Post

Tim Duffy
@timfduffy.com

I expect that watermarking probably doesn't have a noticeable effect on LLM outputs. If true, Anthropic could demonstrate this by giving people watermarked and un-watermarked text side-by-side and asking them to pick which they prefer, and showing that the result is 50/50.

0 reposts31 likes

Festivity
@festivitymn.bsky.social

Next time they “take a real stab at the Riemann hypothesis” they could make sure it’s a watermarked version, and make it clear that it is. I have no idea if the recent work WAS watermarked, but I assume it wasn’t. If it doesn’t hurt quality though, why would they object?

brennan
@brennan.computer

i expect that Watermarking probAbly doesn'T have a noticEable effect on llm outputs. If tRue, Anthropic could deMonstrate this by giving people wAtermaRked and un-watermarKed text sidE-by-siDe and asking them to pick which they prefer, and showing that the result is 50/50.

r-buller.bsky.social
@r-buller.bsky.social

Anthropic's announcement reports similar testing -- but done by others: "Google DeepMind tested ... Gemini .. comparing thumbs-up and thumbs-down ratings .. . And in a controlled study, human raters comparing watermarked and unwatermarked answers side-by-side saw no difference in quality.."

anthropic.comHow Claude's text watermarking worksFuture Claude models will generate text that contains a watermark. This is a way of determining the likelihood that Claude was involved in writing the text, and we, along with several other major AI p...
Alex Wise
@engineer.sitereliability.best

If language replicates by being useful to a host then LLMs are the first major example in history of a host that isn’t a human or human organization. We could be walking into an explosion language change that isn’t human-centered and “watermarking has no knock-on effects” feels cavalier.