AlphaZero crushes Stockfish 2017 Game 8 - Positional

[@PowerPlayChess]

👍 *like* 🗯 *comment* ✔️ *subscribe*

[@tensor131]

reminds me of a Karpov game when he trapped the opponent's knight out of the game in the same corner of the board (what was that game???). It is truly amazing. The "gather up the lost sheep" phase was mind blowingly simple yet seemed like we entered a wonderland of chess; the pawn structure that Alpha zero established in the opening ensured that this "sheep manoeuvre" was always an option ... and the way the d file was out of bounds for the black pieces was just incredible - a precise placing of the pieces in perfect harmony with the pawn structure, yet at the same time ensuring that the pieces remained ACTIVE .. in a human game at super GM level that sort of strategy would win every time .. we can learn from this. In awe of all this stuff.

[@AlonsoRules]

AlphaZero is doing something that computers in the past have yet to do - understand not only short term tactics, but long term strategy and planning.

[@dannygjk]

I'm assuming he meant AZ did what amounts to "understanding" by humans. I doubt that he thinks that AZ actually understands it as if it was an AGI.

[@Badbentham]

It "understands" the winning chances (as probability) in a given position, while traditional engines are bound to their evaluation functions, and their (pretty wide) horizon. It teached itself over the course of trillions of games that the Bb7 is a bad piece that dramatically increases its own chances: "Very likely winning" after 13.b4, while SF etc. show 0.00 for a further few dozen moves. ;)

[@gJonii]

Understanding strategy is probably too strong a claim. Rather, AlphaGo has same tactical intuition as humans. It doesn't plan ahead beyond brute force, and it doesn't really build strategy either. It's just brute force aided with intuition of sorts.

[@vohbovohborian28]

What does it mean "to understand" something. We can say "to interpret" or "to comprehend", but we can't really explain it very well. Does understanding require consciousness ? Then according to neuroscience, it doesn't exist, because are only observers of processes that happen elsewhere in the brain. Perhaps we can explain it as: we see something in a certain context, and can then extrapolate from the specific to the general. But that is not unlike what Alpha Zero does, only that it needs thousands of games to "get" something, whereas we would only need a few examples. In this case I would say there is a very rudimentary "understanding" of what goes on, that is not very different from the way we do it ourselves. A0 is certainly much weaker in that aspect than we are, but backed by huge computational power. I honestly don't really see how this is much different from AlphaGo Zero to be honest, just applied to another game, but then I'm not an expert.

[@marten1712]

Honestly, you are arrogant if you claim to know whatever complexities deep neural networks take on.

[@willianbueno7293]

Loved your analysis master.Tks for the video

[@uriah7006]

Thank you for doing a AlphaZero video, love it

[@nilsp9426]

Maybe with some exaggeration one could argue, that Stockfish is a Patzer in assessing what harm moves do to the position in 30 moves, since it is beyond its horizon. Not that humans can exploit it, but there are definitely positional aspects of the position that Stockfish isn't as concerned with as AlphaZero is. And maybe this puts a closure to the never ending mantra of "maybe a supercomputer could draw this, because the "objective" engine evaluation is 0.00, but in practice its almost impossible to hold it with so many positional disadvantages" - well AlphaZero crushes that dream of holding every position that classical engines see as 0.00 or similar.

[@whitenightf3]

Hasbis and his brother both strong English juniors went to the same school as Kingscrusher aka Tryfon Gavriel.

[@universalchesslyfe3813]

Outstanding as usual!!! Thank you :)

[@mpo87-h1c]

It would be great if you could analyze the other games :) Btw, the 4 hours are deceiving. It was using thousands of processors. That is equivalent to years of computing power on a normal machine.

[@authorFreeman]

It used the thousands of processors to build the model. Once the model was built, it didn't need the thousands of processors any more. It ran on a normal machine (albeit with 4 pretty monster TPUs) to actually play the game. This is not unlike other chess programs, in that years of "preprocessing" of various kinds (endgame tablebases, opening books, general refinement of their algorithms through self-play) have gone into the current products.

