In all of this it's a matter of scale.
Products with 100 million users can amortize the costs over a much larger number of deployments. Even the smallest details become valuable at that point. Conversely, at a certain level of deployment, usability is not worth the time investment. Machine Learning researchers can frequently get away with a loose collection of Python scripts with a basic set of instructions that probably worked on their local setup.
Are they really an outlier, or just the best at what they're trying to do? I mean I guess you could always say that the best is an outlier since only one can be the best.
They're the best at what they aim to do. Again, if everyone did it or they were less unique then it would be less exceptional.
The point is, it's hard and rare to get into such a position. Yes, eventually there will be a new Apple, but not yet. So anyone *today* looking to perform well should probably be more practical.