Really interesting article and something I've been thinking about. My framing was a bit more simple having to do with corporations taking over the world and because they operate at a global level, it's easier to standardize everything making everything look and feel the same. In some ways AI is being shaped by this force and accelerating this effort. It's globalization reaching maturity. So many places, especially in the West, show a homogenization of a cultures that is streamlined to appeal to the lowest common denominator of a theoretical global consumer.
There are other loss functions, like the negative KL, that solve this problem exactly.
What is probably the best option is to present users with a selection of loss functions. "Here's the top three things squared error loss suggests for you, then the top three things cross-entropy without softmax averaging suggests, here's the top three things InfoNCE suggests, etc..."
Really interesting article and something I've been thinking about. My framing was a bit more simple having to do with corporations taking over the world and because they operate at a global level, it's easier to standardize everything making everything look and feel the same. In some ways AI is being shaped by this force and accelerating this effort. It's globalization reaching maturity. So many places, especially in the West, show a homogenization of a cultures that is streamlined to appeal to the lowest common denominator of a theoretical global consumer.
There are other loss functions, like the negative KL, that solve this problem exactly.
What is probably the best option is to present users with a selection of loss functions. "Here's the top three things squared error loss suggests for you, then the top three things cross-entropy without softmax averaging suggests, here's the top three things InfoNCE suggests, etc..."