That's what I'm saying. The fact that you didn't see "responsible AI" as something to question to the same degree is exactly the point I was making.
That you don't realize that "neutrality" doesn't exist in a conversation about responsible AI is the point that sgift's comment was making. There is no purpose to going into a conversation about responsible AI trying to be fully neutral about what good and bad is. The word "responsible" implies that there is such a thing as a good and bad outcome, and that the good outcomes are preferable to the bad ones.
But to restate GP's point, any conversation about user agency, privacy, equity, responsiblity, bias, accuracy, etc... all of it assumes an opinionated philosophy about what's good and bad -- and in fact, it is desirable to have those discussions working from the assumption that some things are good and some things are bad. It is impossible for any of that conversation to be neutral, nor should neutrality be a goal -- because a conversation about responsible AI is a conversation about how to influence and create those preferred world states, and how to bias an AI to create those world states.
There is no such thing as a "neutral" aligned AI. AI alignment is the process of biasing an AI towards particular responses.
----
And again, I just want to point out, these types of conversations never crop up when we're talking about stuff like user agency and privacy. When I say "people should have privacy", nobody asks me to define ethics from first principles.
And I mean, people can say "well, we just want to approach first principles in general, this isn't about minorities" -- but people can also read the thread on this page and they can see that all of this conversation spiraled out of someone complaining that Mozilla was calling out models that reinforced existing biases against minorities. That was the context.
There is no sibling thread here where someone complained about the dichotomy between safety and agency. And people can take from that what they want, but it's a pattern on HN.
That's what I'm saying. The fact that you didn't see "responsible AI" as something to question to the same degree is exactly the point I was making.
That you don't realize that "neutrality" doesn't exist in a conversation about responsible AI is the point that sgift's comment was making. There is no purpose to going into a conversation about responsible AI trying to be fully neutral about what good and bad is. The word "responsible" implies that there is such a thing as a good and bad outcome, and that the good outcomes are preferable to the bad ones.
But to restate GP's point, any conversation about user agency, privacy, equity, responsiblity, bias, accuracy, etc... all of it assumes an opinionated philosophy about what's good and bad -- and in fact, it is desirable to have those discussions working from the assumption that some things are good and some things are bad. It is impossible for any of that conversation to be neutral, nor should neutrality be a goal -- because a conversation about responsible AI is a conversation about how to influence and create those preferred world states, and how to bias an AI to create those world states.
There is no such thing as a "neutral" aligned AI. AI alignment is the process of biasing an AI towards particular responses.
----
And again, I just want to point out, these types of conversations never crop up when we're talking about stuff like user agency and privacy. When I say "people should have privacy", nobody asks me to define ethics from first principles.
And I mean, people can say "well, we just want to approach first principles in general, this isn't about minorities" -- but people can also read the thread on this page and they can see that all of this conversation spiraled out of someone complaining that Mozilla was calling out models that reinforced existing biases against minorities. That was the context.
There is no sibling thread here where someone complained about the dichotomy between safety and agency. And people can take from that what they want, but it's a pattern on HN.