"AlphaZero, the game-playing AI created by Google sibling DeepMind...

>"AlphaZero, the game-playing AI created by Google sibling DeepMind, has beaten the world’s best chess-playing computer program, having taught itself how to play in under four hours."
>theguardian.com/technology/2017/dec/07/alphazero-google-deepmind-ai-beats-champion-program-teaching-itself-to-play-four-hours
this some cool shit fags, get in here

Other urls found in this thread:

lichess.org/study/wxrovYNH
en.chessbase.com/post/the-future-is-here-alphazero-learns-chess
arxiv.org/pdf/1712.01815.pdf
ratings.fide.com/card.phtml?event=401307
en.chessbase.com/post/alphago-vs-lee-sedol-history-in-the-making
blog.openai.com/dota-2/
youtube.com/watch?v=IKVFZ28ybQs
twitter.com/SFWRedditImages

now make it do something useful

it is cool. but i would like more human v ai tournaments and having the moves analyzed by humans. ai vs ai matches would be good but i think it is too complex

top humans don't have chance in chess these days even with handicaps

I thought Chess already was perfected

The key being it taught itself in 4 hours

this one is tricky when you have developed models and supercomputer...

That's passed off as if it's impressive. It sounds impressive to a human since a human can't master chess in 4 hours. Whatever massive machine google has to "teach" their neural network can do thousands of games concurrently, all completed within seconds.
I don't know if it actually is anything special when you think about it.

which is a shame because i enjoy competitions and mind games. and i value competitive integrity too much to want to have an ai that is purposely gimped outside of a show match. looking forward to the starcraft II match tho

Heard about this, pretty neat

It didn't lose a single game either, meaning it probably has ELO hundreds of points above stockfish. Impressive as fuck desu.

Chess computers that are gimped for showmatches are usually stripped of things like their endgame tables which realistically, no human would be able to memorize anyway.

Either way, the two really co-exist in separate realms. Chess computers are more of a tool rather than a realistic opponent. Sort of like how a basic table calculator would win every time vs a human doing mental arithmetic.

>thousand games concurrently in seconds
>not special
C'mon buddy... This shit is going to solve all million dollar math problems and solve al viruses problems humanity have. It'll inproove the human brain for sure as well. This is the beginning of a new era.

How can we even compete, bros

list of games
lichess.org/study/wxrovYNH

>It didn't lose a single game either
You didn't even read the article...
Pretending to be retarded, or a legit brainlet?

>"It took just four hours to learn the rules to chess before beating the world champion chess program, Stockfish 8, in a 100-game match up."
>"AlphaZero won or drew all 100 games, according to a non-peer-reviewed research paper published with Cornell University Library’s arXiv."

???

it played 44 million games on TPUs.

So I was watching some analysis, it's very interesting. They always do the same opening and Deep Mind always turns it into Stockfish having blocked pieces and unable to develop properly. Deep Mind turns flexible vs blocked position into small figure advantage and then win the endgame with this edge.

You could pick up any mid-tier engine out there, if it ran on a TPU it would crush Stockfish just the same.

This just simply isn’t anything impressive. They are just going to try to throw out bigger and bigger numbers hoping to impress people. That’s what “google” means after all.

It's pretty amazing, how deep learning, provided the problem is concisely put, basically beats everything humans can come up with.

I mean, if you think about it, traditional chess programs took decades of development to reach their current state. Now comes some really naive, primitive, fucking model and beats the fruits of all that work in hours. It's just incredible.

Seriously though, I don't think you guys understand what this is. Not only is it better than current chess programs, it also (a) took considerably less effort developing (b) is orders of magnitude less complex and (c) uses a tiny fraction of the resources of a chess program, in every aspect.

It's amazing.

the heuristics of chess engines doesn't need to be complicated, if you just went simple materialistic (counting pieces up) and invested resources on depth your engine could probably be at GM level.

>uses a tiny fraction of the resources of a chess program

lolwut, try and run AlphaZero on a laptop.

