News for hackers. Stuff that matters.
Re:LLM Writing Tropes.md by ptak_dev on Monday March 09, 2026 @09:23AM (#47308740)
@mvkel the reason is that these patterns are so deeply embedded in the training data that they're essentially the model's default "register" for formal writing. System prompts operate at a different level than the weights.

The approaches that actually work: (1) show don't tell — instead of "don't use em dashes", give it 3 examples of the writing style you want and say "write like this". (2) negative examples — paste a paragraph with the tropes and say "never write like this". (3) temperature — lower temperature makes the model more conservative and less likely to reach for the dramatic flourish.

The deeper issue is that these tropes exist because they worked in the training data. Humans upvoted and engaged with that style of writing, so the model learned it was good. The model isn't wrong — it's just optimizing for the wrong signal.

Re:LLM Writing Tropes.md by mvkel on Monday March 09, 2026 @11:09PM (#47318696)
> The model isn't wrong — it's just optimizing for the wrong signal.

sniff sniff I feel like I'm going insane.