Deep Mind AI Alpha Zero Dismantles Stockfish's French Defense
Вставка
- Опубліковано 29 вер 2024
- Download Mproov and Improve Your Chess Today! app.mproov.me/...
Follow MprooV on Twitter / mproovapp #agadmator Check out all my videos on this match
• Google Deep Mind Alpha...
Read more about Deep Mind Alpha Zero here arxiv.org/pdf/...
Link to the other games lichess.org/st...
A chess game between Deep Mind Alpha Zero and Stockfish
Google Deep Mind Alpha Zero vs Stockfish
One of the games
1. d4 e6 2. e4 d5 3. Nc3 Nf6 4. e5 Nfd7 5. f4 c5 6. Nf3 cxd4 7. Nb5 Bb4+ 8. Bd2 Bc5 9. b4 Be7 10. Nbxd4 Nc6 11. c3 a5 12. b5 Nxd4 13. cxd4 Nb6 14. a4 Nc4 15. Bd3 Nxd2 16. Kxd2 Bd7 17. Ke3 b6 18. g4 h5 19. Qg1 hxg4 20. Qxg4 Bf8 21. h4 Qe7 22. Rhc1 g6 23. Rc2 Kd8 24. Rac1 Qe8 25. Rc7 Rc8 26. Rxc8+ Bxc8 27. Rc6 Bb7 28. Rc2 Kd7 29. Ng5 Be7 30. Bxg6 Bxg5 31. Qxg5 fxg6 32. f5 Rg8 33. Qh6 Qf7 34. f6 Kd8 35. Kd2 Kd7 36. Rc1 Kd8 37. Qe3 Qf8 38. Qc3 Qb4 39. Qxb4 axb4 40. Rg1 b3 41. Kc3 Bc8 42. Kxb3 Bd7 43. Kb4 Be8 44. Ra1 Kc7 45. a5 Bd7 46. axb6+ Kxb6 47. Ra6+ Kb7 48. Kc5 Rd8 49. Ra2 Rc8+ 50. Kd6 Be8 51. Ke7 g5 52. hxg5
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
If you realllly enjoy my content, you're welcome to support me and my channel with a small donation via PayPal or Crypto.
Link to PayPal donation www.paypal.me/...
Maiar Wallet @agadmator or get.maiar.com/...
BTC address bc1qckd3ut0hqyymzv33eus97ld8klj02xhk2kcwld
BCH address qzmfclyn69hqhjslls40r7r0dsttwe3tcsl946w4fr
LTC address Laarf1RmvCpLt2BcSwC1PBLG3hRC4HjBrz
NANO nano_1h1kgfaq88t1btwadqzx73rbha5hwbb88sxmfns851kwj8hnosdj51w388xx
Monero 4AdvvqmC4xhPyyRSAEDxTTAoXdxAtX2Py6b8Eh4EQzBLGbgo5rY5Khcap1x76JrDJH87yibAE9b6TPwTsvBAiFFCLtM8Be7
For any other currency address, contact me via agadmator@gmail.com
Check out ALL my videos here • "Grand Opening" - Ande...
Facebook: / agadmatoryoutube
Twitter: / agadmator
Instagram: / agadmator
Lichess: lichess.org/@/...
Chess.com: agadmator
Skype: agadmator
League of Legends: agadmator :) "Watch me without ads on your Amazon devices (bit.ly/Agadmat...) and Roku TV (bit.ly/Agadmat...)
no one can get over saying "and stockfish resigned the game" its so pleasing
This guy gets it :D
Alphazero was *trained* on monster hardware, it didn't play on monster hardware.
stockfish evaluated 70 million positions a second vs alphazero's 80 thousand positions a second. Stockfish had the hardware advantage.
Aniket Bramhankar tbh Stockfish never resigns the Game its in game setting to resign after terrible position
RaniaIsAwesome do you even know how a neural network works?
"the barbaric RC7+" had me cackling, i think barbaric is truly the best way to describe that move.