[@johnd0e25]

Yes! I was hoping to see your insight on those games. Thanks Daniel :)

[@lemoncake7565]

Fantastic game and video, thank you!

[@drob9673]

Thank you for this Daniel. Very interesting to see these games again analyzed from your perspective. I wonder if seeing how Alpa plays will impact your own style?

[@johnkom2339]

I think the advances here that are most impressive are: priority to space advantage, and priority to the idea of cutting off pieces from the action, or reducing their scope. Those are concepts that humans have understood for many years but it seemed that they became demoted in the face of the monstrous tactical power of engines. Now we see that they are still key principles.

[@Youtube_Globetrotter]

I started AlphaZero on my computer, but instead It is a Skynet loading screen 6% done, -Please wait.

[@oozecandy]

Great analysis. One thing I haven't heard mentioned yet concerning "chess's future" is that in contrast to "dry" classical computer vs. computer games up to now, these games are actually pretty exciting and that may not bode well once people begin losing some interest in human vs. human games.

[@weakrecession]

You say the technology isn't available to everybody, which is correct regarding the AlphaZero algorythm. But the open source community has taken up the challenge and there are already fantastic projects, where you can yourself contribute to develop, like for instance Leela and others.

[@cliffactually1379]

That was exciting to watch. Perhaps even more so than many of the recent London Chess Classic games(!) AlphaZero's style of play is definitely easier to get behind than other machines. A game changer (excuse the pun).

[@horrortackleharry]

Still the question remains: would AlphaZero run out of the building if a fire broke out during a game?

[@roqsteady5290]

I'm sure it could handle Alexei Shirov without having to run away!?

[@testthewest123]

Only after it has tried staying in there a couple of times.

[@hansgruber5243]

DeShawn 'Dawg' BNBG Actually, DeepMind has AI that takes on the task of locomotion so walking, running and jumping wouldn't be a problem if AlphaZero was given an extension of that AI. XD kzbin.info/www/bejne/Z2XdnJl6ibOSp9E

[@JimJWalker]

It would probably figure out how to put the fire out.

[@cliffactually1379]

+Jim lol Or figured out how to prevent a fire before it even has a chance to start

[@MrBrain4]

Great analysis. Small comment (applicable to all your videos): It would be nice if the pieces had some motion when moving from one square to another, or some indication of original and destination squares. I always find the videos on this channel to be the hardest to see where the pieces have moved from and to.

[@kyawhtetsoe3407]

thank you!! i ve been waiting your analysis for alphazero, i thought you dont follow machine games. if possible i want to watch all 10 games haha

[@elrenegau0912]

No! He write a book of the match Kasparov-Deep Blue!

[@NawinMallika]

Very nice analyses Dan, thanks

[@danjeory3659]

Best analysis yet for me, thanks. By the way, would be intrigued to know how you arrived at the position on the board over your left shoulder... ;)

[@rafaelq.8506]

man, they really did it. a machine that goes around solving thousand-year-old puzzles like it's nothing. i wonder how long until a version of this fully understands humans emotions and calculates what is the best possible novel that can be written.

[@DarkSkay]

So if I understand this correctly AlphaZero figured out top-notch chess opening play (which took humans ages) without a book just from the chess rules?

[@contriturate5375]

Yes, though in those 4 hours it played more games against itself than a GM would play in 200 lifetimes.