All you did was use a naive probabilistic model and feed it tons of data. It’s the definition of naive.

it's not a fair match up in processing power...this is nothing but google posturing

en.chessbase.com/post/the-future-is-here-alphazero-learns-chess

stockfish was running naked without opening database.

Its still intersting that alpha was analyzing ~80k positions/s while 70mln/s.

en.chessbase.com/post/the-future-is-here-alphazero-learns-chess

brainlet, it a proof that it can see implications on a different level.

>AlphaZero, the game-playing AI created by Google sibling DeepMind

Wait waitwaitwaitwait.

So AlphaZero is an AI, that was created by another AI, and is better than the parent AI in a specific task.

Has the singularity started?

It also moves the same piece multiple times in the opening instead of going for development, which is against the basic rules of chess but somehow it gets an advantage out of it.

arxiv.org/pdf/1712.01815.pdf
full paper

>non-peer-reviewed
TOP KEK. I'd rather take Trump's word.

Its a game of statistics. Moving the same peice over and over may be against our normative rules. It would make sense to develop multiple pieces so that you have options as the game progresses. But, if the program decides the best way to win is to push forward statistically, the stats dont lie.

what training data did it use?
if it just learned from playing itself that would be mind numbing

Deepmind is a company that makes neural networks and got bought by google.
They first made an AI that excelled at the chinese board game Go called AlphaGo, and then excelled again with a new iteration called AlphaZero that is even better than AlphaGo at the game Go.
That new AI "AlphaZero" is also now an expert at Chess.

...

How many games does a human chess player complete in their life? If a human starts at 5 and plays 1 game every other hour for 8 hours every day until they're 50 that makes ~65,700 games. I want to see that AI take on a human opponent with the handicap of only playing the same number of games as the human has.

>mfw i solved thousands of captchas so some program can beat another program at fucking chess

using 5k TPUs which is about equivalent to 75k-150k GPUs

so no shit it's going to be fast

I want to see an i9 Grid Ripper system with a Titan XP beat a human now.

It didn't use any training material

Obviously it wouldn't do as well. A human player has the advantage of having the game explained to them. Additionally, a human player can extrapolate about the effectiveness of different courses of play from the start. The AI systems we're talking about can't do that. They rely solely on throwing things at the wall and finding out what sticks.

> blitz games can last 5 minutes or less

Are you seriously trying that this thing trained itself in four hours? You do know that as it trains, it needs fewer and fewer cores to run, right? The original AlphaGo ran on 48 TPU and this new version shit stomped it 100 games to 0 using 4 TPU. So it may train on the whole network but only use a small portion of it to run. This means that these systems can be trained on a central server to perform tasks, then deployed to much less powerful clients to actually perform those tasks.

It actually is very impressive.... If you really are thinking.

You know what... The level of ignorance in this thread is incredible.

5k TPU is not even equal to 5k GPU let alone 75k. They're built on 28nm, run at 700mhz, and consume no more than 40W. It's far less processing power than you've invented in your head. I guess you're confused because it's an ASIC.

Some user once said
>programming is like telling a retard what to do
I would say there is no advantage because programming a computer is like teaching it and explaining things.

The programming in this case is teaching the computer to learn things. Since all the different tasks that a human performs are all very different, they would require different programming. Each nuance in similar tasks could require rewriting the software. Developing software that can train itself to do different tasks means that the programmers no longer care what the task is. They can focus on making a software that is very good at learning and apply that software to any task at all with minimal rewriting required. In the case of this chess game, they showed that they're getting to the point where they don't even need to rewrite the program at all; you can give it a task and it will eventually figure out how to do it on its own.

Very nice, I like it.

Okay, now let's see it teach itself to carry a game with 4 South Americans on its team in a Dota 2 pub match.

There's already an AI that stomped a bunch of pro mid players in a 1v1

Now that's pretty interesting, any link to a vid?

Just search Open AI Dota or something like that. There are a ton of videos. A lot of players tried to beat it in several matches. I think pajkatt was the only one to take a game-- the AI was so dominant it had never trained for an aggressive TP and it just didn't react at all. Kind of hilarious, really.

