Google's Artificial Intelligence Alpha Zero Conquers Chess Only 4 Hours After Learning The Rules!

What can I say?
Lichess.org is an amazing Open source site which allows you to play and learn chess for free. It'll always be free!
I forgot to mention: the match between AlphaZero and Stockfish was a 100 games long. 50 games with each color. AlphaZero Won by a high margin without losing a single game!
Link to the games: lichess.org/study/wxrovYNH
Link to the article by deep mind: arxiv.org/pdf/1712.01815.pdf
Thanks to my patrons:
Matthieu Aebischer, Shaul Weinstein, Jed Read, Marco Molina, Yaakov Ilowits, Julien Beauviala, Darcy Linde, Paul Firth, yairwein, Liat Marshanski, Nils Bossaller.

[@NathanOkun]

We have graduated to position-pattern chess, no longer piece-wise-power chess. You have to understand a 64-block 8x8 board as a single thing, not as individual pieces with different strengths. This is why Alpha Zero needed only 80,000 patterns/sec to out-play an opponent with millions of piece-moves/sec. Since there are only so many patterns that can be generated by single legal moves from an existing pattern, this "whole-board-image" method taken out to 60 x 80,000 = 4,800,000 images in its allowed minute (usually much fewer need be predicted from a given position, taking all possible enemy moves into account) and matches it to known winning patterns from its learning process. Once it finds the best fit for a known winning position it has previously used, it can, in effect, forget all prior moves and use that as its new start situation, "inch-worming" itself to victory by always starting with a known winning set-up, no matter what the enemy does. Humans do NOT think this way!

[@Censeo]

If I understand it correctly. It is actually the first time a computer is taking a shortcut and analyse patterns in the game instead of just using brute force. This is a huge revalation in ai science and is partly explaining why it took so long for computers to beat humans in the first place. We had the very handy shortcut, which now computers use too. Much will happen in the future of ai. Scary and exciting.

[@Chris-io2cs]

A0 is an amazing implementation of AI and machine learning but your'e underestimating the effort put into engines like SF. Chess engines haven't been brute force since the deepblue days when they actually used asic units so it could find moves faster. I don't wanna waste too much of your time, but if you wanna know more about the match settings and why it was more of a measurement rather than a definitive match, just read the response by a stockfish author near the bottom of this article www.chess.com/news/view/alphazero-reactions-from-top-gms-stockfish-author. Not taking anything away from A0 but stockfish isn't a bad engine for searching more positions, that's just what they've found to work. There are lots of concepts aside of brute force being used.

[@hypersonicmonkeybrains3418]

only problem is most real world problems are far more complex than a game of chess.. Chess is like a very very boiled down problem... So we now just need Alpha Zero to learn to make the most efficient microprocessors( far more complex problem than a game of chess but maybe not impossible). And then use these new processors to train Alpha Zero on real world problems like, how to achieve the highest crop yields, or how to make the best material, or how to author the best genetic code to cure all disease and evolve humans...

[@hypersonicmonkeybrains3418]

Ah but it does not need to solve it... it only needs to find a strategy that fairs better than humans. There is probably no right way to solve most complex problems, the complexity is maybe infinite.

[@violetlavender9504]

HypersonicMonkeyBrains Teach an algorithm to optimize hardware production. Then teach it to manage the marketing using big data. Then have it run the companies finances. Then teach it to program. Teach an algorithm running on a computer to build better algorithms and computers, and then those will be better. That will create the singularity.

[@hypersonicmonkeybrains3418]

Solve artificial intelligence. Then use it to solve everything else.

[@stigcc]

Great video!

[@MrPikkabo]

@6:05 "as a human I would prefer to be on the white side"

[@alonefrati4724]

Hi Tal, I noticed you tried to go live here on youtube, but for some reason I couldn't connenct. Can you upload the stream so I can watch it soon? Thanks

[@GMTalBaron]

Alon Efrati it's unwatchable. Even the recorded version. Sorry.

[@dilyan-2904]

Tal Baron I was hoping for you to play at title lichess tournament yesterday...

