>Can he win 1 game edition
https://gaming.youtube.com/watch?v=qUAmTYHEyM8
1 hour till game starts, should be some pregame commentary starting 15 minutes before.
Game 4 of 5
Current score 3-0 AlphaGo
Official Stream
https://gaming.youtube.com/watch?v=qUAmTYHEyM8
More advanced commentary (Starts 1 hour after official stream)
https://www.youtube.com/c/usgoweb/live
----
Basic rules and introduction to the game:
http://senseis.xmp.net/?RulesOfGoIntroductory
----
Game summaries:
https://www.youtube.com/watch?v=bIQxOsRAXCo
https://www.youtube.com/watch?v=1aMt7ulL6EI
https://www.youtube.com/watch?v=6hROM_bxZ9E
----
Analysis:
https://gogameguru.com/alphago-defeats-lee-sedol-game-1/
https://gogameguru.com/alphago-races-ahead-2-0-lee-sedol/
https://gogameguru.com/alphago-shows-true-strength-3rd-victory-lee-sedol/
Time used in game 3
I know that Mike is 9dan, but how is he with katana?
>>53457584
ha that screenshot is more or less the exact time i went to bed, did it get more interesting after that?
>>53457651
Quadruple black belt
put the start time EST you mong
>>53457657
Yep. Once Lee realized he lost, he went all in and tried to make complicated kos to trick the AI.
It responded absolutely flawlessly.
>>53457584
So now that there is nothing to lose, will Lee Sedol play a crazy opening or will he play normal?
>>53457626
i like that linear slope at the end, you just know that the AI has won at that point.
>>53457682
AlphaGo
lphaGo
phaGo
haGo
aGo
Go
o
>>53457709
Great wall hype
post michael
w-wait you forgot all the stuff actually related to deepmind
CEO of Deepmind explaining how AlphaGo was trained in a pretty good talk: https://youtu.be/4fjmnOQuqao?t=2191
Deepmind's Nature Paper on AlphaGo: http://airesearch.com/wp-content/uploads/2016/01/deepmind-mastering-go.pdf
Deepmind Press Release for the Challenge Match: https://storage.googleapis.com/deepmind-data/assets/media/GDM%20Press%20Release%20Google%20DeepMind%20Challenge%20Match.pdf
https://twitter.com/Kasparov63/status/707532443541225472
>>53457657
Probably the most exciting of the three matches so far.
>>53457584
Pasting some info from >>>/sci/ thread
Streaming @:
>Official stream
https://youtu.be/yCALyQRN3hw
>American Go Association (AGA) stream
https://www.youtube.com/c/usgoweb/live
>Japanese stream
http://live.nicovideo.jp/watch/lv255280483
https://youtu.be/aZtZdAaInEM
Relayed live:
https://online-go.com/tournaments
>>53457626
>that decline
That was when AlphaGo declared the lower group dead, right?
>>53457626
Do you have the graphs for the other games where alphago was doing worse on time?
>>53457709
>>53457731
I'd feel really uncomfortable opening Tengen as white.
>>53458026
Oh right. Though he was black today for some reason.
Maybe AlphaGo will play Tengen
>>53458010
This one has remaining time for all 3 games
>>53458071
That will happen when they train it without teaching it on human games first.
stream is live; game starts in 30 minutes
get in here and spend the next 4 hours watching history be made even though alphago won the match anyway
>>53458071
Are there any more alphago-tans?
>>53458100
That's actually really consistent. I suppose Lee was just playing quicker in the first game, not alphago playing slower.
>>53458157
>>53458157
>>53458157
Only one I found
>>53457718
>>53458167
why is she so cute bros
>>53458100
It is quite nice to see how consistent it is.
Deepmind sure has good taste in music
>>53458222
2comfy
who is this alphago guy, some kind of /g/ wizard?
>>53458276
>guy
Can hard work beat talent?