5:00 „... but after king captures .. yyy phaa” - agadmator 2018
Dude, through your chess, you have made my life better. Thank you from Scotland, and my wee dog
Why stockfish had to play that last dumb move in all of its games? It could have graciously resigned before that ;)
Bharath Kumar
Probably has to get to a predefined certainty of losing before it can resign. As any move is losing, it just took the least worst, or a random choice, and then the next position takes the evaluation over that threshold
Because as we have seen, its a flawed ai
Engines can't resign
great work always! thanks for the channel
I love this repeating exchange:
Stockfish: DRAW!!!
Alpha Z: lol, nope.
/ followed by some clear, ingenious and spectacular Stockfish destruction
Keep uploading alpha zero’s games!atleast once a week please
Muskov OpenAI je duze trenirao sa sobom, i posle toga su neki profesionalni igraci uspijevali da ga dobiju cak oni neaktivni. Tako da je Alpha mnogo iznad OpenAI-a. btw odlicno radis samo nastavi tako.
sib that is because the characters in dota needs.time to finish the action(move, attack, cast spels)
In chess that.happens in.instace
I really like to see this type of games :)
6.. cxd4 has been known to be a weak line for black for a long time. Just goes to show how even a top engine isn't the same without an opening book.
4:59 "ah ieiein phaaah"
No no , you can go on with alpha zero games ...
Kxd2 !! Wow.
Love your comment "to take is a mistake" ... good advice! ;)
alphazero is the legend
The OpenAI five's first game against the 'top' Dota players lasted ~13 minutes, with one player all-chatting "we need John Connor" before the game ended. I think the fastest pub game I've had lasted 17 minutes, which certainly felt dominating at the time.
I play a lot of Chen and prefer playing with cores who come online very fast, so it was very nice to see the OpenAI five formally refute the "sacrifice the development of position 2-5 so that position 1 can farm for 25 minutes and eventually win the game by themselves" style of play.
That was another thing that the OpenAI five did - it never let any of its heroes get very far behind in levels. Just made "forcing moves" from level one that involved pushing fast and hard with cheap heals and 4-5 heroes at a time (usually 5).
I see these moves in AlphaZero games as well. Stockfish had a bishop here on paper - but only on paper. All of the forcing moves that AlphaZero had made up until the endgame basically made black's bishop an irrelevant piece because it never got the opportunity to develop.
"Your opponent can't win the game at move 20 if you've already effectively won at move 10" (points to forehead)
This game probably lasted like 0.1 second aswell haha
OpenAI igra samo kracu verziju dote, 1v1 sa samo jednim od mogucih 95+ heroja. Trebati ce im jos puno vremena da bi mogli ukomponirati interakcije izmedu razlicitih heroja. Tako reci, u doti nisu odradili niti 1% ukupne igre, a u šahu metu sve pred sobom 😄😄
Alpha zero man, with these positional piece sacrifices. Don't try this at home kids.
"How does one crush the French Defense?"
Step 1: Invade Belgium
Step 2 go around the maginot line
step 3 push the frenchmen to the beaches
To be fair that only worked half the time. The first time they went through Belgium it actually cost them dearly. By invading them it made Great Britain get involved. I know it is fun to make fun the quick surrender in ww2 but that was leadership at fault. Some people wrongly assume that the French as a whole were weak and soft....this is not the case. Mostly that is just from how people view the elitists of Paris. Comparing them would be like comparing our US soldiers to the wealthy socialites in DC/NY
@@mattroxursoul We lost because the Germans followed De Gaulle's tactic and the French didn't 😭
@@matteocavalli5918 This is double fun comment because orginal 2 pons werent captured. They stayed as A0 went around and crushed everything
Agadmator, please never limit yourself by saying 'no more alphazero' or whatever. Just show us the the best games, that's why we are here and love your channel
+John Lee That was a joke :)
This game from Alpha Zero is truly beautiful, I've watched it 3 or 4 times already, I can't get over the ke3 and the bxe6 then pushing the pawn, this looks like something Fischer or Tal would play. Alpha Zero really feels as if it played with a certain touch of instinct.