[@DarkSkay]

True. It is amazing how good this learning algorithm is though. Because I know leaving a chess beginner (who knows just the rules) by himself with a chess board won't produce good results, let alone fast results, even after he played many times against himself. BTW it would be interesting to know how AlphaZero played after some human-reachable amount of self-play. Is AlphaZero after say 50 games of self-play better than a dedicated human? So does it learn chess faster (in terms of # of games) than the human brain? I kinda hope the DeepMind software won't do poetry soon ;)

[@DeuceGenius]

@@DarkSkay f that bring on the poetry, lets see how good he does when hes the only one reading his poetry to become better :) im not a big fan of poetry though, let the a.i. make music it will be subliminal sounds to control human brains probably

[@mubasharkhan8815]

Hi Daniel, excellent videos keep up the great work! Just one suggestion: When a piece moves, it might help to "highlight" that square as sometimes one can get lost as to who moved / or whose move it is next - if that makes sense? Thanks !

[@PowerPlayChess]

We had a software update and moves are highlighted in newer videos :)

[@mubasharkhan8815]

Thanks Daniel

[@Martvandelay]

His water drinking skills are amazing

[@elvarg991]

Great job Daniel!

[@aureile01]

Awesome game! Always wanted to see good commentary on computer games

[@testthewest123]

Would you call AlphaZeros gameplay more "human" than the other engines? Do you think it could be a practice/learning tool to get human competitors to a higher level than ever before?

[@hey8174]

OpenAI created a Dota2 bot that feels oddly human in its play style. It baits and switches like a human would to coax over reactions from its opponent. It constantly plays as aggressive as possible on the thin edge of danger, just to take advantage of your first mistake. Things I would think are unique to humans like sarcasm, jokes, deception, etc.. aren't all that unique to a bot using every behavior possible to maximize its fitness criteria. Most professional players have changed the way they play 1v1 matches based on the play style of the OpenAI bot.

[@libertas12]

Thats actually downright scary. Imagine this software implemented in a capable robot-body.

[@hey8174]

Militarize it first, ask questions later.

[@davidblack2970]

The question is how long it's going to be before there are medically implantable devices able to communicate with alpha zero. You'd almost have to CAT scan all of the players before every tournament.

[@roqsteady5290]

It is hard to see how Alpha Zero is doing this: For instance is it just more efficient at pruning useless moves from the forward move tree of possibilities or is it's evaluation of each tree node much more profound or what? And what is it exactly that it is learning when it plays against itself?

[@code_explorations]

Alpha Zero doesn't have a tree per se. It's possible that its neural network develops a tree, but that's kind of hard to believe, and in any case it's not explicitly coded like other engines are. What does it learn by playing itself? It is gaining experience by trial and error. After a massive amount of this training, it has developed supreme positional analysis, but nobody can say how it is evaluating each position. Look for "3blue1brown neural network" on KZbin for a brilliantly presented introduction.

[@roqsteady5290]

Sure it's gaining experience by training, that has got be obvious. But just saying that doesn't really add anything of itself if you don't know what that experience consists of: What are the weightings and how are they applied and stored in the neural network? Its been known for some time (forget the name of the guy who did the research, some dutchman I think) that good chess players remember a vast number of patterns and tend to look at the board in terms of pattern chunks rather than individual moves. Even so calculation is a necessity in some positions - so it seems certain that AZ must be doing that too and that the 80k positions per second is going to be largely involved with concrete calculation. But maybe, like humans, AZ is much more efficient at knowing what moves are relevant to a calculation.

[@femioyekan8184]

We don't know. Check out the video mentioned by Gavin and pay attention to the concept of "hidden layers."

[@ipudisciple]

AlphaZero does have a tree. It's a Monte-Carlo Tree Search, with move selection and evaluation done by a (single) neural net. It looks at about 80 thousand positions per move, as compared to Stockfish's 70 million. Read the paper. arxiv.org/pdf/1712.01815.pdf

[@authorFreeman]

