napping_penguin

3 ms·

1. 1.▲
   Prompt Privacy from LLMs(snwagh.com)

   4 points by napping_penguin 2mo ago|1 comments
2. 2.▲
   by napping_penguin 2mo ago

   I recently came across a really interesting piece of privacy technology. Suppose you have a model M and a prompt P. The technique allows you to create an obfuscated prompt Q such that: - M(Q) is nearly the same as M(P) - P is hard to rever
3. 3.▲
   by napping_penguin 2mo ago

   Ah, sloppy language on my part. It should be 5 of fewer questions. I'll fix it, thank you!
4. 4.▲
   by napping_penguin 2mo ago

   You are right, your argument establishes the lower bound: D(4) >= 5 (you will need at least 5 questions in the worst case). The interesting bit is to come up with a strategy to make 5 adaptive queries that will guarantee you know all the
5. 5.▲
   by napping_penguin 2mo ago

   Thank you for the feedback, I will address this in the post.
6. 6.▲
   by napping_penguin 2mo ago

   Good catch but it's intentional. The engine behind the first simulator for 4-cards is adversarial and hence appears a bit deterministic. - Adversarial engine: Will always respond with an answer to your query such that it maximizes the
7. 7.▲
   Can LLMs identify 16 cards in 45 bit-queries?(snwagh.com)

   9 points by napping_penguin 2mo ago|8 comments
8. 8.▲
   Show HN: Privacy-first work journal with no backend(scribe-notes.com)

   3 points by napping_penguin 1y ago|2 comments