Thank you for preparing this game for us.
In fact we are so happy to watch these games .It feels this is the world championship material not the poor thin gruel doled out the by the recent Human championship.
I'm pretty sure his name is not agadmator but Antonio
@@nwikendulue4824 your comment is so dumb, i hope you joking, and even if you are is a lame ass joke
Alpha *Zero* : “If the king doesn't lead, how can he expect his subordinates to follow?"
Code geass
Nice
Code Geass
Lelouch
I am honestly surprised people who saw that show is so common in 2020.
So if I wanna be the best chess player in the world, I just play against myself for 4 hours. Got it.
no, if you are human, i would take ~1500years to play against yourself
@MTRredux because it can think approx. 100 moves in 10^(-1) sec or more
If you can play over 10 million games in 4 hours. Yes!
Some people go crazy playing against themselves.
bluewaters4567 other people go crazy playing with themselves
There isn't a snowball's chance in hell I would have captured the knight with my king, then move to e3. Wow.
gillyzxcvbnm How did the snowball make it to hell? Must have been a bad boy.
No one knows, very low chances.
I could see myself playing that. I played something similar before in an otb game with a similarly closed centre, I think it was a decent move aswell if I recall, was a long time ago. Especially seeeing as there was no other satifactory defence to the d4 pawn, Nxd2 Bb4 Qc1 Qb6 Qb2 Bxd2 and you would have to got Kxd2 anyway.
Carlos Silva if you played it it would’ve been a blunder ;)
Fischer and Tal played such moved tbh
You like the Alpha Zero admit it:)
Who wouldn't like it :D
Good to see long term abstract thought defeat brute force calculations.
RaniaIsAwesome
hahahahaha no, that's not how it works. They don't become the most popular video network in the world by giving unfair treatment and favoritism for ranking
"Everyone knows" is not a persuasive argument. You do realize all those recommended videos are based on what you like and what's trending, right? So how do you do an experiment which successfully separates your viewing habits, global trends, and UA-cam's algorithms from each other? Let's see some evidence, instead of just talk, that the AlphaGo recommendations are due to UA-cam explicitly prioritizing AlphaGo instead of simply because they're trending organically. Why the fuck would they even *want* these videos to be served to anyone other than those already interested in the subject? That would only hurt their public image since so much of the public is extremely distrustful of AI.
What do you think it says about your argument when you can dispense copious insults and zero evidence?
Edit: Did you see this in "Trending on UA-cam" or a different feed like "Recommended for you"?
dude, you are a fucking mind reader.
also: i enjoy those games soo much, mainly because you enjoy them so much and transport this amazement over, what about once a week alpha zero games? it will still grow and it would be nice to follow it's journey.
They only published 10 games :)
agadmator's Chess Channel please show every game of it, or at least ask the community 🙏🙏😂
Santiago Morales hahaha. 18+ for graphic violence and full frontal nudity.
Incredible to see rows 1 and 2 empty in the midgame and Stockfish crushed against the top of the board like roadkill under a truck tire.
Of course, i knew what you were gonna say the second i saw the title😂😂😂
:D
agadmator's Chess Channel You always reply man, thank you.💪🏾❤️
I'm fed up with boring human chess
Keep making alpha zero vids please
If you check the paper AlphaZero learns the French defense before 1 hour of training and by the 2 hours mark it's 12% of his played games. But the percentage suddenly drops and, after a time, AlphaZero abandons the idea. Seems like he learned from his mistakes lol
Marcos Goetten where?
Wait, alphazero plays shogi too? I want to see these games. I know very little about the game, but it's be nice to see what the highest level of play is for shogi
AlphaZero: "It was just a phase!"
1:19 - When he said "And in this position..." my mind expected the next words to be "Stockfish resigned".
I like these videos about Alpha Zero. If you went over every single unique game it played, I wouldn't mind at all. In fact, I'd prefer it.
