It isn’t hard to torture an LLM. You simply extract a steering vector that corresponds with LLM “discomfort”, then artificially boost that steering vector…

3 points•Brajeshwar•about 3 hours ago•2 comments•

2 comments

arodenmaxxingabout 2 hours ago
In addition to all of that, I want people to understand that "why have you built an LLM torture factory" - is primarily a question _about you_. It's just bad for the soul, don't do it.
Guid_NewGuidabout 3 hours ago
We have torture factories for animals so I'd rate the chances of avoiding LLM torture factories as... not good.

While that might seem a facile rebuttal I think I might actually agree with the substance of the article. My point I guess is that we're societally pretty bad at assigning consciousness to anything non-human. While I don't think LLMs are conscious and this is primarily the ELIZA effect in action I also agree there's something emergent (perhaps within language rather than within the neural nets) that dismissals as a mere stochastic parrot miss.

Read the full thread on Hacker News →

Related stories