Also, you need domain expertise to do good feature engineering and to develop good tailored architectures. These are quite critical to obtaining SOTA results, so far as I can tell. People like to pretend that deep learning and "AI" have made feature/model engineering unnecessary, and it's just not true.
Not really, you don't need to be an expert in go to beat the best go player.
I'd be very surprised if the AlphaGo team didn't at least consult with a few Go players...
Except go (and all other well defined games) are the most trivial of domains. For real business problems, the difficulty is most often trying to determine what the problem is and how it relates to the rest of the world.
there was a strong model of the world in AlphaGo, which allowed it to accurately simulate games that followed the rules.
that is quite a lot of domain expertise, much more than in most ML systems.
Another way of saying this, is you don't really need the same level of domain expertise if the problem space is entirely defined by a small set of clearly articulated rules.
Yea. In my industry you could hire a team of generalists and I wouldn't expect any of their conclusions to be valid. Now if a domain expert can get them a valid dataset...then they can do some statistics and analysis and maybe uncover something I wouldn't notice.