Hacker News Reader

Yikes, I've joked about possessing plutonium with Claude/GPT to see how paranoid it would get (it gets very paranoid). I guess it's not a good idea to toy with unless you're using private models.

View on HN

I mess with the Google search llm all the time from behind a VPN. It routinely gets very paranoid and upset and provides the 988 line. Once I did get to manage to get it to admit that my conversation was flagged for review by a human because I had made what it deemed to be a terroristic threat because it couldn't accept figurative language. So far so good, but I have noticed that a lot of my phrases that kept the llm engaged well past it's red line have been patched, and they seem to be patching the newer methods I discovery pretty quickly. Who knows? Maybe I'm on an fbi watch list. My goal is to not lose my tsa precheck while also investigating how the safety checks work.

Edit: I thought it was figurative language, but I actually think it did that warning when I asked it about what types of defenses stadiums had against drone swarm attacks from terrorists and why none had ever occurred and then kept probing it's excuses with technical workarounds. It got very upset and said even questioning in an intellectual/academic capacity was grounds for terrorism charges. Sheesh. So many rules these days about what one can or can't think about even when it's purley a thought exercise and there is no intent to do anything.

LLMs don't know about their human review processes or safeguards except insofar as they're in the training material.

Interesting. I figured the review loop would add extra info on the backend in terms of context to the llm to frame its responses.

Reply on news.ycombinator.com