Giuoco Forte | DeepMind's AlphaZero Game Changer 4 | Stockfish 8 vs AlphaZero

Daniel King shows the 4th game of the Game changer series between DeepMind's AlphaZero and Stockfish 8.
Support on Patreon: ► / powerplaychess
►Support via PayPal ($): www.paypal.com/cgi-bin/webscr...
►Support via PayPal (€): www.paypal.com/cgi-bin/webscr...
►Support via PayPal (£): www.paypal.com/cgi-bin/webscr...
✔ SUBSCRIBE: bit.ly/powerplaysubscription
👕 MERCH:
teespring.com/stores/power-pl...
www.redbubble.com/people/Dani...
📺 ALPHAZERO: bit.ly/alphazerostockfish
📺 WORLD CHAMPIONSHIP 2018: bit.ly/worldchess2018
📺 FABIANO CARUANA: bit.ly/fabcaruana
📺 MAGNUS CARLSEN: bit.ly/wcmagnuscarlsen
📺 BOBBY FISCHER: bit.ly/rjbobbyfischer
📺 SPASSKY bit.ly/borisspassky
📺 TAL bit.ly/mikhailtal
📺 KORCHNOI bit.ly/viktorkorchnoi
📺 CHESS CLASSICS bit.ly/chessclassics
📺 BRILLIANCY: bit.ly/brilliancychess
📺 TOP10 2017: bit.ly/top10chess2017
🎥 NEW: bit.ly/newchessvideos
💿 Power Play DVDs: bit.ly/chessbasedvds
►Facebook: bit.ly/danielkingfacebook
►Twitter: / danielkingchess
#alphazerostockfish #alphazero #deepmind

[@PowerPlayChess]

If you are interested in the peer-reviewed paper from Deep Mind, then here it is: science.sciencemag.org/content/362/6419/1140

[@eh01uil]

PowerPlayChess ／

[@patrickbertrand6366]

Danny is simply the best!

[@kolektivmozak238]

Great game. Daniel it would be nice to see Engine evaluations on the screen so we can see when Stockfish figures out that things are not going it's way, and we can see difference in understanding positions by looking at evaluation from both sides.

[@proflaxis6968]

You are very generous with your early Xmas presents GM King. 20 games so 20 days of Xmas for us chess fans. The Game Changer series title is so appropriate.

[@rb5955]

Daniel, you are spoiling us. Such fantastic treats every day... I agree, alpha zero is using its bishops fantastically well. It's a lesson for me. I should keep my bishops in my games. Looking forward to the next one...

[@Phurngirathaana]

Fascinating analysis Hard to appreciate these subtleties on my own Thanks for these videos Danny keep them coming!

[@AnandSivaram22]

"But we hope so anyway" -Daniel King, 2018

[@northshores7319]

Spassky once said of Fischer that his strength was so direct that it was also his weakness. It looks like thanks to a computer that Chess will return to the Golden Age of sacrifice and winning games rather than the precision draws so common. From move ...18 on it reminds me of the play of Tall and Fischer. Thanks for the great videos.

[@muhammadnurarifns1134]

Gonna use Daniel logic from now on. Alpha zero definitely a drunk man from random pub.

[@tomandband]

aha a drunk alien maybe

[@Cronus66]

love powerplaychess!

[@ElColombre27360]

That "Gioco Forte" came unexpected and made my day 😂!!! The pun was appropriate and very good, as well as your italian pronunciation! Well done!

[@sumitstir]

These videos are some of the most entertaining content on the internet. Best commentary covering the best chess player.

[@ardweaden]

I'm loving this series. The games are extremely exciting and the commentary is fantastic as usual. I must say that these games give me a lot of hope for the future of human chess, because it seems that it's still possible to have an aggressive and attacking style even at the highest level of chess.