Just to expand on +ipudisciple's reply: a Monte-Carlo Tree Search basically plays out a huge number of random games from the current position, and chooses the move that led to the highest winning percentage among the playouts. The key to making it work really well is to make the "random" moves more meaningful, and that's where the neural net comes in. For every turn of a random playout, it suggests the mostly likely good moves for that position. So in way, you could kind'a sort'a think of it as a conventional tree search, but with pruning of the tree that's so radical, you're actually reaching the end of the game. What I have no idea about is how they arrive at the 80k positions evaluated, because as I understand it, Monte-Carlo does no real evaluation of positions. It just applies the rules of the game to determine who won a given random playout. So do they mean 80k playouts per move? If that's the case, then the number of positions is actually much greater: it would be the total number of moves in all the playouts. Or do they mean the total number of moves, so the number of playouts is much smaller? Neither is very comparable to a conventional chess program's position evaluation.

[@nichonifroa1]

Love all your videos.

[@Varvitski]

Wonderful! Thanks for this fascinating analysis. I'd love to see a tournament with classical time control. You mentioned how at times 'Magnus-like' AlphZero's play was. I have heard others describe it as playing like Morphy. Are they correct, and if so why?

[@2jomok2]

Thanks, Daniel, for an interesting analysis. Having followed Stockfish's development for a few years, I find the AlphaZero games fascinating. One thing that always raised a question in chess programming is that humans are deciding on what is good or bad and giving it a point value. It seems that in true chess programming, the engine should determine for itself piece values, positions and tactics. I think AZ has shown that there is value in space and keeping the opponent locked in. A human values a pawn as 1 and a bishop as 3, but, as in this game, if a bishop is locked in it is not worth 3, and AlphaZero was able to "sacrifice" elsewhere for a won endgame. Same for the connected, passed pawn on c4, traditionally a valuable piece. But AZ determined that the other two pawns in the chain were locked in, so c4 was never going anywhere and thus had very little value. I think that AZ will change the world of chess programming. And letting the new engines figure out their own tactics will eventually spill over to human chess.

[@Chesskana]

Thank you mr King!!!

[@TheMattTempest]

Really dumb question here, compared to some of the brainy stuff below, but when you say at move 68 "the plug is pulled on Stockfish", do computer engines actually resign games, or are they programmed to go on 'fighting' until checkmated?

[@K4inan]

matthew tempest It was programmed to resign. If you play vs stockfish it will resign after 150 moves.

[@xxingww]

Thanks for the video. Would love to see your analysis of other AlphaZero games with SF if possible. Also would love to see Deepmind put AlphaZero into a chess program tournament.

[@guest_informant]

A lot/All of the published material so far is basically a Deep Mind press release isn't it? There's plenty of food for thought there, but I don't think we should be uncritical. A few random thoughts: (1) The first time you see one of these long term positional sacrifices it looks amazing - the fourth or fifth time it begins to look like a bug, or at least a (relative) weakness in Stockfish. (2) I thought I read somewhere that Stockfish's programmers wanted to make adjustments mid-match but weren't allowed to (shades of Deep Blue(?)). Can't find a source though. (3) As plenty of people have pointed out (including Vishy) the claim "four hours" should be taken with a pinch of salt. four hours on their hardware is about a decade on my hardware.. (4) Only ten games (all AZ wins) were "released". What about all the draws? And perhaps even more to the point: What about the handful of defeats in the thematic games played? (5) Stockfish was playing without book or tablebase I think. (6) The time control meant its time management - important to its playing strength at classical time controls - was irrelevant. etc

[@roqsteady5290]

As I pointed out elsewhere, if SF was playing with no book it sure did a good job of replicating some current mainstream theory in the queens indian, especially given that Alpha Zero was first to deviate. And as to tablebase - irrelevant, because SF was already lost before the conditions to use a TB were met.

