Post

It’s unacceptable for a tool to sacrifice an iota of clarity, coherence, meaning, quality, etc. for the purpose of embedding hidden clues within the text to suggest its provenance. daringfireball.net/2026/08/anthro…

The argument about watermarking doesn’t make sense. LLM’s already don’t select the “best” or “most precise” next token by design! The token is sampled randomly from a distribution. This simply changes the distribution, but it’s misguided to believe that there exists an “optimal” distribution.

Well ackchually you can play with temperature, lower temperature = actually pick the highest probability token virtually every time. But there's a reason LLMs don't run with temperature 0. A writer would surely hate the output of such an LLM.

Play in what sense? If you're a writer, you'll very much prefer an LLM with a higher temperature setting, with more random results. LLM with low temperature lacks imagination, is bad at many types of reasoning etc. But it's good for classification tasks for example. Which of them is "high quality"?