[@pazdziochowaty]

"How many humans would play Ne7?" If I played after 5-6 beers and would decide that I want to push for f5 I would play that. Often times I go for my plans and do not care about opponent's threats. Unfortunately those threats happen to kill me in the process which seems to not be the case for Alpha Zero :)

[@PuzzleQodec]

Oh I would play it too. I'd miss that b7 was hanging.

[@chrisiver8506]

many 1200s would play it lol

[@shatraoctober6093]

thanks Danny for this beautifull videos

[@KrishnaKumar-np3tw]

All of these are extremely interesting (I wonder how many were dull games out of the 210 or so games played) . Superb commentary demystifying many of the moves.

[@fingerprince_]

Games and commentary both sublime

[@aljehine]

Thank you again :)

[@davatrof]

I LOVE THESE GAMES THANKS DANIEL!

[@kikaze2000]

The only player I think would even consider these types of moves would have been Tal. He about the only super elite player that would do sacrifices without calculating to a mate just based on the fact that those sacrifices compromised the opponents kings safety and gave him superior piece activity. A lot of Tal greatest games similar to what Alpha Zero does on steroids is that the evaluate the resulting position after the sacrifice :" I'm not sure if it's winning but worse case scenario it's going to give my opponent a lot trouble"

[@nilsp9426]

I love the sacrifices of A0, where Stockfish has additional pawns, which are still far back. So A0 has about 5-6 moves to win on the other side, before these pawns get dangerous. But in a real game as a real person, it would be so hard to judge if you can actually win the other side of the board. Especially in these genius A0 games where the other side seems still relatively even and there are still a few moves before an attack even begins.

[@Guitare_picking]

2018 : when computer games start to become more exciting than WCC.

[@moontan.]

Another outstanding video! When Black wins, why not turn the board around so that we can see the game from Black's point of view?

[@sashi247]

Another common theme with A0 is that it prefers tucking the king in the corner and does not hesitate to go all out on the kingside... another apparently in-human trait

[@yorick021]

Mr King, what I am really curious about...has AlphaZero exclusively learned from playing against itself and therefore started the match against Stockfish as a "finished product" or does it keep learning from every game it plays? In other words, can it adapt to its opponent over time? I wonder because it seems that AlphaZero has no problems sacrificing pawns and where most engines are really materialistic it then may have learned to beat Stockfish by sacrificing material and using this time to bring disharmony and loss of time. And would it then also play differently against a very agressive human player like Kasparov or Shirov over like 2000 games? Today I tried to incorporate some of the general AlphaZero strategy into my own blitz games and so far playing a bit more agressively, trying not to worry about material too much (without getting too crazy) is paying off ;)

[@JediJess1]

NGL, I found this video while trying to look up if Giuoco Forte was a legit opening or not. I didn't really get my answer, but I was not disappointed. Great match! I was really impressed with how AlphaZero was down a Knight in material and still managed to keep up huge pressure. I felt that AlphaZero was winning or in a better position a majority of the time despite the lost material. Grande Giuoco!

[@PowerPlayChess]

Giuoco Forte is just my pun on the usual Giuoco Piano :) So, no, it’s not a ‘legit’ opening as such, just my comment on the power of the game.

[@rjp63vip]

I would love to see alpha zero play the grob,the Kings gambit,the Evans gambit, I think alpha zero plays like paul Morphy , fast development, play on both sides of the board and sound sacrifices. Alphazero is bringing sexy back to chess via gambit style play! Draws are becoming extinct!😉🙄👍🏿💃🏽 great job! I love your channel!

[@HiReeZin]

Perhaps the great attacking masters of the 19th century were right after all. Professional players of today who must take care of their living dare not to sac unless their Stockfish tells it's 101% safe. One could think that was also the problem for some expert consults of past top programs like Deep Blue. Too much risk avoiding does bad to your mental health.

[@drutgat2]

Thanks for this, Daniel. Has AlphaZero actually created its own preferred openings and lines over the the time since it learned chess (presumably based on what works and what wins)? Or, does it base its ideas on ideas about valued pieces, such as its preference that you mentioned for using two Bishops?