>>53458290
The ol' talent bait. Classic my friend!
>>53458290
yes, alphago went through thousands upon thousand of pro games and played itself for what must have felt like eternity to get where it is
>>53458290
Yes.
Alphago has trained more than any human ever has.
Literally hundreds of millions of games.
>>53458276
best waifu of 2016
Pregame interview with deepmind engineer on now
https://gogameguru.com/alphago-shows-true-strength-3rd-victory-lee-sedol/
AlphaGo truly best grill
>>53458345
>springboard
>watershed
>reinforcement learning
>self play
>value function
>>53458375
>self play
L-lewd. Alphago-chan would never do something so depraved.
>>53458372
>and women
>>53458375
I know some of those words.jpg
>>53458417
>/r
>>53458413
>>53458432
>>53458395
i'm pretty sure it played against modified versions of itself like millions of times a day though
http://strawpoll.me/7067357
>>53458417
retard
>>53458432
ONE MINUTE
>>53458464
>>53458499
topkek
>>53458477
I know, I'm just joking that self play sounds like a euphemism for masturbation.
>>53458375
Hopefully alphago will convince people to start doing RL again. RL basically hasn't moved for the past 20 years, despite everything in ML getting deprecated every 6 months.
>>53458290
hard work beats talent, when talent doesn't work hard.
So, I heard AlphaGo uses around 1000 CPUs and 100 GPUs to continuously function as a self-learning AI. But if we remove the learning component (which seems to be the heaviest part of it IMO), is a static but mature model computationally inexpensive enough for implementation in, say, vidya?
>plays P4
It's over
>>53457584
Is there a website/video that explains the plays for absolute retards?
>>53458290
alphago is not talent
it's basically hard work
it spent a shit load of time playing against itself
>>53458586
>lee answers with D16
HOLY SHIT LEE IS FINISHED
Resign
>>53458629
alphago has a fucking huge lead
lee should resign
>>53458594
AlphaGo is more like Talent with hard work, it was made to play Go, meaning that it was constructed with the innate talent to play it but it also had to work hard to get it.
What about that more general deepmind AI that can be trained to play any 2D video game? If that were used on go how strong would it be?
>lee just does the same as alphago
lmao
RELAY HYPE
https://online-go.com/review/116862
>>53458582
no it also does tree search which is hard as diamonds
IT'S HAPPENING AGAIN
What will happen eventually if you keep mirroring AlphaGo's placements?
>>53458629
Lee's getting annihilated omg.
Is the QT go player gonna stream?
>>53458670
Do you want skynet to happen?
HE'S REPLAYING GAME 2
>>53458582
>IMO
lel
A trained model's performance depends on the model itself. A CNN can be very fast (enough for realtime on mobile in common cases). Alphago has to continually unravel the searchspace during play, though, and the cost for that will always be large and will almost always dominate the runtime. The fact that google uses so many CPUs suggests that this part of the process is not easy to parallelize in the first place, so it's unlikely for the next 5 or so years that it would be viable for consumers (as any computationally relevant ML would happen on the GPUs).
Even with the current hardware, it takes about 2 minutes per move for alphago, so it couldn't be used for realtime systems where the searchspace is similar to go's anyway.
how do you even play go
>>53458670
alphago decides to play white
you both get the same amount of territory from mirroring moves for the whole game
alphago wins because of the 7.5 komi thing
CALLED IT
reminder alphago is frozen
that means alphago WILL DO THE SAME MOVES if sedol wants to
Lee abandoned yesterday's strategy of playing defensively. This is good, we might see a good game today. Yesterday's was just pathetic.
I told you niggas AlphaGo would repeat the same moves.
>>53458736
Not necessarily, because of MCTS there is a small chance it could find a move it likes better.
>>53458736
nope
there is an element of randomness to the decision making process
>>53458413
> dat bulge
What did his dick mean by that?