[@alonefrati4724]

Tal Baron oh, ok thx

[@skdkskdk]

Was wondering what happened to you Tal, after Nicehash got hacked I could not afford to lose any more :) So, considering a career change now that DeepMind WILL KILL US ALL? :)

[@MKD1101]

What was the final score? Where can I find all the games?

[@4xelchess905]

72 draws, 28 wins, no losses for Alpha Zero. dunno where to find the games though

[@dipulbharalua]

Alpha Zero doesn't calculate useless or avoidable lines because it is intelligent enough to ignore them. That's why it is better even with less calculations.

[@dannygjk]

I bet Tal and Morphy would say, "of course white is winning" before the game even gets to the endgame.

[@johndoe1909]

It actually Only under 4 tips which is impressive

[@antoinebournel6116]

that's unreal

[@antoinebournel6116]

I feel the joy in stockfish engineers

[@Nemesis_HEX]

now king of chess

[@KungFuBlitzKrieg]

Apparently, Alpha Zero is the Alpha and Omega of chess.

[@andresrossi9]

That proves that nothing can't play perfect chess in the world, even alpha zero. It has too many different combinations that even computers can only calculate a microscopic part of them. So, even humans still can beat computers if they find a new effective way for the thinking process!

[@AYstrength]

Who would have believed that the engineers were so wrong for years and yers thinking that brute force and monte carlo calculations would solve chess which has so many possibilities... All it took was a computer that metaphorically sat down in front of a chessboard and start asking itself by playing how can i improve from my last game and which are the winning moves against myself. 80 000 calculations per second. Its freakin Nothing. Even more remarkable is that deep mind seems to have personnality in it's moves as it almost intuitively guesses good positions without programs which have higher power finding the winning lines. 4 hours. Will chess be broke in one year of deep mind calculations?

[@willywonka6487]

what a strange sentiment this is! do you not realize this machine learning process itself relies on brute hardware strength to be performed? also during these games, a great deal of CPU power was utilized

[@Cscuile]

SF8 was handicapped computationally during it's fight with A0. Not to mention A0 was up against a weaker version of SF without a book.

[@dannygjk]

The book will not save SF. AZ outplays SF after the opening.

[@Cscuile]

The book is not the only factor in SF’s handicap. For one Google used a weaker, older version of SF. SF was also allowed only 1 GB of hash to use which is not nearly enough for 64 cores. On top of all that A0 had access to the full resources of Google during its first stages in training, approximatly 5000 TPUs. Overall Google has been overly secretive with this whole ordeal. Why would they only release 10 games out of 100, why not steam the event, why use a time control that favours A0, why refuse to provide additional comments on the specification of the match? Google has been doing something fishy...

[@MrSupernova111]

This match was rigged. Check game 3 when SF8 moves Rf8. Tell me which decent engine will recommend a blunder like that one? If you are rated 1500 + you don't need an engine to know its a major blunder.

[@Nemesis_HEX]

fabulastic move

[@drakkkkho17]

MAtojelic analize some of thegame and he emphasizes that some stockfish moves aren't the strongest or seems lack of logic. He alzo suggest that stockfish it's not using the strongest level...an dthis it's a media effects from google and other particular interestings...who knows..maybe this is some commercial issue...

[@evgiz0r]

Stockfish is strong enough to "learn from" IMO. so if a player could strive to be like stockfish, striving to being Alphazero is not really different. Both of those are just beyong our play

[@4xelchess905]

Sometimes Stockfish gives unhelpful advice because they assume you can defend a position only a super computer can defend. Alpha Zero will probably do the same, on occasions. Still, I think there is a huge difference between SF and AZ : SF uses finite depth min-max algorithm (or refiniment, like alpha beta pruning combined to iterative deepening), while AZ uses Monte Carlo method, meaning it imagine variations down to the very end of the game when evalutaing positions. It means that, if only for it's analytical method, SF is completely devoid of long term strategic understanding. It is actually not totally devoid of it because we hard coded strategical understanding for it (fine tuning of heuristic function, like evaluation function). So SF has little to learn us as for strategy is concerned. It still manages to beat us because its monstruous forsight still is better than ours, but computation power is not something a computer can teach a human. Alpha zero, on the other hand, plays game to the end, and, if I'm not mistaken, the outcome helps him fine tuning a neural network which basically understand the game of chess, better and better. Will it be able to teach us? Hardly, just like intuition is hard to teach from human to human, but I don't think it's impossible.