[@Andrew-0815]

In my opinion these computer games are much more interesting than all games of the last championship. Unfortunately. Maybe computer chess is the future...

[@hoahuynh2734]

This is absolutely not good comparison. Cars in Daytona 500 runs at 300+ mph, but track and field events in Olympics are still wonderful spectacles.

[@GNU_Linux_for_good]

No, no - *not maybe* ! We're history..

[@Macgki]

Well out of a thousand games this is one example, also alphazero "only" won ~150 of them, so don't expect computerchess to be a spectacle like this all the time ^^

[@DeuceGenius]

how about computer intelligence is the future? it wont be just chess :)

[@MEME-qe4ze]

you and alpha zero are awesome!

[@feartactics]

Sounded like there was a story behind your analogy of pub chess. Perhaps a separate video will be forthcoming on that subject?

[@clairedemorgan5695]

Well Danny i am just going up to get the round in, what can i get you and more importantly what is AlphaZero drinking?! It is on another level!

[@lorendisney5068]

I used to want to write a chess program that would evaluate positions by subtracting the total legal moves for each side at any level of analysis. I wonder if alpha hit on something similar in its training.

[@HiReeZin]

I read from wikipedia King's chess journalism is not so much of a topical type but rather instructive one. Now I could not imagine what is more topical than this. Chesswise and otherwise. As a funny side note, as a non native English speaker I never would remember that 's thing right without the knowledge of King's Gambit and KID.

[@roqsteady5290]

Agreed DK covers all the major chess events and gives insight into current chess ideas. Always my first port of call, even given there are now many other good chess channels.

[@gurjassinghbatra5758]

Hello Mr. King. Do you also have any games in which AlphaZero lost( it lost 6 games). And if not, could you request the people at DeepMind to give. By analysing them, we would be able to understand AlphaZero's weaknesses and more about how it thinks.

[@coyote000]

How does the timer know how long the recording is going to be while you record it?

[@moonboy5851]

I could watch AlphaZero games all day!! Thanks for this. Question - is AlphaZero getting stronger after ever game? Or is it at its maximum?

[@TalEl59]

Good question!

[@Zardoz37]

The a2-g8 diagonal looks like a good one for White's Bishop - can White take a leaf out of Black's book and play for piece activity instead of material with a move like 13.Ba2?

[@AnandSivaram22]

Danny I'd like to ask why your 'tone' went in a direction as to black is the one clearly pressing right after the opening? Is it objectively the case or is it just that we are viewing it from an A0 perspective? I find it hard to believe that an entity like sf will slump into a slightly worse position right after the opening..

[@stiandybbi838]

AnandSivaram22 jhjb

[@julioandresgomez3201]

What I didn´t like from A 0 is giving up the g3 pawn. That was so much more than a pawn to give it away so easily.

[@S3v3n13tt3r5]

Bishops > knights confirmed

[@JJ-kl7eq]

Making bad puns on chess videos is my joke-o forte.

[@Kobs.A]

😂😂😂

[@rjp63vip]

Hello Daniel have you checked out the chess puzzlers videos on three of these alpha vs Stockfish games? Please do ASAP! I've watched three of his videos on these games and I'm convinced that Stockfish has not been using its best program. Why wasn't Stockfish 9 used and why has Stockfish 8 blundered 3 times in various games but very rarely blunders in the TCEC tournaments? I would like to see alpha zero compete in the TCEC against the top engines at their highest level. In a nutshell could this artificial intelligence MATCH be ARTIFICIAL? If you gave the top 5 GMs in the world the critical positions of the three suspected games The Chess Puzzler did videos on, would these GMs make the same blunders Stockfish made in each game? I think Karpov,Magnus,and Kasparov should take a close look at these games. The games are enjoyable but I'm wondering if they are authentic ?

