Over the last week, since Anthropic's announcement about watermarking Claude's text, I've seen a lot of questions being raised about the effect of watermarking on the quality of generated text, as well as debates about the efficacy of detectors. I built this interactive demo where you can actually try 3 different watermarking schemes, along with their detectors. I found it helpful to experientially understand how these watermarks work, and I hope you find it interesting too.
You note that your biggest takeaway "was how elegant the core watermarking schemes are"—and I agree. I've engaged in R&D surrounding similar schemes for nested data in the past, but the constraints of hiding signal in plaintext present an additional conceptual challenge and it's interesting to see how various implementations meet that challenge.
I made this to remove watermarks. I tested it with your algorithms and it does a decent job except for secret race - for some reason this one did not seem to work. I am not claiming this to be a solution to the watermarking - it definitely degrades the quality of the output as my synonym lists are not contextual but it is a POC and works https://yoshimi.space/bleach/
3 comments
[ 0.23 ms ] story [ 13.4 ms ] threadYou note that your biggest takeaway "was how elegant the core watermarking schemes are"—and I agree. I've engaged in R&D surrounding similar schemes for nested data in the past, but the constraints of hiding signal in plaintext present an additional conceptual challenge and it's interesting to see how various implementations meet that challenge.