[@dannygjk]

SF still manages it's time well at various time controls. By the way time management is a non-factor at fixed time/move controls. The book-Even when Stockfish followed theory in these published games AZ outplayed SF after the opening. EGTB-Stockfish had lost positions before the EGTB would be of any use. Hash transposition size-Try it yourself give SF a big hash and see how long it takes SF to see that AZ's sacs were sound. Based on what I have seen in the published games my theory why AZ outplayed SF is that AZ has vastly superior move ordering. This is supported by the fact that SF was doing 70,000,000 nps while AZ was doing only 80,000 nps. Even if SF has a huge transposition hash table that won't be enough to compensate for much inferior move ordering. Inferior move ordering results in too much time wasted on pointless variations. SF will miss crucial variations because of that.

[@carlosladen]

Hi, Mr King. I remember you saying that computer chess doesn't interess you, glad you changed your mind, is chess after all. Please continue.

[@robertofigo6266]

a perfect melody!!!

[@jbennett1618]

different types of closed systems sound funny same algorithm, that was the most intelligent thing I have ever hear a chess commenter say around thirty seconds

[@jbennett1618]

now matter how complex simple duhs in correct sequence

[@jessearthur6776]

At about 10 mins in, u say that as a human we can see for various reasons that black's position is losing but I'm not so sure anybody would be able to analyse that position so beautifully without the benefit of hindsight...Stockfish is better than us ( that's the Royal us, of course!! ;-) ) and it didn't see that so I'm pretty sure that no one on the planet would've been able to take white's position at that point and beat stockfish... Alpha zero seems to have a better understanding of what defines the quality of pieces and we're only able to harp on in a vague manner about quality and synergy of pieces but not really understanding it in clear terms...

[@dannygjk]

There are positions which can be won by a human even tho they can't see all the variations merely because of their positional understanding.

[@stevewild374]

Most beautiful game I have seen a computer play. Reminds me of Bronstein. Sure this will impact human chess too. Great analysis. Thanks

[@michal3141]

It would be cool to see if AlphaZero can beat Stockfish in TCEC format (with a normal time control, tablebases, an opening line to start with, and equal hardware - which is crucial). I think it would be much harder but still A0 has shown an amazing positional understanding. This game is just an illustration that regular engines are still quite bad at recognizing spectator pieces and long term strategic disadvantages. I mean a lot of progress has been made to make chess engines better at this sort of deep strategic play but still not enough. A0 is clearly a revolution in this respect. It would be fascinating to see it winning even against fully equipped SF (with tablebases and more computing power). To me the revolution is in my understanding of chess. I thought it was mostly about exponential search. However, A0 has just proven that that at least 99.9% of search done by engines such as SF, K or H is simply irrelevant. Like you can prune almost everything and still by considering only tiny fraction of possible lines you can beat a monster that is calculating much more. This is what makes me believe that ultimate solving of chess or at least playing near perfection is possible by divine understanding of chess concepts without the need for almost infinite computing power.

[@contriturate5375]

A0 can't run on CPUs, so TCEC would have to get some extra hardware. A0 needs about $6000 in GPUs.

[@michal3141]

They can make it to run on CPU, can't they? I mean A0 is a regular computer program. I don't know the details but I would assume that training A0 requires a lot of computational power but running should be possible even on personal computers. If it isn't possible to run the match on TCEC hardware, they can at least run SF on a supercomputer with similar power to the GPUs on which A0 is running and give SF tablebases. It will make for a perfect battle what is the strongest chess entity of Earth.

[@contriturate5375]

CPUs have a handful of very powerful cores. A0 and similar need thousands of cores running in parallel, though each core can be much weaker than a CPU core. Google used custom hardware, but retail GPUs provide the same thing with a bit less efficiency.

[@dannygjk]

SF still manages it's time well at various time controls. By the way time management is a non-factor at fixed time/move controls. The book-Even when Stockfish followed theory in these published games AZ outplayed SF after the opening. EGTB-Stockfish had lost positions before the EGTB would be of any use. Hash transposition size-Try it yourself give SF a big hash and see how long it takes SF to see that AZ's sacs were sound. Based on what I have seen in the published games my theory why AZ outplayed SF is that AZ has vastly superior move ordering. This is supported by the fact that SF was doing 70,000,000 nps while AZ was doing only 80,000 nps. Even if SF has a huge transposition hash table that won't be enough to compensate for much inferior move ordering. Inferior move ordering results in too much time wasted on pointless variations. SF will miss crucial variations because of that.