[@PowerPlayChess]

At the beginning of 2018 Deep Mind contacted the team behind Stockfish and asked for their most advanced version. The matches described in the academic paper were played in January. The paper was submitted in March 2018 to be peer reviewed. The paper was peer reviewed and published last week. Please forget your conspiracy theories and consider exactly what Deep Mind have achieved.

[@Socrates...]

Would be nice if they had the AlphaZero engine coupled with a chess tutor to explain this paradigm of chess from the bottom up.

[@Prometheus4096]

How? The neural network doesn't know why it is doing something. A chess engine does (though it might be wrong).

[@Socrates...]

@@Prometheus4096 If a neural network doesn't know why it does something, how does it learn and improve? I don't think what you wrote makes sense.

[@Prometheus4096]

@@Socrates... Because the code executing the neural network is different from the code that trains the neural network. And the code that trains the neural network only adjusts weights of the neurons in the network so that it wins more often. It is in no way obvious to know how changing a certain weight will result in different moves for certain positions.

[@roqsteady5290]

Unfortunately, the days of pubs devoted to chess are all but over. I used to play for the King's Head club in the pub of that name in Bayswater, London. Unfortunately, we were thrown out a very long time ago now. I reckon they thought we were a drunken, disreputable lot who had a negative affect on new customers. Maybe there was some truth in that, but was a lot of fun while it lasted trying to beat drunken masters at 5 minute chess.

[@BenNCM]

My God!! Danny, you must realise that you haven't been this excited about about chess in years!!! Has AlphaZero not brought us into the Neo-Romantic age of Chess?

[@davidblue819]

We have come around in a great circle. Judging by AlphaZero, Adolf Anderssen was right about how to play chess.

[@sam-lz6pi]

Now I know why I never became a strong player. Too sober. BTW, I would imagine M. Tal play a move like Ne7.

[@burt591]

AlphaZero loves his boshops

[@MistahGamah]

AlphaZero is the ultimate barbarian

[@Kobs.A]

Kasparov mentality

[@phamduc1987]

Mr King, Could you show us one game where Stokefish wins, and explain why Alphazero got beaten ?

[@hooplehead1019]

Most of the times it was - and it still is as shown during TCEC Season 14 - because of her shallower horizon: Her calculating depth is on average 10 moves less than Stockfish or Komodo or Houdini. In certain positions, this is fatal.

[@skakdosmer]

What are your thoughts on these reactions: »Grandmaster Hikaru Nakamura was less impressed, and stated "I don't necessarily put a lot of credibility in the results simply because my understanding is that AlphaZero is basically using the Google supercomputer and Stockfish doesn't run on that hardware; Stockfish was basically running on what would be my laptop.« and: »Stockfish developer Tord Romstad responded with "The match results by themselves are not particularly meaningful because of the rather strange choice of time controls and Stockfish parameter settings: The games were played at a fixed time of 1 minute/move, which means that Stockfish has no use of its time management heuristics (lot of effort has been put into making Stockfish identify critical points in the game and decide when to spend some extra time on a move; at a fixed time per move, the strength will suffer significantly). The version of Stockfish used is one year old, was playing with far more search threads than has ever received any significant amount of testing, and had way too small hash tables for the number of threads.«?

[@PowerPlayChess]

I suggest you read the paper given in the link at the top of these comments to get details of hardware and time controls of the matches played in January 2018.

[@skakdosmer]

Thank you.

[@skakdosmer]

@Thinking Zzzzz.... Excuse me?

[@elvarg991]

When Daniel stops trying to look like Richard Gere, I will subscribe to his channel :)

[@jodymccullough5891]

Cmon now, Daniel looks WAY better than Richard Gere.

[@betabenja]

Please use technical terms in your videos - a 'mop up operation' is called a 'moperation' in elite chess circles.

[@bulentkirca3311]

Personnally, being a decent player, I do not undrestand well the chess played by these machines :) :)

[@nenadkasnerov2656]

Alphazero was better than stockfish last year and this year. So why leela chess zero isn't best in TCEC?

