Anthropic shares more details about how Claude's new watermarks will work

001-_- about 7 hours ago 2 commentsRead Article on techcrunch.com

⚡ Community Insights

Discussion Sentiment

50% Positive

Analyzed from 144 words in the discussion.

Discussion (2 Comments)Read Original on HackerNews

voakbasda•about 3 hours ago

I cannot imagine a scenario where this does not significantly constrain in the output in ways that degrade the quality of results. It feels a bit like forcing a painter to work from a sketched drawing, rather than letting them work with a blank canvas. But you aren’t told what it looks like. How is this not viewed as a form of outright deception against the user?

This is not what I want. I cannot imagine this is what most users want. Instead, this seems to aim at placating the anti-AI crowds, as those are the folks I see benefiting from this the most.

I expect to be in charge of my AI. The more guardrails like this that are put in place, the more I will look to open models and harnesses that do not impose such bullshittery on me.

throwawayk7h•about 2 hours ago

My understanding of the watermark process is that it only kicks in when it's not the case that one particular token is overwhelmingly likely. In the case where many tokens are plausible, we already choose with PRNG -- the "watermark" is essentially a more easily reproduced PRNG, but it still looks like PRNG.