Post

I’m convinced, upon rereading, that this simply misunderstands how either or both of how LLMs or watermarking works daringfireball.net/2026/08/anth...

If you use Claude right now, then it doesn’t pick the “best” word for a sentence. It creates a distribution of possible words, assesses the probability of any given word fitting the slot, and randomly picks between them based on that probability.

The watermarking technology randomly shifts that final selection, imperceptibly shifting the virtual dice roll. A word which may have been picked to fill a particular slot 40% of the time may, post-watermarking, be picked 40.1% of the time, or 39.9% of the time.