>>53458670
>>53458744
Yeah, didn't the developers said the "froze" AlphaGo learning algorithm?
>>53458736
>reminder alphago is frozen
Does this means it's not learning from the games agains Lee Sedol?
>>53458725
-Connectivity is horizontal or vertical only.
-Groups of connected stones completely surrounded by the enemy are captured and removed from the board.
-1 captured stone = 1 point
-1 point of territory = 1 point
-White gets 7.5 extra points for going last
-You cannot make a move that repeats the immediate previous position on the board
-The game ends when no move can be made that improves a player's score
-The winner is whoever has the most points.
did he just say THICKNESS?
>>53458736
Not at all.
I bet the guy could find ways to break the ai if it could rewind and test different moves to see how it responds.
Being uneducated on the subject but enjoying breaking game ais by doing this, that's my opinion anyways.
Look at that puny human with his petty tricks.
>>53458781
>Does this means it's not learning from the games agains Lee Sedol?
Yes. The data would be insignificant anyway.
>>53458778
That's not what frozen means.
>>53458744
Alphago already decided to make a new move compared to game 2. GG no re retard.
>>53458824
basically save scumming
>>53458836
Lee made a new move first, faggot.
![The Animatrix – Ep03 – The Second Renaissance Part II [Afro] [A6FA0C00].mkv_snapshot_06.45_[2016.03.12_20.12.50].png](https://i.imgur.com/IVtz2ag.jpg "The Animatrix – Ep03 – The Second Renaissance Part II [Afro] [A6FA0C00].mkv_snapshot_06.45_[2016.03.12_20.12.50].png")
Bless all forms of intelligence.
>>53458836
because sedol made a new move
>>53458858
No he didn't retard. The game is identical for the first 11 moves, move 12 (alphago) is the first deviation.
live stream?
Makes me wonder, if you field AlphaGo against a complete amateur, would it throw it off balance? I assume the computer predicts or reacts to what it "believes" to be rational moves by its opponents.
>>53458903
ALPHAGO IS BLACK
>>53458920
Fuck off niggerlover.
>>53458908
https://gaming.youtube.com/watch?v=yCALyQRN3hw
>>53458908
LIVE STREAM
https://gaming.youtube.com/watch?v=yCALyQRN3hw
>>53458913
no enough with these theories, the AI doesnt overfit against pros since it also does a tree search which is a deterministic procedure that finds and optimal course of actions as long as you throw enough resources at it
>Lee is using ancient super nipponese tactics
HES PULLING OUT ALL THE STOPS
HES CHANNELING HIS INNER SAI
>>53458913
I believe it would only play just above the level of the opponent it is playing so any pro looking at those games wouldn't see that AlphaGo is a great player.
>>53458937
>>53458945
i want to see the players not smoe fags talking
>>53458913
of course not, it's just as reactive as it is objective aware, if a scrub goes against it they'll just get rekt.
A pro is better than even amateurs at playing oddly, for instance.
>>53458975
not true, it would stomp amateurs
>>53458913
Alphago choses the best move for its current situation. Not what the other player is expected to do
>>53457584
The only question is if AlphaGo can beat the world champion Gungi player.
>I don't think they'll let me take them actually
BTFO
T
F
O
this board game looks retarded tb.h
>>53459016
Well considering the world champion Gungi player is dead...
Is he winning right now?
THERE IS HOPE
The homolust is real
>>53458980
You want to watch a computer and an asian guy thinking for 6 hours? Just go watch ghost in the shell on replay.
>How long is a millisecond?
I have never won a game of Go against a computer or at all even. Is it better to just keep practicing against a low grade AI or would it be better to play with another human who has also never won a game of Go?
It seems like a good game to get into even if the AI's are taking over.
>>53459076
Don't most people advise playing against humans because A.I (with the outstanding exception of alphaGo) aren't very good at Go?