It's really no that interesting. The Dota2 AI doesn't have any reaction time, it's literally instantaneous and isn't worth any consideration.

Besides, training an AI to play a single hero in a faceoff against the same hero is nothing to write home about, it's really simple shit. Not get an AI, or multiple communicating AI's, to play as a team with enforced human-like reaction times and it'll be a whole different ball game.

how can I use it?
I want it to play shit like WoW runescape or hearthstone to grind away

I love hxh so much. One day autism will beat AlphaZero.

Stop making excessively boisterous conclusions about research you didn't play part in. Stop exaggerating potential outcomes of something you don't fully understand. If they're going to do it, they'll do it. There's a lot of complications that you're throwing under the rug by stressing "well it's possible". Sure, it's possible, but no one knows how likely and no one really knows what will come of anything.

We don't need you telling us what they may do and then getting upset when we call you out on being full of shit.

Yes, well, there's no start. But we're getting a bit terrifyingly close

call me when it can play magic.

Cool, let me know when it has the intelligence of a cat.

A C2D can easily beat the top human under standard time controls.

Everyone talks of the AI becoming "smarter" than a human, but it's always tested with finely tuned axioms in simple frameworks like chess or go. No one even knows what axioms a cat truly follows. One may suggest something along the lines of reproduction and survival as the closest things to axioms, but how do you train for that? How do you set the will to survive as an axiom? How do you "encourage" an AI to WANT to do something? Even if you're successful, what if there is a whole lot of potential existential anguish in a now willfull AI? What if they get so far and suddenly this wonderful machinery simply becomes depressed and lethargic at its own vacuity?

>Alpha zero: 4 TPUs (equivalent to around 720 8-core processors)

>Stockfish: 64 thread standard processor with 1gb hash

explains why alpha zero one

>playing against severely restricted OSS chess engine
>using the term 'superhuman'
>posting 10 cherrypicked games

jesus christ, AI fags are so desperate for hype
they are probably right though, normies just love to gulp that shit

you seem to be projecting user

by having bbc's that women will always crave

Dude a neural network is just a couple of matrix multiplications + some non-linear element that is easy to compute.
The hard part is training and running the optimizer not the playing the game part.

And stockfish can analyze the statistical effect of every move in the probabilities of winning and the advantage of each player.
Tell me what you will do with a bunch of weight matrixes and biases that isn't play games of chess.

>artificially cripple Stockfish by using an older version, giving it only 1GB RAM to work with and no book
>zomg it sucks

More Jewgle propaganda to be slurped up by the morons.

the ''''prodigy'''''s fide rating if anyone interested
ratings.fide.com/card.phtml?event=401307
spoiler, not even IM, let alone GM

if anything is impressive, it is the sheer number of operations modern hardware are capable of
>hurr, it was 4 hours
no shit, with more iron you can probably do it in even less time

Are you confused or something? All I said in that post is information that's freely available on the internet. The Deep Mind team spoke about it.

Machine learning isn't capable of cognition the way you're thinking. This software trains to do a task in a similar way that a human brain does, but that doesn't mean it has a the same structures or software that a human brain does. The only think this can do is learn to do a task. It will never ask why, it will never say no. Think of it as the part of your brain that knows how to play chess and learn to play better in complete isolation from the rest of your brain. The software here contains nothing else; a conscious computer will never result from this research because that's not the intention here.

That's not a fair comparison. A TPU is an ASIC designed specifically to run the libraries that Alpha Zero uses. The fairest comparison is power usage, since a watt is a watt no matter what you do with it. I'm that metric, a TPU uses 40W at most, so you're talking about 160W of compute against what? 200W on the conventional processor?

I don't think any of the Stockfish devs are even NMs, you don't need to be good at chess to program chess.

>win 50% with white
>win 6% with black
>no losses

tbf these numbers are very suspicious and look like to reflect some basic error in the setup and do not correspond to real world data. Even if you drink the kool aid and accept the 0 losses, white does not have such an advantage

Tell that to the Africans

>This shit is going to solve all million dollar math problems and solve al viruses problems humanity have
I think you're legitimately retarded.

