Amazing article! I have a habit of categorizing people into five categories (rather than three) and I would give examples of what “tasks” comes under which category. But you have articulated the categories very well, and now I can rely on your description than direct examples (which sometimes may end up hurting people. Which is why I would avoid talking about it).

As of today, AI is still in the “Third category” (or beyond), but this will not be the case soon. It will come down to the second or even the first, after a while. I’m sure of that. Probably in three years or even 5 years…

Even frameworks haven’t stabilized. For example, TensorFlow developers now realize that it would have been much better if their graph was dynamically extendable; and I have read that PyTorch (I haven’t used this a lot) implements dynamic graph by cannot beat TensorFlow when it comes to distributed computing.

So the day when AI moves from the 3rd category to the first category will be the same day when, most of the AI related paid solutions have out-of-the-box opensource alternatives; and yeah, the framework design for AI libraries would have stabilized the same way how OS Kernel design have been standardized today. It is the “1994" for AI, today (like it was for OS (in the form of Linux) back at that time)!

One more thing: If you look at papers published by people belonging to the third category (or beyond), they restrict themselves to a very narrow scope. This is likely because, the game at that level requires a lifetime of effort-investment. And once achieved, it would take the same amount to move to a different one.