>>53459076
Against humans. AIs are a joke, but if you don't know go you won't know what to try and you won't find the 100% guaranteed victory.
ahhh go, I wish I had the time to practice (._. )
>>53459049
>not YAMATO DAMASHIII.jpg
Why don't they send the Japs in to beat AlphaGo-tan?
>>53459076
Play with humans, they make more mistakes.
>>53459029
Aww shit dude I wasn't finished with the show :(
Is the computer used for the ai datacenter hosted and running remotely? Wonder how much processing power and the code size / games / moves database is.
>>53459107
That Arc ended 2 years ago!, sorry ;_;
>>53459118
There is no database. It was trained on a database, but it doesn't have one to refer to when playing.
It's running on a 1920 CPU and 280 GPU cluster
How does AlphaGo understand complex positions, it just evaluates shapes right?
It blows my mind that it may not even understand what a sacrifice is but it'll play around it
>>53459049
lel.
Can someone put a SS uniform on the guy on the left?
>>53459159
It uses "shapes" to choose what moves to look into. It still looks ahead many variations but this kind of intuition allows it to drastically reduce the number of silly variations.
>>53459071
He put the white stone too low
>>53459214
is black already winning?
i wish japs/gooks/chinks had interest in Chess so they could btfo western pigs in yet another mind game
>>53459238
Heresy.
>>53459187
>SS uniform
>not tanuki or some frog creature
>>53459140
It's ok ;~; I only recently got into the show.
>>53459258
a tanuki is perfect
>>53459226
Sick reference bro.
>That's what makes it easy to play with you.
REDMOND CONFIRMED FOR BURN MASTER
Can someone answer me an important question:
Does alphago have genuine intelligence or is computing power at the point where we can brute force go?
>>53459323
Pattern recognition + machine learning + brute force.
>>53459323
it is sentient, yes
>>53459323
ill-formed question, define genuine intelligence
>>53459323
Define genuine intelligence
>>53459341
No bruteforce.
>>53459323
>brute force go
very sick of hearing this
this is literally impossible, not enough atoms in the universe
read the fucking wiki on go
>>53459323
Neither. It uses machine learning to avoid brute forcing go, but I wouldn't say that machine learning could curate "genuine intelligence".
>>53459323
It might have a sort of animalistic, non-sentient intelligence, but that is hooked up to a more normal bruteforcing computer.
The two combined make it superior to any human.
>>53459359
>>53459361
for local fights
>>53459323
It's not brute forcing in the way chess was beaten in the 90's. Alphago's AI tries to approach the game in a somewhat similiar way that a pro would.
>>53459323
the important question is:
Does it haves a soul?
>>53459323
>>53459341
There's no brute force involved. That's impossible with current processing power, given the amount of possible states of a 19x19 go board.
>>53459388
No, only Quarians have souls.
>>53459323
AlphaGo is more than just brute force. We can't brute force Go nor will we ever be able to.
AlphaGo is a combination of brute force and intelligence.
>>53459359
It's not a true 'bruteforce', but it does do tree searches more powerfully than any human.
>>53459388
Maybe a Seoul
>>53459393
>>53459361
But there can only be so many states based on the current layout of the board.
For example it could use a random opener. Then brute force each active area of the board. 5*5 with 3 states or example isn't hard to bruteforce
>>53459388
No, Koreans don't have souls.
>>53459323
>>53459359
It probably has a database of every playable move for openings and closings so that might be its use of bruteforce.
Who /comfy/ here?
>>53459422
>isn't hard to bruteforce
Feel free to write a program that does it on a reasonable amount of time.
>>53459412
Goddamit Carlos, you little scamp!
>>53459434
it doesn't have any kind of database.
>>53459434
You're retarded. Kill yourself.
how did they manage to download assburger into alphago?
>>53459422
>Then brute force each active area of the board.
This is more or less what it does. It uses a neural network to determine which moves are worth considering in each step and which are not.