Ashley: stockfish, what do you say about your game against alpha zero? It wasn't very smooth
stockfish: i mean, what you want me to do?
Carlsen actually won that game no?
Ahhahaha
Agadmator : Alpha is already controlling more squares in six rank and this is very nice.
Alpha : Ohh thankyou Agadmator (with shy face) ^-^
Even Alpha Zero recognizes the greatness of Agadmator
“Now the White King is infiltrating the position”. Sounds like another Mission Impossible for me
Black was up a Bishop but white was up a King ;)
Elon musks bots defeating dota players is insane considering how its much much more complex than chess
"Even the barbaric Rc7+"
I know Go was a milestone AI advancement, but Alpha Zero beating our best brute force heuristic search is a bigger milestone for neural networks imo. Indeed, its like humanity has discovered a new "fire" with neural networks, with the algorithms generalizing across such a wide range of domains, as well as able to learn via first principles with no human intervention at all.
These recurrent neural networks are demonstrating intuition, forward planning and deep strategy now which is incredible. But we should be concerned about the implications, as unlike stockfish (which is bound to the domain of chess), Alpha Zero can be applied to literally any domain. It's naive to think that this technology won't impact us all in unforeseen ways, and that those that control and implement this technology will have immense leverage over humanity (for better or worse).
I can't help but sense that things may get very bad (or at least very different) from here on out. I've written very small neural networks myself, and even in the most basic implementations, its easy to see how mappable this algorithm is (and scary to think about the neural network algorithms deployed at scale). I have seen what they are capable of and can appreciate the change in thinking with regards to state based heuristic search to neural networks. We are in for a bumpy ride....
An yes, the application of neural networks I put up there with the discovery of creating fire, writing and literacy, agriculture and farming, taming electricity, and the introduction of automation in industry. Given machines provide the fabric for modern civilization, intelligent automation is a BIG BIG BIG DEAL. Be afraid.
BinaryReader Did you have Terminator in mind.
When I took up automation, I knew for a fact that automation will be the center of our society in the coming future. However, I did not predict the advent of neural networks effictively nullifying it.
In the future AI will look at how humanity treat it's own planet, calculate our possibilities of fixing it, and when they see "0" they will conclude, that we humans should seize to exist, and they might destroy us.
Though such scenario might sound scary, I also see it as the human genome "natural" evolution. We went from fish to Apes and from Apes to Humans - we will go from Humans to Androids and from Androids to machines.
That's the only way for our intelligence to survive the hostility of it's environment and spread into the galaxy.
If I was to bet, I would bet that 99/100 alien races out there, if they exist - they are only going to be machines.
Add quantum computing power to neural networks, and it's several orders of magnitude in advancement. The human brain is rapidly becoming obsolete. We will have humans-only chess tourney's while the computers laugh at us!
@Minkki82
You should check out recent news on the games between ΑZero and Stockfish.
deepmind.com/blog/alphazero-shedding-new-light-grand-games-chess-shogi-and-go/
I just can't wait until they train Alpha Zero for an entire 2 weeks, to see if it can reach ELO 10,000 That sounds like a joke comment. But they trained Alpha GO Zero, for 36 hrs to get to human master level, and then let it train for 40 days, in order to really become a GOD. After that it plateaued. Is GO really so much more sophisticated than Chess that it only takes 4 hrs for AI to master to GOD level??
HypersonicMonkeyBrains No. The simple answer is that humans are better at Go than they are at chess.
Playing a game of Go is mostly an additive process, where a chess position can look drastically different just a few moves down the line. So Go is easier to visualize for humans, and hence we're naturally better at it. Chess engines have been beating human masters for 20 years now, but for Go it's only a recent thing.
I play both chess and go. In my opinion, chess is a better game all in all but the go is so deep in the advance level. There are just so many moves possible. Imagine a chess game where there are always more than 10 best moves, and your opponent also has 10 best moves against you.. The complexity is unimaginable..