I think what's interesting about the challenge is alpha-beta/brute force vs. monte-carlo/ML, not stockfish vs. alphazero. If we want to decide which method is better we have to give them the same hardware.

1. Stockfish has a wide range of contributors
2. No one called them prodigies
3. Assuming you don't just want to strawman(2) pointing out you don't need to be good at chess to program chess should be aimed at the guardian article which cites one of the authors as a chess prodigy

pick all three

kek, I remember watching the queen of katwe, they literally made a movie out of a back women learning the basics of chess

Still a decent rating.

The whole time they talk about her brilliance and genius. Turns out she's 1500 FIDE.

Anyways, this was a really shitty matchup. They gimped stockfish just to build hype. Call me when it beats it with an access to an opening book and more than (((1GB))) hash.

>lichess.org/study/wxrovYNH
Game 4, AlphaZero v Stockfish
1.d4 e6
2.Nc3
Okay, I'm calling bullshit right here. There is no way that 2.Nc3 is going to be played here in a top competition game. Playing 2.e4, transposing to a French, is a natural (and good) move. And playing 2.d4 is an obvious invitation to transpose into a Nimzo or QID. But seeing 2.Nc3 suddenly pop up invalidates the whole experiment for me.

I didn't read the article on the guardian, read other sources about the subject. You're right he's not a prodigy.

it is only decent if you compare him to amateurs, certainly not in competitive chess. And nowhere near prodigy tier.

I recall the same discussion regarding some dota bot, which was later shown to be using some tricks to beat pro players. Nothing to see, post it on facebook and collect likes

No, I mean he's nothing compared to real pros but he's still probably in the top 0.1% of rated players. Not a prodigy, but a weak master.

...

his profile clearly says World (all players): 14381
it is up to taste if you think that's anything to brag about

Huh, that's surprising, I wonder why the median is so unreasonably high. I'm guessing that there's some sort of selection bias where only stronger players play in fide tournaments as opposed to uscf/dwz/bcf etc. Iirc uscf is quite a bit lower, with median somewhere around 1200, which is like 1100 Elo.

I consider 2250 respectable for someone who is obviously a casual player nowadays. Supposedly he was world #2 U14 behind Judit Polgar, which is a decent indicator of potential. A lot of people consider her a prodigy.

Are you fucking retarded? It's a standard opening, exactly what a computer would play. Computers don't give a shit about the current metagame in chess.

>Hassabis, a former chess prodigy in his own right, rated no.2 in the world in the under-14 category just 35 Elo behind Judit Polgar
en.chessbase.com/post/alphago-vs-lee-sedol-history-in-the-making

You sound salty.

>shit stomped it 100 games to 0
Alpha zero won 28 games and drew 72

He's talking about improvements in the Go project, not chess.

develop the Dota 2 and Starcraft AIs already

t. not a chess player

Seriously, 2300 is something only a small minority of chess players ever achieve.

Also white to move, mate in 2

google's new captcha
>white to move

there's literally only one move for white, how is this even a puzzle

My sentiment exactly, you have a sequence of not just forced but literally only legal moves.

Here's an actual puzzle and a pretty one at that. White to move and win.

where is this prob from

Dota 2 AI already exists.

blog.openai.com/dota-2/

actually it is the very opposite, I'm not making a living out of chess but played with plenty of 2300+ IMs and GMs. Honestly, it is more about dedication than talent.

youtube.com/watch?v=IKVFZ28ybQs

Black knight from A5 to A6.

Bullseye

It's a composition, Lazard 1928.

Nope.

That doesn't diminish his ability to walk into most chess clubs and beat everyone. Just because he isn't IM+ doesn't mean that it's not a decent (ie way above average) level of skill.

...

Are you sure? I don't see anything that can counter the next unavoidable move, knight from A6 to G6 with the exception of moving the queen. That in turn leave the Queen vulnerable to the pawn at H5 or the knight depending if you move the queen before or after. I don't think too far ahead in chess because I suck at it but this effectively leaves white's king out in the open with no queen or bishop.