Hacker News new | ask | show | jobs
by ai4ever 685 days ago
whats also puzzling is giving away the foundation model

the llamas will be used to pollute the social media reducing their value to advertizers

2 comments

I am surprised to see this: https://www.404media.co/where-facebooks-ai-slop-comes-from/

dont know how valid this is, but encouraging ai-slop as a business strategy to keep sticky customers !

I wonder if they have a watermark built in that allows them to detect text generated by Llama easily.

https://www.youtube.com/watch?v=XZJc1p6RE78

https://arxiv.org/pdf/2301.10226

I'd be suprised if they weren't making sure that Facebook could recognize content made with their tooling. Watermarks will always be on the honors system though, people could always remove them before posting.
For all the talk about safety it's surprising to me how little watermarking is going on, these AI producers are content with generating unmarked garbage.

I figure there's a lot you could do with interwoven zerowidth space characters and other unicode tricks (using other code blocks for otherwise normal characters like you would when spoofing a url) - sure it would be easy to write software to normalize it back to ascii but at least that requires intent to deceive - we could even write laws against stripping encoding schemes meant to identify automated content

Prior art is that time Genius encoded "red handed" into morse code by way of alternating apostrophes to catch Google copying their lyrics: https://gizmodo.com/genius-claims-it-busted-google-stealing-...

If I'm not mistaken, didn't the major LLM companies try to band together for watermarking type features only to walk it back later and admit it wouldn't work?