[@evgiz0r]

You are right. But i feel that this REALLY deep "strategic" search always involves a lot of tactical vision. Chess is pretty tactical, even in strategical positions and endgames. That is why I feel that AZ doesn't give much added value for the human player. We cannot really explore this strategic depth, except few moves deep, because our brain is pretty slow in this sense. Even if you KNOW a position is winning according to tablebase/AZ, it doesn't help you much unless you play precise 25 moves. An extreme example is really if you have 32man TB, it does not really give value bigger than Stockfish, probably even much worse...as the evaluations are just useless.

[@canary5555]

RIP chess ,😩

[@AbhishekKumar-cn2bu]

canary5555 yo,

[@AbhishekKumar-cn2bu]

canary5555 Chess will never end

[@dannygjk]

Note that the way AZ defeated SF is eerily similar to how a strong human defeats a weaker player. For example Morphy vs weaker player or Tal vs weaker player. People think the lack of an opening book is why SF was crushed. Nope, AZ brutally outplayed SF in game after game. Even when SF played a solid opening AZ outplayed SF *after* the opening.

[@richardfredlund3802]

They still should have tested it against Stockfish with book also.

[@dannygjk]

I hope the SF team gets that chance. I look forward to another annihilation by AZ. Check for yourself. Do you really know many book lines? SF was beaten convincingly even in the games where it played known theory.

[@timothybolshaw]

+Richard Fredlund If looked at purely as an exercise in pitting A0 against the strongest possible chess opposition, you would be correct that including opening books and endgame tablebases should be used. You need to recognize that DeepMind is engaged in a long term project to develop artificial general intelligence. Comparing a neural network derived with no human interaction against the best algorithms and position evaluation created by humans is a GREAT test of artificial intelligence. The use of opening books and endgame tablebases would have obscured the desired comparison. It is satisfying to the ego to see how your tools stack up against the strongest chess, go or shogi opposition, but not part of the big picture. Note that the approach looks likely to work in any "perfect knowledge" system where either sufficiently large volumes of data are already present, or where such data can be created. This includes games more difficult than chess, like go, but also, already, certain real life important issues. For example, it should work very well for power grid design (and actually has already been used to dramatically decrease power requirements in Google data centers). We are still a long way from true artificial general intelligence. Techniques for dealing with unknowns still need further breakthroughs. Also, the inability to generate sufficiently large data sets for most problems means we need to build in the ability (which humans to some extent possess) to use experience with other problem spaces to address the current area of interest. As with most unsolved problems in the area of AI, understanding how humans do it would be very helpful. That said, A0 is an exciting step forward.

[@tzakman8697]

this is the deference between go and chess....

[@abnerlouischarles]

Tzakman and what difference is that if I may ask?

[@user-pj1sy9jd2m]

interesting

[@dilyan-2904]

Hey Tal, let's fire a stream tonight, some bullet action,?

[@hypersonicmonkeybrains3418]

why was the learning limited to 4 HRS??

[@asherasator]

8 hours not 4.

[@TourUser9630]

Your "facts" are wrong. AlphaZero won 28 of the 100 games (72 drawn), but the hardware was not identical. (Even if the move calculation was reported accurately)

[@larsyxa]

Well from what ive heard the actual game hardware was exactly the same, but alphazero trained on a "300000 dollar" rig. Well point is during game AlphaZero did 80000 positions a second. Stockfish did 70 000 000... AZ used a NN and SF used brute force tree search.

[@hypersonicmonkeybrains3418]

So you're saying stockfish is quite a bit stronger than Alpha Zero? right...

[@larsyxa]

No im saying AlphaZero doesnt use brute force. AlphaZero is a neural network that doesnt use the same technique as brute force engines. It more or less filter out the best positions immediately and checks them. Its a little bit more complicated than that. Bottom line is, it doesnt HAVE to check 70000000 positions a second. I am a programmer and done abit of nn programming, but im very FAR from an expert but I know the basic concepts of AlphaGo, AlphaMaster and AlphaZero. I havent read up on AZ so i cant tell you exactly how the network is build. But AlphaGo's neural network training was on pattern recognition. Which positions are more likely to lead to victory. AZ is a general purpose nn that can be used on a large array of problems (games).

