3 ms·

Someone needs to train a model where untrusted input uses a completely different set of tokens so that it's entirely impossible for the model to confuse them wi

Someone needs to train a model where untrusted input uses a completely different set of tokens so that it's entirely impossible for the model to confuse them with instructions. I've never even seen that approach mentioned let alone implemented.

jorl17 8mo ago
Perhaps this is in line with what you had in mind? https://patents.google.com/patent/US12118471 https://patents.google.com/patent/US12118471
- LoganDark 8mo ago
  > The input is represented as tokens, wherein the trusted instructions and the untrusted instructions are represented using incompatible token sets. Yes, exactly!
  - aminerj 8mo ago
    [flagged]