Paul Christiano thinks there's a 1/3 chance Tesla gets fully self-driving cars by 2024, and expects that conditional on that their market cap has probably >tripled to >$3T. that's pretty insane commercial value right there.
In scientific applications, one obvious thought is advances on AlphaFold that enable better drug design. I'm not a domain expert, but I think that might require significant improvements from AlphaFold v2-- moving beyond crystal structure prediction to in-solution structure prediction and to protein-protein interaction modeling.
I've heard from two casual programmer friends that AI programming assistants like Github Copilot are impressively good. They make it easier to write various finicky pieces of code, and help fix bugs. It seems to me like this could be really impactful if it turns out to help professional programmers; there's a lot of value to add, and potentially this could be turned towards AI programming itself...
Really interesting, thanks!