Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

>So it nudges certain words to be paired together slightly more than random and you can mathematically derive with some level of certainty whether an output or even a section of an output was generated by the LLM.

hah, every single LLM already watermarks its output by starting the second paragraph with "It is important/essential to remember that..." followed by inane gibberish, no matter what question you ask.



I've always felt you'd be able to tell someone uses Reddit because they'll reply to a comment starting the sentence with "The problem is that..."

Now LLMs are trained on Reddit users.


Sounds like LLMs are trained on my posts because i tend to use both of those phrases :-P




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: