I’m a computer-science professor at MIT and cofounder of Nectry.
My blog Structure and Guarantees considers how to build AI systems that we can actually understand and trust. The usual academic stuff is on my home page.
I work broadly on creating better abstractions for programming. The most-persistent theme of my work is formal methods, with proofs about programs that are credible with minimal trust requirements. We would rather not trust computer processors, compilers, operating systems, databases, cryptography, programming-language designs, individual applications, or theorem-provers themselves—and parts of my research have addressed removing trust in all of the above, including in integrated systems with proofs that cross such layers.
Yeah, the weights of different factors matter a lot for the consequences of changes in price levels! I think you’re probably right that we wouldn’t expect a 100X price improvement to spike usage downward, since the normal surface-level engineering considerations clearly get much more attractive.
I don’t know what the exact weights are in practice, and I agree that they matter for how much of a boost deep learning can get in decision processes, beyond what the engineering fundamentals imply. I’m making a case that the weighting is significant-enough that this factor is worth taking into account because it changes enough important decisions.
The challenge here is to consider both cases of top-down pushes by company management for everyone to use more AI, which may lead to misguided application of expensive solutions; to bottom-up efforts where engineers are picking the best tools they find available today on a problem-by-problem basis.