divyanshusingh

5 ms·

1. 1.▲
   Glyph: A sub-millisecond prompt-injection detector(github.com)

   6 points by divyanshusingh 5mo ago|0 comments
2. 2.▲
   No Free Lunch with Guardrails: Evaluating LLM Safety Tradeoffs(arxiv.org)

   3 points by divyanshusingh 1y ago|1 comments
3. 3.▲
   by divyanshusingh 1y ago

   I’ve been working on building AI safety solutions for a while now, and one recurring theme keeps coming up: there's always a tradeoff. You can’t have perfect safety, perfect usability, and perfect performance all at once. So we wrote a
4. 4.▲
   by divyanshusingh 2y ago

   Agreed, on something like battery of tests for `Safety4all` but it'll be too generic not for an enterprise use case.
5. 5.▲
   by divyanshusingh 2y ago

   Ig, it's there in the paper: training or finetuning the Deberta-V3