future of top engines:how much more elo?

[Uri Blass]

I don't think we know the answer to any of the three questions, but I guess it is "a lot".

Yes, and I believe even the most optimistic outlooks are going to be surpassed. Here's what people were thinking about this on 2008:

The right & interesting question is: which ELO involves the solution of the chess game ?

No this is not the right question. A better question is: With 1 TB memory for an opening book what rating is required for white/black to hold the game?

I think 3100 is enough to hold the game for white, while maybe 3500 is required to hold the game for black. In both cases a very strong opening book is needed.

"The solution for chess" will have to consider openings like the Najdorf, where black might (in expressionistic terms) need a "4000" rating to hold the game. If certain very sharp openings have to be played (as a requirement for both players) perfect play might require a rating of 6000 or even more. These very sharp openings might need to be considered in "solving" chess, but are irrelevant for what rating is required to hold the game.

As I pointed out frequently - though people does not seem to "get it" - the level of play of a 32 table base depends on heavily on how it selects moves in equal positions.

If the moves in objectively drawn positions are selected randomly in my judgment a 32 table base will have a rating only slightly higher than its opposition for players with rating +2200 (in other words even a 2200 player can draw almost all games)

If the moves in objectively drawn positions are selected as generously as possible (e.g. playing 1.Nh3 followed by 2.Nh3-g1) virtually any player can get a draw e.g. by answering 1.Nh3 with 1.-Nf6 and after 2.Nh3-g1 play Nf6-g8!.

If the moves in objectively drawn positions are selected by a chess engine in my view it does not matter that much if the program is weak or a strong. I think that a 3500 engine equipped with an extremely good 1TB database is necessary but also sufficient to hold the initial position for black.

3500 elo is just 50 elo stronger than current strongest Stockfish, and nothing close to what people imagined a decade ago is being perceived. 3100 elo enough to hold the game as white? What I've been seeing is that, no matter the hardware, software, or time control, white has a though time holding the game (and at the top I've seen more games won by black!) Unless both players are happy with a draw (which is so common people perceive "draw death".)

My prediction: We're at the end of the line, and within 25 years a system is going to be developed that is able to pick a non-losing move in any chess position. And that will be it, corr chess players with access to this system will not lose any game, chess engine development will cease because people will never be able to beat the system, computer chess will go to the story books in a similar fashion to checkers, and... I guess the only way to keep things alive will be with honor systems where people pretend the system doesn't exist and play each other the old-fashioned way.

But I hope to be wrong.

It already seems that top correspondence players can at least draw every game.

We had 123 draws out of 136 games that is more than 90% in the ICCF correspondence world championship that started in 2015

https://www.iccf.com/event?id=52852

We already have 122 draws out of 131 games with 5 unfinished games in the following event that started in 2017

https://www.iccf.com/event?id=66745

We do not see a single loss for players who scored more than 50%

Edit:

I doubt if today top correspondence players can beat stockfish most of the time(note that stockfish is not deterministic so it is not going to play the same moves in every game).

[Uri Blass]

<snipped>

[
3500 elo is just 50 elo stronger than current strongest Stockfish, and nothing close to what people imagined a decade ago is being perceived. 3100 elo enough to hold the game as white?

The claim was that you need a very strong opening book for it.

Of course if you have a good book that is big enough no engine is needed because all the book is written in the 32 piece tablebases so we need some size limit about the size of the book.

Rating is dependent on the books people use and unfortunately I do not know about people who try to build a special book against stronger engines to see how much the book can help.

[Ozymandias]

Nowadays, people expect the Elo system to allow them to sort differences in playing strength, from the random player to the best computer money can buy. They also expect to keep using it for future entities. Was it designed to do that? No, it was meant for human players, who will never fit into neither of the two categories. With time, this will be even less so.

[Ovyron]

It already seems that top correspondence players can at least draw every game.

Look at the actual games, not just statistics, they're not trying to win, they're happy with a draw, they play the same old safe moves, their opponents aren't challenging them in new positions. And it seems they got their rating by playing against opponents that just mess up positions and could have lost against anything they'd have thrown at them (as I'm doing, slowly creeping up the ranks, I guess I'll eventually join their group and see how they handle someone not happy with a draw.)

