Claude has become noticeably, painfully worse at writing in the last six months. At this point it’s practically useless for anything except code.
Claude has become noticeably, painfully worse at writing in the last six months. At this point it’s practically useless for anything except code.
I did notice after the fingerprinting update a marked uptake in strange language in responses. Specifically if I ask it to do something sometimes it will replace some of my request language with synonyms that don't actually make any sense. Like my request was fed through google translate twice
Watermarking doesn’t affect writing quality (on average) as long as the implementation is correct.
Love your "on average" qualification. Like the cartoon where the water temperature is fine on average, with one bucket boiling and the other ice.
The interesting question is how to define 'average'. Over what probability distribution?
Hey, anyone remember this from earlier in the week? https://daringfireball.net/2026/08/anthropics_watermark_text...
The qualifier is there because it changes the outputs, so it’s necessarily true that some outputs will be worse.
But it’s just as likely to make an output better.
Take the example from the article. He complains that watermarking might sometimes, for example, choose to say “bananas” over “pineapples” because only the former is on the green list, potentially making an output less precise. But 1. It could do that regardless of watermarking since the model is probabilistic, and 2. The more accurate word choice of “pineapples” is equally likely to be on the green list instead, further increasing its likelihood!
Overall, the article is pretty silly because he’s complaining about the possibility of Claude not always choosing the most “optimal” token, even though LLMs are probabilistic so that will happen anyways.
It does. There is a marked difference in certain word choices that sometime stick out like a sore thumb.
It's not good for writing code either, despite the many claims to the contrary. At best you come out even on speed as you have to review everything it does. At worst it actually slows you down as you clean up its mess.
Reverting to Opus 4.6 is much better than later models, though that is still full of annoying tics as well.
Perhaps a one-trick pony is all we need.
[dead]