Peter Svidler and Sergei Shipov discuss Alpha Zero

Dec 8, 2017, 4:28 AM |

What do you think of Alpha Zero destroying Stockfish?

I'd like to clarify a few things first. For instance, I read that Alpha Zero played on much stronger hardware than Stockfish. But other people write that Alpha Zero trained on much stronger hardware than Stockfish usually trains on, and then they played on equally-strong hardware. This is an important moment. Also... I don't like wearing tinfoil hats, but I looked through some games. The games were absolutely fantastic, phenomenal. But it was said that neither chess engine had opening books. Alpha Zero won several incredible games in Queen's Indian, with Qc2, c5, d5. This is central theory... (Laughs) The central theory that was developed by Borya Gelfand, Lyova Aronian and others from the ground up ten or so years ago. And now a computer just plays like that on its own? This is absolutely central theory!

It just improves theory...

Yes. It's an absolutely central theoretical line. We're told that it has no opening book, it's just so devilishly strong that after training for a few hours it's able to replicate things that took humans years to develop. This was a breakthrough in Queen's Indian, you remember. This line was a breakthrough. I was in awe of the machine's games, but I was just astonished when I saw openings. I thought, "Damn, if it can actually..." I can believe that it can play equal positions greatly, but if deep learning can actually replicate opening lines and improve upon them, it's just stunning.

Basically, as Peter Nielsen, Anand and Calrsen's coach, said, "the aliens came and showed us how to play chess". The first reaction is obvious - that it's fake, deception, that it's just a show put up by developers. Something like Slyusarchuk's blindfold games...

I don't believe that.

Yes, Google and DeepMind are serious companies, they shouldn't deceive people so blatantly. Even Garry Kimovich [Kasparov] is in awe!

I don't believe in any deception - we already saw it in Go.

Yes, AlphaGo basically solved Go, and now they solve chess.

The fact that it was harder to learn Go can be easily explained: chess is much more calculation-based game, and it's clear that if you already have an AI that learns so quickly, it will learn chess much quicker. This is all logical.

Chronology is important. Go happened two years ago, and now...

Yes. They learned to do that, learned to give it the rules and say "Go ahead"... I'm not amazed with the fact that it learned chess, but I was stunned by the games' quality. I expected something non-principial in the opening, and then cosmic-level middlegames. But in addition to cosmic level, it shows things that it should not have possibly known. Very interesting.

I'd like to make a couple of remarks. First of all, the hardware comparison doesn't make sense. Neural network of Alpha Zero, and an ordinary computer used by Stockfish... It's completely different things.

OK, your opinion here is much more valid than mine, because I've never seriously studied computer hardware.

But the main question is, can the things done by Alpha Zero be adapted for personal computers? Can humans use it in the future, or this AI will exist in its own league? Of course, some people already ask whether they can borrow this Alpha Zero thing to study openings. If it studied chess for just four hours, as they say, and created such an impressive opening book for itself that its novelties amaze even supergrandmasters... Look at the Candidates' tournament participants. For instance, Sergei Karjakin: he would probably be glad to work with Alpha Zero and amaze everyone with opening novelties. But I think this will cost a lot of money...

This is simply a question of resources. Some people have it, some don't. But it all boils down to whether it's technically possible. Can it be adapted for PC, or it requires special infrastructure?

Even if it requires special infrastructure and hardware, you can still have it study some opening and develop novelties...

You know, Google probably don't need money all that much...

So, even with all Karjakin's sponsorship, he just won't have the money to use the services of this thing...


They spent four or so hours on chess, then destroyed shogi in two hours... you know that story, right? They needed just two hours to defeat the strongest shogi program. It's totally clear that they'll be working on very different things now, chess isn't a priority. The machine spent some hours on chess, defeated chess, and now works in other fields.

Yes. Chess and other games are just a publicity stunt for them.

Great publicity, I must add.

Publicity not in the sense that it's all just a show, but in the sense of making headlines, etc. This is the right kind of publicity: they achieved something and show it to people. But it's clear that they're working on artificial intelligence not for the purpose of solving tabletop games, so I'm very skeptical about borrowing Alpha Zero's services to prepare for the Candidates'. We shall see.

Of course, we would like to have a duplicate of Alpha Zero somewhere around the Central Chess Club in Moscow, to work on chess specifically. Russian Chess Federation would probably be interested in something like that, but it's unrealistic. This is too cutting-edge science, too exclusive thing, and it can't be pirated online (Laughs). Still, I'm quite worried by the question: What would this monster actually do now? I hope it's not military-related...

The question of what awaits us in the future with AIs that improve by such leaps and bounds... well, you know, it's a chess show here. (Laughs)

You know, everyone watched Terminator...

OK, if you insist on talking about that, I'll be concise. We should worry.

Yes. Because humans indeed look puny in comparison to such achievements. Before, we were at least proud of something, but now...

Terminator and Skynet are one thing, but the problem is that this improvement is uncontrollable, and I've got a feeling that... After reading the press releases, I came to the conclusion that this is now an important commercial project. Various companies compete between each other, they all want to be the first to build a real, honest-to-God AI, and... in this race, nobody really asks the questions of what happens when somebody wins. The questions of security, of what will it start to do when we build that machine. It's clear that people who want to create an AI don't want a downer ending, I don't doubt that their intentions are good, but... one little error in coding made because it was important to finish ahead of another company, some mistaking in testing may lead to global consequences.

Sci-fi writers can probably describe it better than us.