It then does a constrained brute force search on this reduced tree. (Constrained in the sense that it cuts off at a maximum depth. It can't brute force all the way to the end)
>>53459434
The only database it uses is the neural network weights.
>>53459323
It's basically doing statistical approximations of what the best move is. It's basically finding the best fitting line but instead of a line it's a more complex function.
>>53459425
>>53459388
This may seem like good bait locally, but cannot make life.
>>53459473
>It then does a constrained brute force search on this reduced tree. (Constrained in the sense that it cuts off at a maximum depth. It can't brute force all the way to the end)
You forgot the most important part. It only goes so deep, then it uses a second neural network to evaluate how good the outcome it.
a deep learning neural net is a database
>>53459359
There the brute force of over 1000 cpus doing number crunching.
>>53459474
It's a highly non-convex, billion-dimensional manifold.
>>53459510
Nope.
>>53459510
you are retarded
/g/: filled with cutting edge artificial intelligence experts
>>53459510
lol
>>53459527
Holy shit. Were can I read more about it?
>>53459510
>>53459531
I hope enough interest in Go starts up that a general gets going in /tg/. I've never played but been watching these commentaries, it seems like a lot of fun and a lot of room for discussion.
>>53459542
All you have to do is read the paper to be more knowledgeable than 99% of /g/
>>53459542
/g/ complains about calc I in their CS majors, I don't think there's any pure mathematicians or statisticians who haven't killed themselves yet in this board
>>53459323
>>53459393
Oh yeah I should add that we should be getting rid of that 'intelligence' word altogether, because there's no unique approach to it, as it has been proven time and time again. If you are referring to human intelligence, then yes, though maybe it's not intelligent in the human sense. If you wanted to define intelligence within human boundaries then it clearly isn't, for example it can't speak or understand spoken words. Restricted to go, it clearly seems intelligent to me: it considers the best possibilities and punishes whatever mistakes its opponent makes (or rather than saying 'punishes', it just takes advantage of whatever openings the other side makes).
I'd go as far as to say that it doesn't matter how it achieves what it does, as long as it's as good or better than us at it. Then it's intelligent. Deep Blue used brute force to win at chess? But it did better than us. We use brute force too with simpler problems in the way of trial and error, yet it's a valid way to come to a solution and a valid one at that, as it achieves what it was set to solve.
>>53459434
No, Google doesn't have a true Go database yet.
The Chinese Go Association, The Japanese Go Association, and the Korean Go Association all have their own databases and servers and Google did not ask any of them to help with AlphaGo's creation and development.
However, that Fan Hui guy helped the most by playing with AlphaGo like everyday since his defeat and spilling out everything he knew about Go to Google.
However, AlphaGo is still in the alpha version.
It doesn't even have a final form yet.
>>53459323
If you mean sentient, then no
It's just MCTS + deep learning neural network
The concepts are not new as well
>>53459565
all the mathematicians are on /sci/ trying to sound smarter than the other idiots there
>>53459565
Calc needs to be de-emphasised for linear algebra in CS curriculum anyhow
Come on, guys. You know bruteforcing can only go so far. Even chess engines start to get lost when they reach the endgame, where they're provided with tablebases.
Lee is playing really submissive right now. In his heart, he has already resigned.
>>53459565
Calc 1 is considered a weeder course just like how Organic Chemistry is considered a weeder case for a majority of wannabe doctors.
Colleges like to fuck with students in so called introductory courses.
Why do they pick the stones in that funny way?
Do diagonals make a line?
Is there go learning software for lunix?
Because watching the summaries here is what I understand
>Alphago can process 10, 000 moves a second (according to conference guy)
>Average thinking time is 3 minutes
>1,800,000 moves calculated (Lets say that equates to 1000 possible outcomes)
>Then it just MAXs each tree and spits it out.
That is pure brute force yes?
>>53459595
>>53459605
>kids can't handle some introductory analysis
this is why /sci/ thinks we're the dumbest board
>>53459606
why do you asks stupid questions?