[@pazdziochowaty]

"This year" was actually only one month after the first match so it was still version 8 of Stockfish. SF has improved from that time over 100 elo points

[@simohayha6031]

Leela hasn't played as many games as alpha yet and therefore isn't as stronh

[@HiReeZin]

.

[@simohayha6031]

@@HiReeZin why is everyone crying about that match of last year with unfair hardware. Read the new alpha0 paper on chess.com. They played 1000 fair matches and even with time odds did A0 beat SF

[@HiReeZin]

@@simohayha6031 At least you could have provided the direct link for the information. But I do the work for a lazy asshole: science.sciencemag.org/content/sci/suppl/2018/12/05/362.6419.1140.DC1/aar6404-Silver-SM.pdf (p.17, Dec 7th 2018) Google matched their AI against 2016 TCEC Stockfish 8. Possibly in the future they match it against 2019 Stockfish? The point is, If you demand odds in chess or in life, don't brag if you happen to win. #positivediscrimination

[@FMDD168]

No mention made that the Actual losing move for Stockfish was 51. Ke2? when 51.Bd1! Rg8 52.Kf4 would have held the game. The Devil is in the Details, Danny.

[@kb7890]

Yeah sadly, my moves are not at all backed up by anything more than alcohol.

[@suyashsharma7949]

Bishop>Knights.Danny had known this way before.

[@smashu2]

White Ng5? was a mistake white should have keep normal development.

[@sujiththiyagarajan4290]

Nice..

[@iconicon5642]

a0 plays like a drunk because like a drunk it doesn't care whether it wins or loses....until of course the drunk punches you in the face. Perhaps a0 would change its tactics if somebody pulled the plug if it lost.

[@uniquebunnie333]

Is it true that Google was helping the team Caruana and team Carlsen was aware of this fact because of which Carlsen was compelled to play the way he played in standard time control games??

[@V8SupersQirreL]

May be you're in love with AlphaZero - but you have to be careful! This partner is not easy to handle. She makes strange decisions...

[@jprw]

Title misleading looks like AlphaZero is playing White

[@HiReeZin]

For one thing as Nakamura said, if you give Stockfish a laptop and AlfaZero a supercomputer hardware to work on, you can't expect Stockfish to survive. We can believe that Google has achieved a good feat, but much of it is exaggerated and Google as a company naturally is after the profit from media hype (of what this video naturally is part of, good or bad). If I'm not wrong they gave SF 1 GB, when in CCC engines do have 256 GB and so on. Further, IIRC Leela is a community project and gathering enough volunteers and calculating power to train the neural network equal to Googles AZ was not possible. So they have scaled the system down to speed up the optimisation process. Also the harware is not free. Those TPU's alone are some 10,000 euros worth apiece. Thirdly these games are selected by Google from the best achievements of AlfaZero, leaving out all the blunders. If you recall how say Alehin worked to secure his right to challenge Capablanca in the eyes of public, you are not surprised. Alehin even twisted the game moves to make him look better.. but certainly Google wouldn't do that. Nevertheless this is interesting developement of things. Please do not take the above to read I don't believe these neurotic entities can play strongly. Also they have created a base for a succesfull universal gamer that is not an expert system an sich. But everyone must understand we have here a duplication of 1997: the 1st time for 20 years some big company saw it profitable to invest millions of euros to chess Turk to get publicity for their AI research. No-one in his right mind argues that if you install Stockfish to modern top supercomputer it could not achieve any ELO you ever like - 4000, 5000, you name it. That said, Chess is still a worthy hobby to us mortals. Intellectually challenging, an intriguing sport and an excellent road to selfdevelopement.