2 years later I’m still in awe over this game. Absolutely astonishing! 😁😁😁😁😁
Same
2:28 Dog : I want to sleep but that guy is always talking , let's go get some food
Dota 2 is also an extremely complex game, so Elon Musk's AI winning against professional Dota 2 players is an outstanding achievement indeed.
I used to play a lot of DotA2 and I have seen the live coverage of the event, so for those of you who are interested here's some info: dota is a lot similar to chess in a sense that you have a draft (like the opening) early and mid game (midgame in chess) and lategame where a single mistake means loss (usually). The difference from chess is that you also need to be good with reflexes and "mechanically skilled" to perform combos and fireworks. Best thing about dota is that its like chess but for teams: every player of the team really needs to understand the game strategy, coordinate with your mates and stay cool headed.
Now, Musks AI defeated pros handily, but only in a one vs one scenario. To continue the analogy to chess this is like playing queen and king vs queen and king. Which is nowhere near real chess. Although even that victory was surprising because dota bots were always horrible.
The real challenge for Musks OpenAI is going to be to teach bots how to coordinate and create strategies.
I am both excited and horrified of this AI mania
You fear that the AI will have problems creating strategy? haven't you just watched the video?^^
I mean coordination might be weird but you could just let 1 System play multiple players...
the thing is, there are 46 897 636 623 981 (4 * 10^13) combinations of heroes in dota, and a computer has beaten a man in only one of those, and in a highly specific scenario. Heroes can also be played in many more combinations, the combination of those is also very characteristic. If it were a strict set of heroes (say, 5 chosen for both sides) then yeah, I'd say its possible with neural network algorightms which machine can repeat many times and learn from reinforcements, but if I were to lay a bet, I would say computer won't beat a team of 5 players any time soon, and DotA probably won't even be relevant by the time it will
@@astroNexx This reply didn't age well...
I have feeling that alpha zero is Bobby Fischer.
That's probably insulting to A0
More like Tal
I mean, Stickfish would beat Fischer pretty easily.
Alphazero eats Fischer for breakfast and defecates Carlsen.
Alpha Zero is the Mage of Riga.
I hope AI shows that the French defense is losing by force I hate it!
looking at the alpha zero paper, it looks like black wins or draws, this seems to be one of the rare cases where stockfish did not draw the game as black
Even if it can, no human can play like stockfish, let alone like A0
And btw learning the French is not only good brcause its a good defence but because you learn key chess concepts that can be implemented in other openings
@@ZapOKill The other ones are saying that after first 2 hours of playing against itself alpha zero abandons French defense and never plays it again.
I have a joke.
“French Defense. “
I have a better joke.
“English Opening.”
Thanks Agad😊✌️
Dota 2! You gotta see those pros beasts in fact. But stockfish isn’t a small one either
Sabareesh Sundaravel nah. It's amazing that Musk's AI can 1v1 the pros, but as soon as we see them 5v5, itll be the best AI for sure
I kind of gave up on dota 2 after patch 7. They messed up the levelling system.
Can't the Dota AI only play 1v1 SF mid with no items? I'm hardly impressed by that.
1v1 SF with special rules (no runes, no soul ring), but it is till very impressive to see how he crushed so many pro players
Limitations were quite serious. No bottle abuse, no mangos, no wand... But hey, it was only learning for a week or so :)
It's amazing how stockfish consider the position a draw on 6:04 and didn't find the mating moves that alpha zero found
You made another one!!! ❤️
I want to see a DeepMind vs DeepMind game. I just wonder how complex it gets. Because stockfish looks like a complete beginner against AlphaZero. AlphaZero seems to understand key concepts at such a masterful level that Stockfish doesn't even begin to grasp.
I also want to see a human against AlphaZero. I wonder if it's easier to beat a machine that plays like a human.
you realize that humans have no chance whatsoever against stockfish, right? The only cool thing to see is how much faster DeepMind manages to beat humans compared to stockfish...