[@vittoriopaternostro]

Thanks!

[@arnieus866]

Doesn't matter if the match was strictly fair as some claim it wasn't. AlphaZero accomplished in a matter of hours what it took thousands of humans to accomplish in 600 years. This is a critical juncture in the history of humanity. Will this kind of technology be used to solve humanities many problems or will those who would rule the planet employ it for that purpose. The implications are both promising and terrifying.

[@K4inan]

Arnieus Yeah... but AlphaZero played hundreds of thousands of games in those 4 hours. How many games have you played? :p

[@dannygjk]

+K4inan Irrelevant

[@Stefan0v1]

It played almost 20 mil.

[@BixenteFabregas]

Is one minute per move for Stockfish relevant for an engine game ?

[@dannygjk]

Irrelevant

[@maxpheby7287]

First Go then chess now the world? I wonder which top pro will get there hands on AlphaZero first? We should be able to tell when one of them go's on a winning streak

[@K4inan]

Max Pheby It won't be able to evaluate moves like how stockfish does. But I really want to see A0 pkay versus a National Master or something, just to clearly see the difference in play from an engine, and skill.

[@DarkSkay]

Fighting and amazing this algorithm can beat (in a couple hours) 1) hundreds of years of human chess experience partially coded into Stockfish evaluation function 2) paired with all the engineering talent it takes to make a good engine!!

[@JimJWalker]

I want to see a match where Stockfish gets to use its opening book. Or better yet, have a strong GM using Stockfish assistance to play AlphaZero!

[@helpless4625]

Even when SF followed theory AZ dominated after the opening. Also AZ has apparently discovered that some opening systems are so inferior that it stopped playing them. If SF gets a rematch then imo the SF team had better make big restrictions on which openings and lines it will play against AZ.

[@pytokland]

Beautiful game. If that's how their 4-hour-old babies play chess, I don't want to be around at their terrible two's - or "Skynet phase", if you prefer.

[@raamshankar4121]

If Alpha Zero starts loving ones' girl friend...switch off the power (if self charging initiation not programmed).

[@balazsio]

As I heard Stockfish wasn't the latest version.

[@dannygjk]

SF 8 was the latest version at the time of that match.

[@Wind.celestial9842]

How could stockfish panic and play c3 ? Stockfish can panic too? :3

[@koroshiya5758]

Btw that is a 2900 level skin & hair game

[@thomasradar1860]

You should look into getting an audio consultant and or a new mic.

[@dannygjk]

I hear him quite clearly.

[@bowskiechessplaya3337]

68 moves and stockfish was dead for 64 of them...RIP stockfish

[@bruceli9094]

Any A.I smart enough to pass the Turing test will know to fail it.

[@dannygjk]

How about giving the original person who stated that credit?

[@flamingxombie]

It was sort of like a Geller Gambit.

[@koroshiya5758]

I bet the power socket thing behind him is a British 3 hole one. Just an observation... and some trauma... never forget to bring an adapter... note to self.

[@praveennarayanan9451]

Although it is a fascinating - very human even - display by AlphaZero here, I doubt if any reasonable human player would play like stockfish, allowing the LSB to be trapped forever, even with the extra pawn.

