If you try to use this definition, and other people use AI alignment to refer to things that they think are relevant to whether advanced AI produces good outcomes, you can’t really object as a matter of linguistics. I have no sympathy and am annoyed at the complaining. You can see the last LW discussion of this here.
I think “intent alignment” is a more useful concept, and complaining about how intent alignment isn’t the same as existential risk from AI reflects a misunderstanding of how language works and what its purpose is. I think MIRI people are basically doomed to be perpetually frustrated because they are trying to use language as a tool to shape discourse in a way that’s just not feasible. (The whole point of the choice of “alignment” was to sound good and bring in a bunch of connotations that aren’t the same as the proposed meaning of the term.)
Alignment seems to mean something that’s alternative to control because of the ordinary meanings of the words. If you stipulate that you are using “alignment” to mean something else, you are going to face a perpetual uphill battle. If you define a dog as a domestic pet that either barks or meows, expect people to be confused.
Eliezer defined AI alignment as “the overarching research topic of how to develop sufficiently advanced machine intelligences such that running them produces good outcomes in the real world.”
If you try to use this definition, and other people use AI alignment to refer to things that they think are relevant to whether advanced AI produces good outcomes, you can’t really object as a matter of linguistics. I have no sympathy and am annoyed at the complaining. You can see the last LW discussion of this here.
I think “intent alignment” is a more useful concept, and complaining about how intent alignment isn’t the same as existential risk from AI reflects a misunderstanding of how language works and what its purpose is. I think MIRI people are basically doomed to be perpetually frustrated because they are trying to use language as a tool to shape discourse in a way that’s just not feasible. (The whole point of the choice of “alignment” was to sound good and bring in a bunch of connotations that aren’t the same as the proposed meaning of the term.)
Alignment seems to mean something that’s alternative to control because of the ordinary meanings of the words. If you stipulate that you are using “alignment” to mean something else, you are going to face a perpetual uphill battle. If you define a dog as a domestic pet that either barks or meows, expect people to be confused.
Example:
https://www.lesswrong.com/posts/ykccy6LXfmTjZcp6S/separating-the-control-problem-from-the-alignment-problem