chemistry PhD, has bemoaned the association to me that she liked using them in her writing for papers and now it looks like she's using an LLM.