AI Reveals Secret Watermark Hidden in Generated Text! #shorts

Transcript

1. 0:01When When you generate a response When you give a when you give an LLM a prompt It's a probability calculator and it's
2. 0:15and it's calculating What's the next most probable token to to to be the answer, right? And then it just keeps giving you the next most
3. 0:24probable token. The way the entropic watermarking works is instead of delivering the most probable token They're delivering the second most
4. 0:36probable token. And so Because Because that's a mathematical thing You can analyze that. So you could look
5. 0:45at a piece of text and say, you know, I can look at this text and if this was generated with an LLM Um all of these words in here are the
6. 0:54second most probable words and therefore I can tell that this is AI generated. Um so that's that's kind of the
7. 1:01technique they did. I don't know if it's every token or every 10th token. I don't I don't quite know how it works, but
8. 1:06it's a it's a really interesting thing that That another another AI can look at that and figure out what the probability is
9. 1:13of these tokens in this order and recognize that oh those are all second order Or you know, second probability
10. 1:21Um tokens and that's how they're doing it. >> Watch the full replay at community.thesalon.ai.