This was apparent after importing the full ICCF database of games and checking what people played. They're just GM draws at a higher level, and it seems people are scared of trying something adventurous and losing.

I claim the incentive to win isn't enough. Put a money prize of 10 million dollars and I bet someone that really wants them is going to come up with something to win them, and it's not going to play the same safe openings people have been playing for decades, that have been examined to death so both sides looking for a draw will get it.

I doubt if today top correspondence players can beat stockfish most of the time(note that stockfish is not deterministic so it is not going to play the same moves in every game).

I can beat Stockfish on a regular basis at corr time controls. In fact, I'm winning games on the ICCF against people that just followed Stockfish recommendations like parrots. But the opening is critical, and since Stockfish can't play its own openings I guess it'd come down to a Stockfish wanting a draw wouldn't be defeated by an human wanting a draw...

Of course if you have a good book that is big enough no engine is needed because all the book is written in the 32 piece tablebases so we need some size limit about the size of the book.

No matter how big is the book, you can always get it out of theory with little effort, after knowing their variations it plays against you.

Plus, I haven't seen any correlation whatsoever between a book's size and how strong it is. The strongest book I've ever seen is 27MB in size, relevant positions you'd want your engine to play aren't heavy, so you'd probably be able to reach the theoretical strength of that 1TB book in just 30MB, that's why the strongest books of an era don't get increasingly bigger than in previous eras.

Rating is dependent on the books people use and unfortunately I do not know about people who try to build a special book against stronger engines to see how much the book can help.

People have been building special books against stronger hardware, and apparently the conclusion is that with a book strong enough you can equalize any stronger hardware (at least for A/B engines, NNs are a different beast despite ending with same elo...) If hardware can be made irrelevant and everyone is using the stronger engine then the book is the most important part of the equation.

[Uri Blass]

If you put a lot of money then I am afraid somebody is going to cheat in correspondence games(meaning he is going to pay another player to lose).

[Ozymandias]

If you put a lot of money then I am afraid somebody is going to cheat in correspondence games(meaning he is going to pay another player to lose).

You can be certain of that. With much smaller prizes, I saw this happen in two separate occasions (involving four players) in the only Freestyle event I could ever attend, that took place physically. I suspect a lot more occurrences of the same behaviour in online tours, but it's much more difficult to get people's tongue loose, when they have time to reflect on what they're writing.

[OneTrickPony]

If you have significant budget you can make an open challenge: it's 1k to play a game but 100k if you win it. This assumes you can play chess you believe is unbeatable yourself though.

[Ovyron]

How does Jeopardy do it? Recently some guy won +2 million dollars in prizes by answering trivia questions, and at no point was there collusion to let him win by his opponents.

Perhaps the players would need to be paired against each other anonymously, so you don't know who else is attending the tourney and can't arrange anything. Or something. If you have 1000000 to spare for this, hopefully you can also build an infrastructure where people can't cheat.

But anyway, nobody is going to do this, so, let's assume collusion can't happen and people aren't going to drop games to get paid by the winner. Do you think we'd still perceive draw death?

[Ozymandias]

How does Jeopardy do it?

A TV show isn't probably a good example, remember "Quiz Show"? That was way back in the 50's, so imagine what they get away with now.

[Uri Blass]

How does Jeopardy do it? Recently some guy won +2 million dollars in prizes by answering trivia questions, and at no point was there collusion to let him win by his opponents.

Perhaps the players would need to be paired against each other anonymously, so you don't know who else is attending the tourney and can't arrange anything. Or something. If you have 1000000 to spare for this, hopefully you can also build an infrastructure where people can't cheat.

But anyway, nobody is going to do this, so, let's assume collusion can't happen and people aren't going to drop games to get paid by the winner. Do you think we'd still perceive draw death?

I guess we are going to see draw death with no paying.
With paying there may be a problem when 2 people agree in advance that A lose to B if they meet.

They can also agree in advance about the exact game when the losing mistake is late in the game so random pairing is not going to help because the probability to get the same game with another person is almost 0 and having the same game for the first 5 moves or the first 10 moves is not enough for losing because the losing mistake they plan in advance happen only at move 50.