I feel like better expertise targets tend to be more durable. In other words, rather than expert in a specific technology or technique, the best experts had the ability to develop expertise in any given technology within a space and often overlapped cursory knowledge with other spaces. I met plenty of graphics experts at Oculus and they didn’t care so much about whether it was D3D or Vulkan - originally they were D3D engineers for PCVR and then a good chunk of them shifted to Vulkan once the focus shifted to mobile VR. They just knew how those APIs mapped to the actual HW reality, how things connected, why things were slow, how to improve performance, etc. The mundane stuff of “what is the sequence of steps to do X in Vulkan” is answered by Google/StackOverflow (or even these days ChatGPT). Heck, a good chunk of them were creating their own new APIs. This isn’t unique to Meta by the way. It’s like engineers who say they’re “C” or “C++ experts”. With the exception of authors like Scott Meyers or people working on the C++ language spec who I think can truly maybe claim that title, the kind of thing that is called a “language X expert” is the kind of expertise that a good engineer should be able to develop in any language with 2-3 years of practice and proficient mastery within ~3-12 months because the true expertise is the underlying CS principles (at least for a family of languages - I’ve never done too much with non-Algol families so I don’t know how I’d fare there).
However, I do agree that generally graphics engineer is a niche limited to the few people working on gaming engines, VR R&D, or animation R&D. But those skills, at least today, are generally transferable to AI engineering because GPU compute plays such a huge role. There’s less graphics programming of course and the APIs for GPU compute are a bit different, but AFAIK many of the hardware concepts remain (e.g. wavefronts, how GPUs do threading, etc etc).