April 1, 2023•295 words
Fine, I'll elaborate.
The version of "AI Alignment" I'm talking about here is the kind that a bunch of big-shot internet celebrities and "research foundations" and such are "trying" to "solve", where we build an AI with arbitrary capabilities far greater than human, and then try to make it do what we want. Here's why it's stupid:
For any galaxy-brained constraint that we could place on an arbitrarily powerful AI, there exists an even more powerful AI which could develop an even more galaxy-brained way to evade the constraint.
There, I solved it. All those "research foundations" can give all that money back now. Maybe they could point it in a more useful direction, like figuring out why so many 1000IQ nerds spent so many years trying to come up with a way to make an object so immovable, even an unstoppable force can't move it.
More generally, AI alignment is a silly problem because if you accept the nerd theology that AI will inevitably become omnipotent and take over the entire universe, there's no silly human obstacle that can get in its way, short of the forcible destruction of all AI research worldwide (which actually may not be a bad idea, but mostly for different reasons).
So if you don't accept the nerd theology, then you agree that AI as a technology operates within a set of constraints. This is the way every other technology operates, from hammers to rockets to antibiotics. I think there's a lot of potential for discussion about where exactly these constraints may turn out to be, which is typical for a new technology, but I think it's a productive discussion to have, and likely far more fruitful than coming up with a bunch of thought experiments about machine-gods.