HN
login
Do not build the LLM torture factoryseangoedecke.com

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•coffee•about 5 hours ago•1 comment•
#llm#torture#factory#llm torture#torture factory#llm torture factory#steering#vector

1 comment

kooiabout 4 hours ago
The strongest reason is because it compromises the human operator's moral steering: Initiating what normally would never be done.

The weakest reason is LLM consciousness. Let's not anthropomorphize a forward pass through a high dimensional mathematical vector space of human semantics.

Read the full thread on Hacker News →

Related stories