4 ms·
Alignment is with the user of the LLM not to some fuzzy interpretation of human rights. So solving alignment for the DoW is just "don't refuse to bomb people wh
by vasco 8mo ago
Alignment is with the user of the LLM not to some fuzzy interpretation of human rights. So solving alignment for the DoW is just "don't refuse to bomb people when I ask you".
- _heimdall 8mo agoThat's absolutely not the definition people use for alignment. Safety discussions often circle around alignment because they are worried about AI doing things that are bad for humanity as a whole, not because it goes off track from any one user's goal. That would be terrible for safety if alignment meant I could ask to hack tha TSA and the LLM would do it. Ignoring the definition, what would be required for individual alignment is exactly the same as collective alignment. The only difference is the goals and who writes them, for the LLM it is being somehow forced to follow those rules no matter what.
- vasco 8mo agoThat's safety, not alignment. Alignment is necessarily to the user.