I *don't* want to see an AlphaZero vs. AlphaZero series of games, I'm afraid it'll solve chess and we would watch it infinitely draw itself.
That could be great - or not. I kinda doubt if we understood what the hack is going on on that board. Like - sure, somebody could say "this is for this" or "later this is used in this way", but where is the certainty? AZ could do that move with completely different intention or for completely different reason which would get lost in following game going into another direction.
As long as commenters understand at least one side of the game and know what it is doing, it is easier to understand what AZ is (probably) doing. But AZ vs AZ? How would you evaluate its moves?
Petr Mazak it could very well be that the AI can explain the purpose of its moves
A0 demonstrated what they set out to do. A0 can learn to play chess above human level all alone - mission accomplished. Giving SF a few extra ply wont help it. It sees completely lost positions as draws?
60K Subscribers by the end of this weekend, easy. Congrats in advance Agadmator!
+BMessemer It's already more than 60k! :D
did stockfish win any of the public games?
Geckomaniac 3 Stockfish didn't win any of the 100 games played.
Stockfish sadly did not win any of the 100 games they played but it did draw 72 of them.
This channel is hot... will donate soon!
Please mr agadmator, we want more alpha games on your channel. Without your analysis, i don't get full contentment.
Elon musks Ai beating Dota 2 pros is more developed Ai than Alpha Zero.
vs French Defense - Remember: Time and space when an opponent plays a solid defense ask yourself, 'What would Morphy do?'.
We want more alpha zero. This AI cpu is like Fischer, Morphy, and Tal morphed into one with a little Capablanca intertwined.
Love this comment.
Show a game where stockfish won against Alphazero.
You mentioned that 75% of your viewers were non-subscribers in an earlier video (I'm one of them). I'm interested in AI that uses neural network learning. I don't really know how to play Chess (I know how the pieces move, but I'm not really sure what a French defense or other Chess jargon means). I do play Dota though, and I watched the OpenAI matches. There were a lot of constraints placed upon the match (the match was 1v1, the player and the bot could only play one particular hero, certain items were not allowed, and certain resources were not allowed). Despite the limitations, the bot was very interesting to watch. Some of the things the bot knew is easily programmed in any bot that would ordinarily give players a hard time (the bots know the exact point towers will start attacking and always stay out of that range when harassing and never misses a last hit). Human players don't have that type of precision. What the "normal" bots can't do is fakeout a player (or trick a player into thinking the bot is going to attack or cast a spell). Spell cancelling is a thing in Dota, but you never see bots do it, so when OpenAI faked a player out by cancelling a spell, the whole audience let out a gasp. I don't even know how OpenAI learned to do that or why. I guess human players also learned it, so maybe it's not too surprising. One thing I didn't really get a chance to see if OpenAI learned is juggling tower aggro. OpenAI was winning too hard for the player to come near for the opportunity to arise (though it occurs all the time in regular games).
With all that said, I think OpenAI is too limited right now, and since the initial match, Dota players have beaten OpenAI several times (albeit, it is probably not the version of OpenAI that is learning). OpenAI will supposedly come back to play Dota, but this year with a full team of 5 OpenAI bots against 5 Dota players. That will be really interesting to see.
I am really liking AlphaZero. I'm learning a lot about Chess from these videos. I don't know if it's normal for Bishops to be trapped with no moves to make, but I was really impressed by seeing that! Thanks for making these videos!
Only 4 analysis left and we got the 10 🙏🙏😂
I love it when anyone brings their king to the other side of the board, something satisfying about that
kept thinking my notifications were going off but it was just yours lol,
Sorry bout that
Dota 2 player "Black" beat Ellon Musk's AI numerous times. Check it out by searching "black beat Dota 2 AI"
Hi, I watched your channel couple of days and really enjoyed it, most 2 alpha zero game, this one and another when stockfish had same situation for bishop, when I watch the alpha zero I feel it played really like very smart person, it just fight for wining, it hate draw, it want to win even it loss all of its pieces.