[@harrypounds456]

Stock fish is still stronger, people never mention the time control played on and the fact that stock fish heavily relies on opening books

[@dannygjk]

SF still manages it's time well at various time controls. By the way time management is a non-factor at fixed time/move controls. The book-Even when Stockfish followed theory in these published games AZ outplayed SF after the opening. EGTB-Stockfish had lost positions before the EGTB would be of any use. Hash transposition size-Try it yourself give SF a big hash and see how long it takes SF to see that AZ's sacs were sound. Based on what I have seen in the published games my theory why AZ outplayed SF is that AZ has vastly superior move ordering. This is supported by the fact that SF was doing 70,000,000 nps while AZ was doing only 80,000 nps. Even if SF has a huge transposition hash table that won't be enough to compensate for much inferior move ordering. Inferior move ordering results in too much time wasted on pointless variations. SF will miss crucial variations because of that.

[@untwerf]

Hi Daniel, I have found Roman's Lab, the entire collection of 100+ DVDs, being sold for approximately $1000 online... do you think this would be a good value-for-money investment with respect to chess training material? I think Roman produced his videos in the early noughties, so I'm not sure if some of the ideas are outdated.. I'm also generally not sure about how good a teacher Roman is..

[@tgwnn]

untwerf I'm not Daniel, but from that money you could get GM coaching for several months, I would go for that instead.

[@roqsteady5290]

It would hardly be gentlemanly for DK to blow his own trumpet, but I can tell you that Dzindzis stuff is OK (even if he can be a bit "optimistic" about the openings he recommends), however Daniel King's own material on DvD is better and more interesting. If you have $1000 burning a hole in your pocket you could do worse than to buy a few of those, because quality > quantity. And it does seem a little inappropriate to ask a content author to recommend someone else's content, particularly on a video that has nothing to do with the subject. You'd be better off discussing on chess reddit or something.

[@untwerf]

Roq Steady I enjoyed your reply. Thank you!

[@JimJWalker]

Dzindzi is a great player, but only an average commenter due to his thick accent. This just my opinion.

[@postnubilaphoebus96]

If you want to improve, I also recommend spending a lot of money on tournaments.

[@etienne7774]

They only show us the games Alpha Zero won!

[@canary5555]

Alphapocalypsezero . Period

[@john-r-edge]

IT would be interesting to see alphazero play some chess 960 against a human. Has it learned enough about play in general to deal with a human who would be without their regular Opening repertoire. There are other issues raised here for humanity. Fast learning AI gone rogue, while the trope of much fiction, takes another step towards it being a credible existential threat.

[@louismiller4374]

Sie sind der beste :)

[@tensorcats50]

I came here from this retweet by Demis: twitter.com/demishassabis/status/942344630712053760 This was the first chess analysis I've ever watched, and I learned quite a bit! Has the progress in Machine Learning weakened or strengthened your passion or chess (I mean people reading my comment, not just Daniel King)

[@oldmanc2]

Tensor Cats Same.

[@johnkom2339]

Very impressive game. AI such as shown here is dangerous, Stephen Hawking is right. If this is what AI can do, then we have a lot more to worry about than simply getting thrashed by a chess engine.

[@K4inan]

John Kom No we don't. The whole AI thing being bad in the future is just superstition.

[@dannygjk]

It won't be bad like Terminator bad it will be a different bad.

[@archaeopteryx7405]

Alpha-Zero will kill chess soon. The artificial intelligence algoritm, based on the theory of neural networks is not the real innovation. These algoritms are well known from long time. The real innovation is to create an hardware able to implement such algoritms! When the computers based on TPUs architecture will be available on the market, the chess knowledge will be completely revolutioned. The chess truth will be disclosed to everyone. Then the game will lose any appeal, unless to change its rules. Considering how fast computer technology is grown in the last two decades, I suppose this will happen in less than three years. It is sad to say goodbye to the most wonderful board game of the world, but this is the price to pay to technology evolution.

[@tarunt6872]

Carlsen is nowhere near alpha zeros level. Stop comparing anything and everything to carlsen

[@bferi]

I'm not much of a Carlsen fan but I have to agree with Daniel's assessment - alphazero's style in this game was really similar to that of Carlsen's.

