2026-09-27 11:36 UTC
@mattblaze@federate.social I know, not a real quote, but everything about it is painful and wrong.
1. In order to develop a useful tool, I must understand the specific use it is being developed for. Unless the specific use is "kill people", "murder" should not be a function that ever gets optimized to the point of effectiveness. If someone thinks murder or manslaughter is an acceptable goal or even an acceptable byproduct, I want to introduce them to the legal system and/ or Arkham Asylum, asap.
2. What is an "AI robot"? A virtual agent that operates entirely in software and has no physical end effectors? Or an autonomous robot that incorporates AI algorithms in its sensory or decision making or control processes? Anyone saying this is not someone with sufficient expertise to be making pronouncements about what is or is not necessary.
Not even getting into the whole "it is ridiculously easy to turn a perfectly safe consumer robot into a killing machine without an LLM - all you need is a sign error in an optimization function controlling the equivalent of 'acceleration in direction of bystander'" thing. Knowing that is not a reason to build less safe things! It's a reason to make absolutely sure you didn't make that kind of mistake!
Replies (0)
No replies.