[@naveenanil237]

Nope. AZ is an ANN which uses its own games for drawing its own conclusions (learning from experience) whereas SF have different preset conditions to analyse the position. This would limit SF quite severely. Basically, brute-force engines like SF runs on the Infinite monkey theorem with some conditions in place to recognise when the complete work of William Shakespeare comes up. Whereas, a trained ANN would know from experience that monkey+typewriter+infinite tries=works of Shakespeare and it can then immediately give the solution. With no limits on hardware or time, both would be equally strong. I might be wrong tho.

[@edwardshowden5511]

25 wins out of 28 games (with white) and 0 losses with black were during learning period page 5 arxiv.org/pdf/1712.01815.pdf So why in this tournament Deep Mind had 'only" 28 wins out of 100 games? Deep Mind from tests was better trained or what?

[@edwardshowden5511]

thanks tal. your way of talking is a bit boring but WHAT you're talking is insightful and thats more improtant ;)

[@GMTalBaron]

piotr monn I'm aware of it, sadly :)

[@Fcstfan]

bro, put some moves on the board...

[@fredpennington6180]

There are too many moves made by Black which lead to an inferior position. Not developing the Queen side Knight and Rook, moving the knight to b7 and Q to h7 is to guarantee a loss against almost any Expert or class A player.

[@thom1218]

Such basic criticism of Stockfish's play makes it clear that you're neither expert nor a class A player. Stockfish eats grand masters for breakfast, and until a few days ago was considered the strongest chess playing entity on earth. Now AlphaZero has surpassed that by a good margin. You think Stockfish just "forgot" to bring those pieces into the game? AlphaZero would punish your idea with a quick victory with its early pressure on black's king side, while you "developed" those extra pieces, i.e. essentially wasted time in an opening not conducive to bringing them out early as they were needed in this game.

[@fredpennington6180]

Thom1218: I don't know about Stockfish or Alpha; but, look at the position on the board! The game looks like a computer beating a human player who has no concept of returning material to solidify the position; which is a key strategy.

[@dannygjk]

Fred that's what traditional engines do, they emphasize material in their move ordering, (google move ordering if you don't know what it is with regard to chess engines).

[@MegaTp4]

The hardware wasn't equal. Alphago had much more computing power.

[@SlimShady-ri6ej]

MegaTp4 Well if it would learn chess for couple days it could destroy sf with equal power.

[@jean-marcfueri6678]

Nope. Read the paper: SF had 64threads and was calculating 70kn/s. Alphazero 80kn/s.

[@dariusduesentrieb]

stockfish calculated 70,000k evaluations per second, not nodes, that is something different. and the position evaluation-speed is different because of the design, not of the computer power, at least primarily.

[@MegaTp4]

Jean-Marc Fueri You should read the paper again then! Stockfish calculated 70 million positions per second while AlphaZero only calculated 80 thousand positions per second. However each position required a lot more computing power for AlphaZero. AlphaZero runs on 4 TPU's. Each TPU has 45 teraflops so together 180 tflops. There is no 64 threaded cpu that gets even near that power.

[@jean-marcfueri6678]

Yes but since the programs operate differently I am not sure the argument about hardware is that relevant. We don't ask computers to calculate only 2/3 moves per second like humans when they play against it. So the overall performance of A0 is amazing, which is not surprising given how it crushed Go.

[@soMeRandoM670]

You mean millions of games, how long it takes is irrelevant. assuming age of 3, and playing 100 game a day. would be a 1202 year old human player. like, it would taken years run it on software of normal computer. it's played 44 million games. arxiv.org/pdf/1712.01815.pdf

[@youssefm6061]

This program that use google, you only put the inputs( pieces and rules) and outputs ( the results hacke mate) so this machine use algorithms inside that derivads , probabilits that in his train improve his performan , and while more time train more is the best, so the algorith inside is not created for as, is something magic, we only can the hiperparamentros example epcohs, batch, kfold , time of the train. So it is difrent to program of stockfish!!!!!

[@ziquaftynny9285]

Really mideocre, boring, and superfluous commentary.