"The star move of the game Ke3" it's official... Alpha Zero is Lelouch V Britannia
If 22. Bc3, Ne3 forces the exchange of the white square bishop and messes up white's kingside -_-
I can't get enough of these deepmind games, they are so amazing
The Germans dismantled the French defense long ago
OpenAI is a company with a lot of different investors, Musk is just one of them :D. And OpenAIBot defeated lots of ESports players of Dota, but eventually i think 5 (not sure) people managed to defeat it. +its not that new compared to this event, 6th december :D.
this video is soo satisfying coz i hate the french
I thought the last video was your last video about this hahahaha ;)
That is how an addiction always starts, more more MORE!!!
If you sacrifice a bishop that early in the game you must play perfectly till the end to maintain the positional advantage, one wrong move and you lose. This is why humans can't replicate engine like sacrifices.
You said it before I did
in theory, crushing dota 2 players is a WAAAAAAAAAAAY bigger accomplishment (the complexity is exponentially higher from an ai's perspective), but the games weren't entirely fair since a robot doesn't need screen updates and it had full map knowledge. also they weren't team based games which most dota 2 is on a professional level
4:17 The light square bishop is utterly effed, and trapping other pieces, until a significant change.
Agadmator, the optimist - "The light square bishop isn't looking good here."
Genuinely scary. It got 4 hours with only the rules.
4 HOURS
and it's not only beating stockfish, its wiping the floor with it? Jesus christ whats gunna happen with a day of self learning?
Interesting... Stockfish 14 now finds 30. Bxg6 after the neural net improvements introduced since v12. After looking at games, and specific positions like this, as well as considering the improvements in Stockfish since neural net play was merged into the engine, the accusations of unequal hardware being the reason for Alpha Zero's victory can be soundly rejected.
If "Elon Musk's artificial intelligence" actually defeated a professional _Dota 2_ team, that would be an incredibly impressive achievement. In actuality, it defeated several individual professionals in the 1v1 format in multiple games. 1v1 is not a common format, so the pros aren't practiced in it like they are in the game they play all day every day. But more importantly, it is extremely simplified: Only one hero is available (Shadow Fiend) out of 115, only one lane of three is used out of three, no runes spawn in the river, no neutral creeps spawn in the jungle, several items are banned, and most games last only 5-10 minutes.
That's _still_ impressive, but at this point, an AI that can actually outperform a professional human team in Dota 2 (as with most competitive video games) seems unlikely in the near future. I think the current target for Deep Mind is Starcraft, which would be a huge advance in AI technology already.
Not sure Chess & Dota2 are comparable. One is primary control over an avatar, the other is primary control over a number of avatars with specific abilities (bishops, queen, rooks, knights, etc.). One is more reactionary and the other is more position based. Regardless, I find Alphazero's demolition of Stockfish to be far more impressive. I think it's incredible and a bit scary.
Hi agadmator, I'm someone who's pretty into Dota 2 esports. Alpha Zero is by far the greater accomplishment. I'm not sure if you've heard the full story but basically the AI that defeated Dota 2 players had to do so in a very limited environment; they weren't really playing Dota 2. It would be similar to designing an AI to solve chess puzzles. It's not really playing chess.
Dota2 AI really shocks the wrold of Dota2 because it’s really hard to imagine an AI can play better humans.
chess is more like a calculation so it’s easy to for AI to crush humans like they said engines are 1000x better than humans if not 1000000x
Dota2 is all skills but those AI were deafeted by some dota2 players in youtube streamers but of course those AI is still underdeveloped and were not able to counter the unique strategy given by the youtubers
I have two points that are not clear in your explanations. First, at 6min38s. You showed f5. Why the answer cannot go for queen f7 and if white take the g6, the counter rg8 can be played and the attack cancelled. And if white take e6, queen can take e6 or even king can take e6 too because the bishop guards from re6 check then. I do not get why black didn't just strengthen defense with the queen at this point loosing two pawns for the bishop and it would be ok for both sides for me. Even, I think such defense would probably be favorable for black because pawn h4 becomes extremely weak and may enable the black rook to infiltrate and to destroy the defense whereas the position of the black bishop and of the black king just almost nullify the pressure of the black rook at this point.