[@warrendsmith6832]

"If I'm not wrong" well you are wrong: SF was given 44 cores in parallel and 32 GB and 6-man endgame tablebases and (in one match) also the "brainfish" opening book, in the test matches versus A0. SF in short was given the same hardware & software it had when it won the 2016 TCEC. In most matches they used SF8 (the 2016 TCEC champ) but in one match they used SF9. All matches won by A0 over SF by large margins. The 1000 game "main" match A0 vs SF8 was 155 wins, 6 losses, rest drawn. A0 ran on a machine equipped with 4 TPU chips, which did the bulk of the work.

[@warrendsmith6832]

If SF10 were used and the endgame tablebase with 7-men (both now available) I do not believe it would do much better vs A0. The best SF version versus A0 tried was SF8 with brainfish opening book. Brainfish book helped more than upgrading to SF9. But I think upgrading the memory RAM size to 256 GB would have noticeably helped SF because SF was searching 58M positions per second. If it stored those in hash table using say 16 bytes per position then that is 1 GB consumed per second, so for a 256 second think it would want 256 GB of RAM to avoid having to worry about conserving memory. (SF if it runs out of memory tries to prioritize storing more important positions only.)

[@HiReeZin]

@@warrendsmith6832 A full year in a competitive sport like (computer) chess is a long time. Remarkably Stockfish staff still is doing it for free, even if they could have made themselves millionaires selling the program rights to some commercial operator. SF naturally needs it's opening book because AZ has one as kind of organic. That's just fair. Otherwise SF has no idea what openings are bad what good and it plays openings at a level of some year 1800 La Bourdonnais. But AZ has that knowledge thanks to "big data" (probabilities drawn from tens of millions of super GM level games. No fool's mates there).

[@warrendsmith6832]

@@HiReeZin A0's training games each lasted about 4 seconds which may not have been "super GM level." And A0 undoubtably would have been even stronger if it too had access to a self-generated book; it is true its neural net in some sense acts like a book but not as much so as if it actually genuinely had a book. In particular A0 has to think a long time about even its first move, with a book it would play it instantly. SF's "brainfish" book supposedly is partly self-generated and always-improving, but I'm not sure it has had a whole lot of cycles put into it so far. The same brainfish code could have been used with A0 instead of with SF (it supposedly will work as an add-on to any chess program) to accomplish that self-generation.

[@zastava9412]

They run AZ on machine like 100 times more powerful than opponents machine. Elo calculator gives AZ on THIS machine +52 elo compared to SF8. While sf10 is +110 elo on IDENTICAL machine

[@synchronium24]

"They run AZ on machine like 100 times more powerful than opponents machine. " Where are you coming up with this? Both AZ and Stockfish were run on 44 CPUs.

[@cyrenaique]

@@synchronium24 the training phase of AZ is done on TPU/GPU, the prediction /match isn't computationally intensive... it's like giving stockfish a huge database and cherry pick the best probability of win

[@synchronium24]

@@cyrenaique "the training phase of AZ is done on TPU/GPU" Sure, but I don't consider the training phase part of the actual match against Stockfish. As long as they run on equal hardware during the match, I think Stockfish just needs to be able to use it's usual opening book for the conditions to be fair.

[@zastava9412]

@@synchronium24 4 TPU were used on game itself not in a training phase. Which was done with even much better hardware. According to generally availalable information TPU is x20 more powerful than normal multicore CPU. That is in addition to 44 cores. They didnt say that processors were the same or similar. Even if they do who can verify this.

[@Prometheus4096]

So Alphazero lost several games vs an old version of Stockfish running on subpar hardware/time control while making weak opening moves because no opening book. So we know the Stockfish from TCEC is actually stronger than A0.

[@Prometheus4096]

@Thinking Z Yes, so under the conditions set by DeepMind, A0 still lost 6 times. So under fairer conditions, Stockfish will probably win. Last year we were speculating it might be 1000 elo points higher than the best engine. Big difference ! Still, letting a bunch of engines play 1000 games and commentating on the best one is a lot more fun that two superGMs drawing 12 games.