[@tarunt6872]

You have to understand that alphazero has no specific style, (check out Bg5 sacrifice game) Also Kramnik, is more of a positional mastermind than carlsen if you intend to do some comparision. just because he is currently world no 1, doesn't mean every brilliancy has to be attributed to his style.

[@K4inan]

Tarun T Yeah, no style. A0 adapts and steamrolls. :D

[@sarbasov]

I don't believe AlphaZero. Chess is about calculating a lot of moves ahead, and classical chess engine is best adapted to that. I think that Google cheated with giving higher performance hardware to AlphaZero or not giving Stockfish opening and ending books.

[@Piecemaker1975]

In addition to calculating, one needs skill. And this is the step of so called supervised machine learning. Stockfish needed the input of people regarding e.g. king safety etc. Supervised machine learning works in an other way: play, classify the results, use this classification to improve, repeat, no input from GM's needed. This way SkyNet will eventually wipe out human kind, because it can improve without humans.

[@pmlonline]

AlphaZero ran on an army of PCs. It's like expecting a program on a 1980s calculator to win one running on a high end PC.

[@dannygjk]

Incorrect.

[@ziyadalkilic]

Dont you just wish that someone would decode the unintelligible wisdom in alpha's nodes into conjectures like; "When you have space dont trade down", etc. I think such a thing might become a job title for humans, if likes of alpha prooves useful in too many areas. And also the "rules" of chess were given. Someone will have the fantastic job of defining those rules for applications where this is not clear. Ah, fascinating times...

[@roqsteady5290]

Maybe it is too messy to be condensible into rules. When we think, millions of different neurons may be involved in a decision and each of those neurons are themselves subtly affected by many incoming synapses as to whether they fire or not and support or inhibit some particular response. It seems likely that the difficulty we have in understanding the mind and consciousness stems largely from this. And it maybe that we will never be able to understand mind in the kind of mathematical theorem way that we tend to think of as constituting understanding.

[@ziyadalkilic]

Well, I see your point. But, I think the "neural network" tag we associate with a problem solving mechanism does not necessarily imply that it is a one to one imitation of how brain works. And personally, i would write off anything Alpha or other similar technologies could achieve, as not dependent on conciousness. Take image recognition AIs for example, show them a million cats and cat-nots, they decipher random images for cattiness with a powerfull vigor. So, there must be an ad-hoc definition of the concept of a cat, in those nodes, entagled, not necessarily organized, yes, but it shall be there. However, the AI is not "AWARE" of the concept it just uncovered. Thats why, when you change a single pixel it can mistake a cat for a car. However, i see the value in yet to be organized and distilled raw information in such AIs after they are trained at a specific problem. We may not just merely run them. We may try to uncover a deeper understanding of the problem at hand through them. (Understandee being a human here)

[@contriturate5375]

That's one of the big problems of the deep learning algorithm: it works marvelously on many problems, but nobody has any idea *why* it works, how to understand the result, or how to change its behavior aside from retraining. The AI's mind is an incomprehensible black box after training.

[@ziyadalkilic]

Just hoping it doesn't stay that way for too long. Oh, the goodies that could come out of this...

[@Fiercygoat]

Didn't know we had neuroscientists hanging around on Daniel's chess youtube videos.

[@Piecemaker1975]

Panic is a human emotion. No chance Stockfished "panicked".

[@PowerPlayChess]

Irony...?

[@1stJJ]

This is the second time I've seen a computer "panic" within 2 days.... (The other being Star Wars' C-3PO)

[@dannygjk]

+Piecemaker1975 If you knew how engines work you would know why engines do that. Plus making a move which prolongs the time until the end is the engine equivalent of panic.

[@brandondaniels9471]

Stockfish was hung over from a wild night at the strip club... _Kappa_

[@giorgosplatias4227]

You are passing the variation as if they were ancient knowledge! I did not understand anything. Somewhat poor explaining, arbitrary assessments on the evaluation skills of the machines. Only a cursory glance of the crucial moments of the game!!