This is a really good description of the problem. I’ve been trying to use Claude to get familiar with the mechanics of a new codebase, and there have been so many moments where I’ve stopped after reading the same paragraph five times in a row and thought “Am I tired? Or stupid? Or is this codebase just wildly more complex than anything I’ve seen before?” before realizing that it’s just taken English and smushed it around like a ball of clay into some abstract sculpture that kind of evokes something from real life.
I think part of it might be an innate feature of LLMs, but Claude seems extra prone to it lately. I ran the same query about the same codebase with Codex, and it gave me an answer that was about 1/4 the length and made me realize that it really wasn’t all that complex.
If nothing else, it’s good training for my own writing. I’ve been working on making myself be more straightforward and concise, and Claude’s writing is a good example of how cleaner prose is a functional choice, not just a stylistic one.
Ironically, I've been already active on codex after trying it for the past 1.5 months, from Claude, and I find it to be similarly complex and confusing.
I think they all have the similar styles and tells. If I were to go to Claude, and use it now it would probably be clear for a little before reverting.
And I don't know why it feels to me like the language "drop off" happens after some time with the system. It makes me wonder if my account are getting silently degraded or sent to lower intelligence/lower priority queues after being a member for a while.