It’s a strange feeling as a native English speaker to read the entire sentence and know every word, for the sentence to be entirely grammatically correct, and still have no idea what it’s trying to communicate.

Another problem is that it will open up all sorts of tangents about nits that it encountered, but it will often not tell you that it’s a nit or give you adequate context to realize that this paragraph is exceedingly low value until you’ve spent a bunch of time and energy trying to make sense of it.

I’m curious if anyone has any suggestions for promoting agents to improve their prose. I’ve had some okay results with “optimize for clarity, don’t dump every thought on me, treat my attention and focus as constrained resources, stay focused on the task at hand”.

I think it’s no longer speaking human language. It used to, when it was mostly or entirely trained with imitation learning of human-generated text. Now it’s increasingly trained with RLVR, and there’s far less optimization pressure on actually speaking English (or Punjabi or Mandarin). It still uses English words and syntax in its output, but the semantics are drifting more and more. It’s basically speaking in a code that superficially resembles English.