AlphaZero: Will People Treat Chess the Way they Treat Tic-Tac-Toe? - Chess Forums - Page 8

SeniorPatzer · 2017-12-08T09:13:50-08:00

Chess aficionados have been bombarded (and rightfully so) by the news of AlphaZero teaching itself in a matter of hours to destroy StockFish. Regardless of whether it was a Google publicity stunt, we still have GM's utterly stunned and in awe of AlphaZero's moves. (Eg., GM Peter Svidler

Martin_Stahl

Mar 25, 2025

0

#141

magipi wrote:

PositionalDamage wrote:

To show that Alphazero's win again was LEGIT, that is why.

And Leela Chess Zero is Alphazero's younger sister.....

The were made with the same design for a chess engine.

The Leela of today is a lot better than AZ was back in the day. Obviously. Engines get better over time.

It's extremely suspicious that a super powerful engine supposedly appears from nowhere, plays a secret match or two, and then disappears forever. The fact alone that the match was played in secret with absolutely no transparency is enough to raise the suspicion of anyone.

It's possible that they cheated with hardware. AZ was running on the Google supercomputer, while SF was running on a grandma's laptop.

It's also possible that the games were cherry-picked. They played 10 thousand games, and then the PR guys picked the 100 that was the most spectacular.

AZ was written specifically to run the evaluations on the Tensor Flow hardware. That's not something Stockfish could run on.

One of the biggest steps was that AZ was a completely self taught engine and just through playing itself got to the point it could beat an engine that was very strong and programmed with theory, years of tuning, and brute force calculations.

Sure, they could have tested with a newer engine at the time and there's a certainly some criticisms that can legitimately be made about the configuration of the hardware running Stockfish. That does not take away from the method used to create the engine. The simple fact that LC0 was able to duplicate the results to a similar degree are a very good indication that Deep Mind's results were very likely legitimate.

Had Deep Mind been interested in anything more than a tech demonstration proving their methods could learn from scratch, they very likely could have continued training the engine, against other engines as well, and then had a competition against the best Stockfish version at the time. Unfortunately, that wasn't their goal and we didn't get anything like that.

magipi

Mar 25, 2025

0

#142

PositionalDamage wrote:

This is not hard to figure out. Stockfish 8 was the engine that played Alphazero.

So the best way to compare other NN chess engines. Is to have them also play Stockfish 8. This is why we know 2 things.

1. Alphazero's win was legit.

Your number 1 point absolutely does not follow. I have no idea what logic you see in there, but it doesn't exist. AlphaZero is gone. We have no chance to figure out how strong it really was, unless we accept the word of Goolge - who have no reason for telling the truth and every reason to bend it.

magipi

Mar 25, 2025

0

#143

PositionalDamage wrote:

Alphazero is not gone. It lives on in Lc0. Google gave us the blueprints to Alphazero. And how to recreate Alphazero.

1. We know Lc0 can crush Stockfish 8. Just like Alphazero.

Facepalm. Do you seriously believe that the current version of Leela is exactly as strong as AlphaZero was? If so, I have a nice bridge to sell you.

magipi

Mar 25, 2025

0

#144

PositionalDamage wrote:

Do you not read....

2. The lastest Lc0, or CPU NN chess engines like Stockfish 17, or Dragon . Are much stronger then Alphazero. Because the can beat Stockfish 8 with a bigger winning percentage.

Therefore, matching them against SF 8 has exactly no value. It proves nothing.

At this point I am convinced that you are trolling. I'm gone.

Elroch

Mar 25, 2025

0

#145

magipi wrote:

PositionalDamage wrote:

To show that Alphazero's win again was LEGIT, that is why.

And Leela Chess Zero is Alphazero's younger sister.....

The were made with the same design for a chess engine.

The Leela of today is a lot better than AZ was back in the day. Obviously. Engines get better over time.

It's extremely suspicious that a super powerful engine supposedly appears from nowhere, plays a secret match or two, and then disappears forever. The fact alone that the match was played in secret with absolutely no transparency is enough to raise the suspicion of anyone.

You think that's suspicious, how do you think about the fact that at the time they produced a similar AI for the (computationally harder) game of go and utterly crushed the standard of all existing engines AND one of the best human players in history?

And if that is not suspicious enough, having spent time on playing petty games, DeepMind effectively solved the protein folding model and have now calculated the shape of the large majority of all proteins known to biology. Suspicious, huh?

[For clarity, I am being ironic]

It's possible that they cheated with hardware. AZ was running on the Google supercomputer, while SF was running on a grandma's laptop.

Right, which is why the open source emulation Leela is probably stronger on the sort of hardware AlphaZero used.

It's also possible that the games were cherry-picked. They played 10 thousand games, and then the PR guys picked the 100 that was the most spectacular.

This is reasonable, but there is no good reason not to believe the result of the match. Since then, similar AIs have reached higher standards, but Stockfish managed to leapfrog the pure AIs by combining high numbers of nodes per second with a small, fast neural network for positional evaluation.

playerafar

Mar 26, 2025

0

#146

Elroch wrote:

magipi wrote:

PositionalDamage wrote:

To show that Alphazero's win again was LEGIT, that is why.

And Leela Chess Zero is Alphazero's younger sister.....

The were made with the same design for a chess engine.

The Leela of today is a lot better than AZ was back in the day. Obviously. Engines get better over time.

It's extremely suspicious that a super powerful engine supposedly appears from nowhere, plays a secret match or two, and then disappears forever. The fact alone that the match was played in secret with absolutely no transparency is enough to raise the suspicion of anyone.

You think that's suspicious, how do you think about the fact that at the time they produced a similar AI for the (computationally harder) game of go and utterly crushed the standard of all existing engines AND one of the best human players in history?

And if that is not suspicious enough, having spent time on playing petty games, DeepMind effectively solved the protein folding model and have now calculated the shape of the large majority of all proteins known to biology. Suspicious, huh?

[For clarity, I am being ironic]

It's possible that they cheated with hardware. AZ was running on the Google supercomputer, while SF was running on a grandma's laptop.

Right, which is why the open source emulation Leela is probably stronger on the sort of hardware AlphaZero used.

It's also possible that the games were cherry-picked. They played 10 thousand games, and then the PR guys picked the 100 that was the most spectacular.

This is reasonable, but there is no good reason not to believe the result of the match. Since then, similar AIs have reached higher standards, but Stockfish managed to leapfrog the pure AIs by combining high numbers of nodes per second with a small, fast neural network for positional evaluation.

When something 'disappears' suddenly and suspicioulsy you can be sure that there's a certain five-letter thing at the center of the cause.
M-O-N-E-Y
-------------
yes. That's right.

Martin_Stahl

Mar 26, 2025

0

#147

playerafar wrote:.

When something 'disappears' suddenly and suspicioulsy you can be sure that there's a certain five-letter thing at the center of the cause.
M-O-N-E-Y
-------------
yes. That's right.

There's nothing suspicious. It was a tech demonstration not a product (AZ not the underlying learning code). Things didn't disappear either; there were a number of people that got access to both the engine and the games.

playerafar

Mar 26, 2025

0

#148

Martin_Stahl wrote:

playerafar wrote:.

When something 'disappears' suddenly and suspicioulsy you can be sure that there's a certain five-letter thing at the center of the cause.
M-O-N-E-Y
-------------
yes. That's right.

There's nothing suspicious. It was a tech demonstration not a product (AZ not the underlying learning code). Things didn't disappear either; there were a number of people that got access to both the engine and the games.

I was implying - if it really is suspicious or if there's good cause for suspicion.
If you've now debunked the suspicion in this case - good for you Martin!