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

The problem here is that most treat rating as a comparative system, like 3000 is better than 2500, while some treat it as descriptive. People say things like: " a 1200 should work on his openings, or an 1800 should study endgames, or a 2000 should focus on advanced positional chess..etc. If rating is just a pure comparison, than all those statements have no basis, as it is a ranking system and not HOW GOOD AT THE GAME you are.

That is my point, a 3500 engine crushes easily a 3400, so a 3000 would lose almost every game against a 4000

That is true by definition of the Elo scale. A 1000 point difference leads to the same statistical expectation (about 99.7%). With the stronger players a subtle difference is that you would expect the weaker player to win even less and get slightly more draws to achieve the same result, but this hardly matters much when they are meant to achieve one draw in 159 games. (Such an expectation means you would need thousands of games to get a reliable sample).

The number of games necessary for a reliable sample is no problem when engines are concerned. The problem is that there are some who know quite a bit about engines and nothing about chess. Chess is a draw, and by a wide margin.

You should look at engines testing sites where millions of games were already tested  and get an idea how 3000 engines play vs  3500 engines .

( e.g about half million games are being tested among 3500 rated stockfish vs stockfish everyday.)

(about 1 million games are being tested among 2500-3000 rated Leela everyday)

And also you can follow TCEC  live games of 3500 engines as well, and see some insight into chess.

So, the result of AlphaZero/Stockfish tournament was 886 draws, 290 wins, 24 losses.

Therefore from Elo definition, the Elo difference should be AlphaZero about 430 points above Stockfish, and if Stockfish is rated about 3390, then this gives 3820 for AlphaZero.

Does AlphaZero easily beat a player rated 2820?

If a 2820 rated player, like a human grandmaster, played 1200 games against AlphaZero, what would be likely outcome?

Still 886 draws, but now 313 losses and 1 win?

Or, say, just 300 draws, 897 losses and 3 wins?

This is unproven. And that is another problem with ratings. If prefect play by both sides leads to white winning for example, then 2 equal players will ensure white always wins. In other words, 2 equally perfect rated players WOULDN'T draw, changing the ratings

Wouldn´t draw games one by one. But this does not change the ratings, if there are equal numbers of blacks and whites assigned. Because if perfect players draw then a tournament would result in all draws and no wins or losses; if perfect players win with White then a tournament would result in equal number of wins for each and no draws. Both cases equal Elo.

In that case yes, but then color becomes a factor, not simply who is better.

You can say that it's not proven that chess is a draw, but it's axiomatic for the top players in the world, and even the ICCF. There are serious conversations about changing the rules of the game for correspondence chess to allow a pawn edge to be considered a win! 

There are people who are really not clear on the concept of a theory. A theory isn't proven beyond all doubt, but it's clearly the best explanation for observed phenomenon and can be used to accurately predict future outcomes. 

Chess is a draw. The more people say it's not proven, the more it shows how little they really know about the game.

A rare quality post.

Not rare for @smyslovFan, but rare for the forums in general.

Got any proof of your claim? Remember, White has a tempo.

Well he probably meant chess is probably a draw, not chess IS a draw. But technically he's right isn't he? Sometimes it is a draw. Sometimes it's a win for white, and sometimes its a win for black. So he's right, at least some of the time. As for stating that chess is a draw as some sort of final solution I think it's the same as saying chess is a win for white. It's just an opinion which may or may not be true. I think given that we are in the infancy of computers it's far too early to speculate on whether or not chess is a draw or not because we have so little knowledge about a solution. A few hundred years from now when computers are a lot better I think we can make a better guess. At least by then we'll know how easily a 4000 beats a 3000.

My point was that to fairly rate players, you need an equal game to begin with. So if White has an advantage from the start, he might not have to be as good to win. If I am up a queen against magnus, i will perform like a 3300 by beating him even if I am really a excited 1500.

Unfortunately you have the Elo calculation wrong.

The advantage of the first move equates to about 50 rating points on the Elo scale. 

There's an easy way to equalize the problem of the first move: play an equal number of games as White and Black, thus removing the bias. 

Why dont you check statistics of computer games, there are more than 1 billions games stockfish vs stockfish played already. 

Is this a good time to get some hypothesis?

What's your hypothesis? I think we can rule out black winning a perfect chess game (because white is in Zugzwang from the beginning lol). Either White Wins or a draw.

the higher the numbers get the more often it will end in draws alpha zero outweight stockfish be a lot and still couldnt beat it in 75% of the games

Another problem with ratings is that they start provisionally, so they aren't perfectly accurate. If I am rated 1000 provisional, and then beat a 2200 (cause I am really 2400), the 2200 loses points he shouldn't have. It takes a couple dozen games to see what level a player really is, and in the meantime, screws up other people's solid ratings.