How long till computers reach 4000 ratings ?

That is not true. Older engines, 2500ish(comparable to humans) will always be at that level, modern engines 3000+ will stay at their respective levels once they are superceded by better ones therefore future engines will have to have a higher rating as they will be better relative to  current ones

Older engines will always be at the same strength (assuming no changes in hardware, of course).  But rating changes as the rest of the rating pool changes.  If you add a bunch of super-strong engines, those old 2500 engines will get pushed down to 2450 or lower.  If you add a bunch of weak engines, those 2500 engines will get pushed up to 2550 or higher.

Interesting stuff, cheers Philidor.

MGleason wrote:

krudsparov wrote:

MGleason wrote:

Elo is not entirely a measure of strength.  It's a measure of relative strength.  The top engines have high ratings not simply because they are strong, but because they are strong compared to other engines.

Therefore, to reach 4000, the top engines will have to increase in strength at a faster pace than the other engines they are being compared to.  So far they have managed it.  However, if the other engines improve and narrow the gap, the top engines will see their ratings decline even if they continue to improve.

That is not true. Older engines, 2500ish(comparable to humans) will always be at that level, modern engines 3000+ will stay at their respective levels once they are superceded by better ones therefore future engines will have to have a higher rating as they will be better relative to  current ones

Older engines will always be at the same strength (assuming no changes in hardware, of course).  But rating changes as the rest of the rating pool changes.  If you add a bunch of super-strong engines, those old 2500 engines will get pushed down to 2450 or lower.  If you add a bunch of weak engines, those 2500 engines will get pushed up to 2550 or higher.

.

.

But they wont be pushed down as the old engines 2500ish will still be comparable to GMs 2500+ and therefore better and better engines will need to be rated higher and higher in comparison to previous engines.

Who cares?  What I want to know is how long before a chess engine will analyze my game and tell me, "You shouldn't have attacked Kingside because you had no material advantage. You should have posted your Knight on that fine c5 outpost, then attacked up the middle with your d, e, and f pawns."

When a chess engine can do that, it will help my game much more than a 3500 vs 4000 engine evaluating each position.

Who cares.  I want a chess engine that will actually explain things.  

Indeed it will Mick! But that's not what the OP was asking, so some people are interested in the question.

I read the link that Philidor_Position referenced, but still think 3600 seems a bit low. The top engine now is ~3400 (correct me if I'm wrong). Is it not true that the best centaurs (strong human with engine assistance) can still beat the best engines? If it is true, then the engines still have a lot of scope for improvement

Probably there now.

The old engines and modern GMs are no longer in the same rating pool.  You can't directly compare engine Elo ratings with FIDE Elo ratings.  You also can't directly compare FIDE Elo ratings with USCF Elo ratings, as they are two separate rating pools.

Elo is not a measure of strength.  It's a measure of comparative strength relative to the rest of the rating pool.  If you add a whole bunch of strong members to the rating pool, your rating will go down even if your strength stays constant.  If you add a whole bunch of weak members to the rating pool, your rating will go up even as your strength stays constant.  This is because, while your strength has not changed, your relative strength when compared to everyone else has changed.  The same thing happens to engines; their strength stays constant (assuming equal hardware), but their rating might move depending on the addition of new engines or changes to other engines.

That is already achievable. For example 15 years ago ChessMaster program had simular preprogramed limited analysis of this type. If there is enough money and interest new program can be created with far more words etc. But what you mean i think is free speech, any position. That can be possible with artificial inteligence learning algorithyms - no coding, no pre-programmed stuff. All in real time. The program is half self-aware. Currently many so called agents exist. The best is DeepMind by google, last year beated Go champion , learned , taught and mastered the game by itself with no human help.No coding. Those system will be the future and will change our life in next 10 years or so just like smartphones and touch screens did it 10 y ago, but way bigger change.

In middle of the century A.I will be almost self aware and could do almost  any human job equal of better than humans. For example full Self-driving cars already exist right now someone somewhere is driving one (search Tesla S car )  (in limited locations) and in few years will be the start of mass production (final test are made right now with start by 2020) but not until like 2030+ they will enter most roads, cities and countries.

It's not going to happen because the rating cap is somewhere around 3300. As you play opponents lower and lower in rating, you gain less in a win. Therefore, even if a 4000 is playing a 3000, it will likely gain 0 points.

There is no mathematical upper limit to Elo ratings.  Sure, you won't get past 4000 if the second-best opponent is 3000, but if you've got several 3300 opponents and you trounce them all, you'll get to 3600.  If a few more engines come out that can trounce the 3300 opponents, they'll get to 3600 too.  Then, if you can trounce those 3600 engines, you'll get to 3900.  And so on.

There's no mathematical upper limit; the only limit is that you can only get so high above your strongest opponent before you no longer gain anything from a win.

Even in 40 years computers will not be able to do research

in mathematics, physics, computer science in the same quality as scientists.

I don't know, but I'm guessing, based on my analysis, that Gioacchino Greco would be at roughly 2000ish today (except that he'd sometimes get killed in the opening).  I haven't gotten to Staunton yet.

Elo calculation shows that 4000 Elo engine is an engine that can win against Carlsen(2835) with a probability of 99.99545%; draw 0.004549%; lose 0.000002%. That is(on average) 5 draws, 99995 defeats in 100000 match marathon.
That shows it would be physically impossible to get AI Elo rating by comparing it to human.

Against Stockfish(3391) 4000AI would(on average) win 97 games, draw 3 games out of 100. Stockfish winning probability would be 0.0027963%. Such AI seems unlikely but at least more realistic.
But human and AI Elo pools do not correlate perfectly already, because of a lack of games between them. Also, humans or AI can play for a draw which might affect results.
To be honest, I just wanted to mess with and Elo calculator(https://wismuth.com/elo/calculator.html), so don't take this post too seriously. : )

I don't think that centaurs are any stronger than engines.
In 2014 Nakamura + Rybka played against Stockfish with no odds and lost 0.5-1.5.
Some things to consider is that engines improved somewhat since 2014; at that time Rybka was weaker than Stockfish, so it would be more fair to play Nakamura+Rybka vs Rybka, or Nakamura+Stockfish vs Stockfish; also, there were only 2 games played, who knows what would happen in, say, 7 games.

 10000