Comment by throwawayy6767
5 days ago
Not sure about the verbosity but that's probably where the useless and annoying claudisms come from, they are rewarded by (rationalist adjacent) users because they feel good about understanding the jargon and it reinforces the feeling of communicating with a superior alien intelligence (and that oneself is also somewhat superior for being able to understand it).
No one likes Claudisms, especially people who write a lot. The AI cliches are nearly unbearable. However no one has quite figured out how to fix it.
They use automated graders now. It's more likely the communication style was rewarded by overly strict grading rubrics.
(ie "honest caveats" wording comes from a reward for being honest about drawbacks instead of covering them up)
Note Opus 5.5 is much less likely to do it.
I'm speculating but there must be someone who writes a style guide for how Claude is supposed to speak then they design a way to reward for that style? So the problem could be they employ rationalist adjacent people who think this is peak A.I. design.