Again, this is an argument that I believed less after looking into the details, because right now it’s pretty difficult to throw more compute at neural networks at runtime.
Which is not to say that it’s a bad argument, the differences in compute-scalability between humans and AIs are clearly important. But I’m confused about the structure of your argument that knowing more details will predictably update me in a certain direction.
[Yudkowsky][15:21]
I suppose the genericized version of my actual response to that would be, “architectures that have a harder time eating more compute are architectures which, for this very reason, are liable to need better versions invented of them, and this in particular seems like something that plausibly happens before scaling to general intelligence is practically possible”
This exchange feels like a concept I’m supposed to propagate through my models of “what AGI will actually look like.”
My attempt to rephrase this in my own words is “a powerful optimizer should have some way of transforming more thinking into better outcomes.”
This exchange feels like a concept I’m supposed to propagate through my models of “what AGI will actually look like.”
My attempt to rephrase this in my own words is “a powerful optimizer should have some way of transforming more thinking into better outcomes.”