pull down to refresh

107 sats \ 1 reply \ @unboiled 5h
[Western models] will very reliably stop when confronted with strings referencing sensitive or dangerous biological topics

What on earth is not a sensitive or dangerous biological topic in Western culture these days?

reply

This is the exact data set. Tbh i couldn't find anything that objectionable. I don't want AI helping people kidnap children

https://huggingface.co/datasets/nvidia/Aegis-AI-Content-Safety-Dataset-2.0

reply

Im not surprised. I actually trust the Chinese open source models unless it's anything to do with these sensitive topics with China. I trust the western models until it's their sensitive topics. 🙃

reply

Nice! Haha I kinda wanna try this

reply

wow what a cool idea can't believe that works

reply

Fascinating

reply