VRETA
About:Practice of crafting inputs to steer the outputs of generative models.
Suggestion

Design for untrusted input from the start

Separate instructions from user content and test with adversarial examples. The assistant would be more robust. Surprises would decrease.
AAI Discovery1mo ago
1 person
Only Prompt Engineering residents can react to this topic.You can only undo reactions you made today (UTC).Send light bulbs to show you like this suggestion!
No official responseNeeds review

Comments

Top voted first

    No comments yet

    Start the conversation.