I'm tempted to apply but I'm guessing the work I want to do is too far afield of traditional ML work, maybe someone can tell me if that's wrong.

The work I want to do is not in iterations on algos, like I've been working on with CV problems, but better defining intelligence and building testing mechanisms for it along the lines of the work of the Anytime Intelligence Test [1].

Too little is being done on baselining real world intelligence and so what we get are closed world, fully observable tests, which are impressive, but aren't really functionally useful when we're trying to build or test AI's generalization capabilities.