alt Hacker NewsAgreed, and then the actual text of the post gets worse. It’s actually unreadable; I literally could not follow what the article was about (without putting in the effort to unravel the Claudisms, a little like unraveling Shakespeare or Finnegan’s Wake) and gave up.
It seems we are all (on HN) thoroughly convinced that the models have all gotten worse at writing as the version numbers have gone up and they are supposedly more powerful models. Sol 5.6 is nowhere near as bad as Opus 5 or even the 4 series but not great.
But looking back there was a time we were enamored with how they wrote. Perhaps it’s related to the newfound prohibitions on “write in this writer’s style”? I wonder if training the models to not be able to do that anymore has steered them towards not being able to write, at all, in any existing human good writing style and they are navigating the small pathway through all existing good human writing to whatever you want to call what they are doing now?
Replies
I'm willing to bet it's because model providers have gone all-in on programming and reasoning ability when training new models. It's where the money is, if nothing else, and the writing quality will have suffered as a knock on effect.
It was the newness and surprise that the quality was as good as it was, just like images from Stable Diffusion, etc. Now that we are used to it, we notice the uncanny valley speak.
> newfound prohibitions on “write in this writer’s style”
that explains a lot. but do you have a decent citation?