I think it's because the tablebase solutions are not based on theory, but on exhaustive search. And they have passed the limits of human memory, even the memory of someone like Carlsen, who remember thousands and thousands of positions.
It's hard to come up with the exact number of possibilities for each set of pieces, because some positions are illegal or impossible, but we can make rough guides.
The first king can be on any of 64 squares. The second kind on any of 57 squares (if the first king is somewhere in the middle). Then one rook can be on (roughly) 62 squares, another rook on 61, and the knights on 60 and 59. (Again, this will be a slight overestimate, but not too much, I think).
That's 48,840,445,440 positions! Of course, some are symmetric to others, but even figuring those exactly is going to be hard. And the solution to each of those 50 billion solutions would have to be memorized, because, again, the computer doesn't base the solutions on logic or "chess thinking".
This is happening now much earlier in the game. The computer will recommend moves that "no human would play" (or so many grandmaster commentators say).