A 3000 could easily beat a 2000, but could a 4000 easily beat a 3000?

Middlegame is like an endgame with just many pieces, especially if there are few pawns! Actually chess is just the ultimate endgame of 32 pieces, so a 32 piece tablebase should be able to solve chess.

Yes, a 32 piece EGTB by definition solves chess.

And I'm drunk, so I got you confused for someone else, sorry.

Once again, Elo works the same way regardless of the base rating. A 1000 point difference is the same at any point along the number line. 

The problem is that chess is not a perfect zero-sum game. In fact, it is a theoretical draw. 

I've asked a few GMs, who posit that even Magnus Carlsen could not score more than 1 out of 20 against a perfect machine. Alex Fishbein points out that a truly perfect machine would know to get Magnus out of the opening as quickly as possible, and then outplay him even as Black. He suggests that the top rating for a computer is perhaps over 3800. 

So yes, a 4000 would destroy a 3000 precisely as easily as a 3000 would destroy a 2000 or a 2000 would destroy a (true) 1000 rated player. It's the whole point of the rating system. 

Many people think that 32 pieces table base will be the highest elo.

In fact, 32 pieces TB may proably be lower rated than current Leela or Stockfish running on high end machine running good amount of time , e.g one move per  two weeks.

Why? 32 men Table base doesnt know "winning probability". If the position is draw, the table base will throw away practical winning chance.

Look at this position taken from 5 men Table base. Table base will choose either one of those draw moves randomly even giving away free rook thereby practically throwing away winning chance. So , it will lose elo.  However , Leela will try to play Kf2 and try to win, squeezing chance of winning in case opponent play blunder and squeeze practical winning chance.

I believe that the moves are chosen by the engine rather than the tablebase and are not usually random, so what you say is not necessarily true. (The draws are not even in the tablebase.)

It would certainly be true of the Wilhelm/Nalimov combination, because Wilhelm appears to look for the quickest route to a statutory draw. E.g.

(Obviously it doesn't look very well, it should rather have played 1.Qb5 ... , 2.Qxd7.)  

But is it a theoretical draw. There are like 10^100^100 possible games

This position is a draw too according to the tablebase site, yet stockfish played this against itself and in around 33 moves white won.

Most strong engines could 10000-0 Magnus

Magnus gets beaten a number of times in the blitz championships

Non sequitur

I downloaded SF12. It's only marginally better than SF8 (though reasonable as Black). Here's the mate in 50 example that SF8 drew in 5 also at G120+3 (mins+secs).

 So it's down to maybe LC0+GPU. Anyone?

 Please read the chess rating testing conditions from 

1. Stockfish website

2. CCRL

3. CEGT or other websites.

Rating are based on their testing conditions , not on your custom positions. 

I am pretty sure none of engine authours today will care about 6 men , 7 men tablebase positions it is already part of 100% solved chess.

Once engines hit the tablebase positions, the engine search is stopped, so technically Stockfish 1 and latest Stockfish 13 development versions should be equal rated in your custom position

Depends on your time control and hardware. If you give Stockfish 250 million nodes per position ( which is a minimum nodes to be assumed as quality engine move), I bet Stockfish can draw 99% from black, if not 100%. ( I would like to bet Stockfish can do 100% draw if you give 1 billion nodes per move for Stockfish).

It is a tablebase draw, 1. Qb5 is still draw.

But don’t you think once we make a 32 piece table base, we can figure out some algorithm which can estimate winning probabilities, and ultimately play “perfect” chess?

Winning probablility is based on opponent error, tablebases are constructed from zero error or fully calculated moves , mathematically called N! or permutation.

There is " zero " posibility that those two states can coexit together.

Ok, but what if we look at the accuracy of moves required to keep equality. As in, calculate the amount of playable moves in a variation. Do you think this could work?

Yes I know it's a draw. (But try playing it as Black against Stockfish and see if it's still a draw).

I think you missed the point. In the line he takes after 1...KxQ it's not a draw until 50 moves later, whereas the moves I give force an immediate draw (on pain of a Black loss). So if Wilhelm was looking for the quickest statutory draw he screwed up.

Theoretically, there is no 100% accurate chess except tablebases.

"Engines are based on two inaccurate conditions, inaccurate search and inaccurate evaluations",  less inaccurate in less time, more accurate in more time. 

That's nonsense. As I posted, programs can still select what they regard as the best moves by a forward search but limit the moves considered for its own moves to those mandated by the tablebase, as I understand SF does.