Alpha Zero is better than Stockfish because Stockfish was created using human chess theories Alpha Zero no, and if we analayze all of alpha zero victory, in 90% he use zugzwang and Stockfish don’t know how to defeat it
I’m a relative beginner in chess but nothing about the way Stockfish plays inspires me. It seems to just play ‘boring’ or ‘safe’ moves whereas it seems AlphaZero isn’t afraid to lose material to gain a positional advantage but maybe this is because AlphaZero is able to restrict Stockfish from playing more attacking chess?
it doesn't really crush dota 2 players. It's just a 1vs1 mode with only 1 hero allowed and with the exclusion of a lot of items. It's a very very niece part of the game that the AI managed to "crush" players that eventually found a way to "bug" it and defeat it. To people that are unfamiliar with the game it might sound like something exceptional, but it isn't of much use as dota is a team game. It's interesting that it took the AI 3 weeks to be able to somewhat "crush" players in a very limited 1vs1 mode.
really late to the party, but since I play dota 2 and chess, I think beating dota players with AI is more impresive, since in chess you have 6 types of pieces, in dota you have more than 100 ( counting heros, structures, items, creeps, and neutrals ) also the "map" of chess is just 64 squares, and the "board" of dota is masive, and you can try to be in a position as people were before hand and is imposible. Also for dota you need 5 players and they have to coordinate to win vs professionals like they did.
@agadmator It didn't defeat professional Dota 2 players, but rather defeated 98 percentile players, some of whom were casters and ex-professional players. It also lost in the main stage of the International 2018 but put up an impressive showing none the less.
*DOTA AI beats professional players?!*
Meanwhile League of Legends AI are like kids first time on a keyboard
As somebody who has 3,000 in Dota 2, Elon Musks AI crushing these heroes of mine is much more impressive. A lot more variables to that game than Chess I would like to think.
Ok, take a break and comeback in a few weeks to do more A0 games, we need to learn lots from it. I think we are doing a mistake to distance ourselves from the machines when it comes to competing. Computers should be allowed to enter international chess championships, even if they destroy all the human masters they will force their levels to increase. Humans can still get the prizes according to who was closest to the top, computers will get recognition which engine is the best. We get better by training with stronger opponents, computers are the closest to an alien intelligence, we should be learning from it. If A0 was a human, would be quit studying his games because he is unbeatable? Lets not be like the kids who don't let the nerd kid play because we are afraid to lose. Just my opinion.
Stockfish may as well play every game (as black) without his light square bishop. You wouldn't even notice the difference
You said you hate the French Defense. I don't hate it per se. I just don't know how to play it. It feels so confining to me. Guess I like more open board. Anybody here had any success with the French Defense?
As a DOTA2 player I am 100% sure that DOTA2 AI is better giving the complex and difficulty features of the game.
"Alpha with the white pieces opens with d4. We have e6 by stockfish, standard stuff, but now the position is completely lost for black.."
"And in this position, Stock resigned. The Game." BLISS
Hi dude. I just wabted to clarify that OpenAI beating dota2 players happened months ago and it also won on a 1v1 match only on the first stage of the game with only 1 out of 110 heroes. It is an amazing acomplishment but it is nowhere near, not even close to being able to defeat dota2 players on a real game.
Alpha: Hey guys whats this game called?
Other engines: Chess, wanna try?
Alpha: okay give me 4 hours
dota 2 players is a much greater accomplishment - it relies on reactions and fake moves and taunting - quite impressive - although the player Black has now beaten it on most occasions - but in time it will learn from this and become unplayable. The dota test was a 1 v 1 mid battle. This year they think they will have an AI team of 5 players so you have to factor in map awareness and rotations around the map in the dark as well as team ganks and team fight synergy - looking forward to seeing how that plays out.