> the space of "the hard stuff" is shrinking over time

That's theoretically true of programming in general - and has been true of programming since programming started. Network programming used to be super specialized, but now the standard for applications is distributed over the web. Graphics programming ability used to be rare, but now it's strange to see an app without a GUI. Yet programming takes longer to learn than it used to - precisely because the state of the art has advanced so much.

I've worked with a handful of people who thought they could just use the libraries but didn't understand what they were doing or why they worked - people who could put together fancy UIs in, say, jQuery and such - and inevitably, they would find themselves hopelessly lost because they didn't actually understand what asynchronous callbacks were and couldn't figure out why, when they stepped through their programs with a debugger, the debugger kept "skipping over" their callback function.

This is roughly what leads me to agree w/ the GP. The people who have the training to understand AI at a fundamental level will have transferable skills that will give them a leg up on whatever the next "hard stuff" might be, even if it's not in the area we currently refer to as AI.

I remember when the database guys were the high priests of software development and did their secret performance rituals in secret. Nowadays 99% of developers use databases without having a clue how they work and they are fine.

I expect the same for AI. 99% of use cases will be commoditized and easily accessible for devs and only a very small number of people who understand it in depth will be needed. You already can do a lot of cool stuff by copying code with some tweaks and I see this trend only continuing.