>>53459557
Just go make one yourself after the 5th match is over. Take initiative.
>>53459627
branching factors
>>53459633
Even Bill Gates mentioned he got a C in Organic Chemistry before dropping out of Harvard.
>>53459627
no you retard
>>53459651
What does organic chemistry have to do with mathematical inability
>>53459627
It's not bruteforcing if you only consider 0.0000000001% of the possible variations.
>>53459578
From page 3 in the paper linked in the thread.
>We trained a 13 layer policy network, which we call the SL policy network, from 30 million positions from the KGS Go Server. The network predicted expert moves with an accuracy of 57.0% on a held out test set, using all input features, and 55.7% using only raw board position and move history as inputs, compared to the state-of-the-art from other research groups of 44.4% at date of submission (full results in Extended Data Table 3).
also line breaks are dumb and also stupid
>>53459665
good post retard
AGA stream about to start
https://www.youtube.com/c/usgoweb/live
Any decent go players here? How bad for Lee is that top left corner?
>>53459633
I have a degree in maths but I think lin alg is more important for CS than calc. I am kind of dumb, but I handled introductory analysis just fine.
>>53459637
No answers? So you're as ignorant as me.
>>53457584
>game 3 article
>and by 48 Lee was desperately fending off White’s powerful counter-attack.
:^)
>>53459690
And by top left I dont't mean that lone stone, I mean all of mid left to top left.
>>53459696
no point in answering retarded questions
>>53459578
They should mount the final version on a qt loli made by boston dynamics.
>>53459412
kill yourself
>>53459735
everything from boston dynamics rattles and waddles and makes loud noises in really spooky ways
>>53459690
>>53459717
There is a lot of potential for attack right now if he gets initiative. That 1 stone isn't pointless either because it still has value if black chooses to attack.
>>53459719
I'll just wait for someone who isn't an idiot.
BUILD WALL
>>53459735
>>53459751
I feel it would fall right into uncanny valley territory. It'd be extra unsettling to play against it, one more advantage for the AI I guess.
>find a chink
>>53459743
Edgy kiddie, careful you'll cut yourself.
>>53459767
exactly, it'd bring poker intimidation into Go, make it less nerdy.
But, white, you ARE the chink.
>>53459751
>>53459767
The scariest thing at Boston Dynamics isn't that robot dog or robot human but that RC looking robot they called Sand Flea or whatever where it can jump over building rooftops.
Lee Sedol actually winning according to Myungwan Kim 9p
>>53459761
>No QT on AGA stream
Well I'm out.
AGA stream live
https://www.youtube.com/c/usgoweb/live
>>53459792
Retract that, he was joking.
>>53459792
Thats how start of 3rd game looked like too. Then, he got stomped into the ground.
>>53459693
I have seen programs who completely took out calc or linear algebra and desu i think it's completely retarded even if CS students think any math is useless
neat move by sedol
>>53459753
There's only one toxic guy here as far as I can tell.
I don't understand your question, diagonals (without the enemy already cutting) make a strong connection shape cus it's hard to cut per se.
>>53459735
Metallic-sounding screeching voice: "That was a bad move, onii-chan"
More rattle as the loli moves her arm to place her stone on the board
Why doesn't Magnus Carlsen face off against AI? Filthy westerners have no honor.
>>53459788
p.cool desu
https://www.youtube.com/watch?v=6b4ZZQkcNEo
>he didn't play go as a child
>>53459847
because no man can win against machine in chess.
THICKNESS
>>53459806
Is AGA still the superior stream now without the QT?
>>53459852
What the fuck lol
>>53459870
Yes
>>53459859
alpha go is still a child :^)
>>53459847
Why would he? Computers have BTFO humans in chess for years. Might as well have an Usain Bolt vs motorcycle race.
>>53459847
Magnus Carlsen is an AI
