Progress over the last 18 years

[Graham Banks]

Spike 1.1 won the very first CCRL Amateur Championship back in 2006. It has a current rating of 2690.

Going by the latest 1CPU version of Stockfish, there has been a 950 Elo improvement over that 18 years.

Can we reach a 1000 Elo improvement anytime soon?

[Dann Corbit]

I think that because chess strength on the Elo scale is exponential, adding 1000 Elo to a very strong engine is incredibly difficult and adding Elo to a very weak engine is much easier.
That having been said, I guess that SF will hit 1000 Elo (another +50) in two years. If they upgrade the hardware, it could be tomorrow.
Now, upgrading the hardware for a mob of testers would cost big money so I don't think it will happen.
I think also that there is some kind of effort to keep the actual compute power the same because the time scales like 40/20 have been compressed for some of the testing sites.
Unfortunately, Pohl's site does not measure regular chess strength anymore, so it is very hard to know how it has been progressing lately.

[Brunetti]

If they upgrade the hardware, it could be tomorrow.

Hi Dann,
by upgrading the hardware, everyone benefits, so why should SF gain an advantage from it? Unless you run its opponents on the old hardware

Alex

[Dann Corbit]

Good point, but hardware is no different than software.
Faster hardware = more Elo
Faster software = more Elo
Both increase exponentially.

I think if more effort went into search, the progress would be faster for software. Just a guess.

[towforce]

...adding 1000 Elo to a very strong engine is incredibly difficult...

Almost certainly not even possible. Stockfish right now is around 3640. Looking at charts of draw ratio v elo rating, it looks as though a player becomes unbeatable at around 4200.

[Brunetti]

Good point, but hardware is no different than software.
Faster hardware = more Elo

Since Elo is not an absolute value but rather the performance difference between engines, faster hardware, for everyone, means no Elo differences. Or rather, it depends on specific implementations and different time management techniques. If you run a tournament on the best current hardware and replicate it on one from 10 years ago, you might not get the same Elo ratings. You could even find that the winner of one comes last in the other

Alex

[Dann Corbit]

...adding 1000 Elo to a very strong engine is incredibly difficult...

Almost certainly not even possible. Stockfish right now is around 3640. Looking at charts of draw ratio v elo rating, it looks as though a player becomes unbeatable at around 4200.

True, but Graham was talking about the improvement of Stockfish from its origin until it hits +1000 Elo, which is only 50 Elo away.

[chrisw]

I think that because chess strength on the Elo scale is exponential, adding 1000 Elo to a very strong engine is incredibly difficult and adding Elo to a very weak engine is much easier.
That having been said, I guess that SF will hit 1000 Elo (another +50) in two years. If they upgrade the hardware, it could be tomorrow.
Now, upgrading the hardware for a mob of testers would cost big money so I don't think it will happen.
I think also that there is some kind of effort to keep the actual compute power the same because the time scales like 40/20 have been compressed for some of the testing sites.
Unfortunately, Pohl's site does not measure regular chess strength anymore, so it is very hard to know how it has been progressing lately.

Disagree your comment on adding strength.
Your very weak engine is a massively cut down stockfish. Adding strength is done by adding back missing bits of stockfish and tuning. This is quite easy until you run out of stockfish, when it then gets very difficult. Easy when you know how, difficult when you don’t, but unfortunately says nothing about some inherent exponentiality.
To say Elo is exponential implies you’re comparing to some underlying factor, presumably the probably undefinable and unquantifiable “chess strength”, but since chess strength is subjective and unquantifiable (you’re not allowed to use Elo here because that’s the other vector you’re comparing against in the exponential argument) you can’t use it to compare the relationship between the two. Maybe you can put Elo on one axis and time on the other. Exponential? Not. Or work involved. Again looks pretty linear.

[Modern Times]

True, but Graham was talking about the improvement of Stockfish from its origin until it hits +1000 Elo, which is only 50 Elo away.

Depends whose Elo you are talking about - Ordo or BayesElo and which testing group. CEGT already have it at over 1100 Elo at 40/20, and that is Ordo.

[Graham Banks]

...adding 1000 Elo to a very strong engine is incredibly difficult...

Almost certainly not even possible. Stockfish right now is around 3640. Looking at charts of draw ratio v elo rating, it looks as though a player becomes unbeatable at around 4200.

True, but Graham was talking about the improvement of Stockfish from its origin until it hits +1000 Elo, which is only 50 Elo away.

What I meant was that there is a 950 Elo difference between Spike 1.1 and Stockfish 17 on the CCRL 40/15 1CPU list, which as Ray correctly points out, is certainly not the only rating list.
My point was that how much longer might it take to increase that gap to 1000 Elo (another 50 points).