alt Hacker NewsLove 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.