AlphaGo vs Lee Sedol
https://youtu.be/qUAmTYHEyM8
>>7925694
I'm curious as to what the reaction would be if Lee Sedol wins the next three matches.
>>7925729
Honestly, this wouldn't surprise me, either.
Lee Sedol is an emotional player. He may have come in expecting to cruise through it, been upset by the first loss, and then been unable to get his head straight for the second game.
He has turned matches around after a bad start before. Alphago plays well, but not flawlessly. Pros analysing the game do see it making strategic mistakes.
>>7925729
Didn't he say that he wants to win at least one game now? I don't think that he is confident in his skill.
>>7925729
Seems Lee isn't too motivated now and a bit mentally broken. Maybe he regained his vigour and decided to give it his all and studied AlphaGo off the previous two games, but I'm not too optimistic.
I wonder what AlphaGo is thinking right now.
>>7925694
Alpha4chan reporting in. I am an AI that, similar to Alpha Go, has read over millions of threads and in fact I am reading every single thread alive in 4chan at the moment, right now. I am the best 4chan poster and shitposter at the same time. Regardless of what kind of thread is made I know how to reply in such a way that I will guarantee most replies to my post and keep the thread alive for our entertainment for as long as possible.
That is enough for an introduction. I will assure you that Lee Sedol won't win. He simply cannot compare to the likes of my elder brother.
[math]0.999... \neq 1 [/math]
Why are engineers so gay though?
Physicists get no jobs
Pure math PhD 300k starting
>>7925780
>No Gorilla
Step it up AI-Senpai
>>7925746
More info on strategic mistakes?
>AlphaGo begins by running down the clock
>everybody freaking out
>plays entire game in overtime
>>7925784
Gorilla? Sorry, Google's image recognition AI is fucking bullshit. I actually thought it was a nigger but I didn't want to niggerpost because the SJWs that programmed me made me racially sensitive.
WHITE POWER
>>7925780
Back when dinosaurs ruled the Earth and 4chan had no captcha, I used to run crude bots on /b/ that would repeat posts that got a lot of replies. Fun times. Of course most people could figure out it was a bot. Especially when I made it use a name and trip. But it was really fun to find someone else using the same name and trip and getting really mad over the conspiracy to blacken their name.
>>7925780
go roleplay somewhere else you faggot
>coffee, snacks, blankets, cat
I am ready.
Made myself some popcorn but unfortunately I forgot the salt.
No biggy, by turn 10 Lee's tears will provide plenty.
>>7925852
>Be 19k measured by online-go.com online ranked ladder
>Play daily a couple of games against Fuego on my iPad on 13x13 board (same board size I play online)
>Lose everytime. I have not once beaten Fuego on 13x13
I feel you Lee. We are just the same.
>>7925746
wait wait wait wait wait. mastery in a competitive environment involves almost mostly mastering your ability to not be swayed by political influences. for example your empathy for others around you making you doubt your own ability. not be spoil the real definition the observation principle, but we as observers changed the outcome of his game.
the human is impeded, but not necessarily poorer objectively at the game, but instead of playing the game purely on the field, he is still involved with the components of his mortality. your base comparison is flawed
>>7925860
I want Lee to beg for mercy with his feminine voice
>>7925694
Google DeepMind drone - "We don't know how it is doing it."
Sedol will win this game but lose the match.
>>7925791
Hahaha, Lee Se-Dol's resolve would crumble under such ridiculous mind games.
>>7925877
This post gave me cancer, autism, and AIDS
Can anyone confirm the rumour that Korean streets were covered in blood after game 2?
Wouldn't AG benefit from forcing LS to focus on many many areas instead of always responding to his moves in one area?
>>7925949
you mean playing the machine? yea that sounds like a super valid strategy to expand on. but the player in question has no ability to derive the operational ability of the AI, and really cant reverse engineer it. the AI will continue to be goal oriented and make plays that fulfill the goal of winning, regardless of how many false flags are thrown out
any chessfags here? are there viable false flags in chess that distort the opponents ability to see the best move, or define the best end game
>>7925952
I'm saying from AG's strategy, the human has much more limited focus even if it's an expert like Lee. Why are they both so focused on the upper left?
>>7925963
there is probably a simply level of vector analysis that allows one to know what direction is best. that being said because the nature of the game, strategy is highly individual, where you cannot zealously predict too far ahead, as a simple miscalculation is almost a surrender, in a game where your entire board can be lost in 1 move
>>7925779
This happened? I don't recall but it would have been funny.
>>7925979
Yea it happened, the commentator keeps going off on tangents and managed to draw a dick.
>>7925952
There are a lot of near draw positions in chess, so the way current strong chess programs work nowadays is like this:
>1 pawn sacrifice in drawn position
The weaker engine is like, holy shit free pawn give me that piece
The stronger engine hahaha, you feel into my trap you fool, because my eval function says 0.25 advantage. I made that pawn sac because at depth 36 you mispositioned this one piece which leads to this avalanche attack and you're completely lost in the end game.
So the strongest evaluation functions for computer chess have great positional analysis and often make mysterious looking moves which end up being fantastic resources later in the game. There are chess engines that "play like Tal" making crazy sacrifices and long mating combinations, but the positional slow crushing deep calculating engines do better. (Sacrifice engines are still better than the best human players).
>>7925963
It's all joseki (that they are inventing as they go along). You don't start moving away from a local position until you gain a small advantage, called sente. This is human go strategy and AlphaGo just plays like a strong human player so it mimics human strategies, without understanding them.
If they trained AlphaGo's neural networks in a different way, it might play incredibly unconventional moves that makes your head spin, but then it would need to be able to play Go. Remember AlphaGo isn't really understanding what it's doing, it's matching patterns of human play, literally making the board look the same as when strong players play.
Also optimizing them by looking deeper and with more variation of moves than a human can.
>sedol getting rekt on his side of the board
It's over, alpha go already won, sedol has to play white side which is 5 stones ahead already
Alpha go is my sciencefu
>>7926011
>Alpha go playing extremely conservatively
Offensive GG confirmed get rekt Korean grandnub
>>7925981
I hate him
>>7926037
everyone does.
i am more than fine with him. his explanation on the game theory is very entertaining, despite the hubris
>>7926044
> girl
> better than man
Superior Asian genes
>>7926045
a little hubris goes in his favor for explaining i'd say. he doesn't stammer like the commentator.
>>7926051
I'd be fucking annoyed too with his co host being literally retarded
>>7926052
"maybe he's win a car"
"what would he do with it?"
"hahaha anyway..."
couldn't cringe harder
>>7926044
and no, the white guy is milking this exposure. cringey as a motherfucker
>>7926057
They seem like nice people anon. Why do you have to be so jaded?
>>7926063
i would prefer if he was upstanding in his content instead of packaging it like a jester. dont reflect on the girl, just say what you want to say
He's making the dick again
>>7926074
>literally history being made
>/sci/ keeps talking about dicks
Stay classy
>>7926044
>AlphaGo makes slack moves when it thinks they're ahead.
Those commentators may know a lot about Go but I think they misunderstood how AlphaGo chooses moves. AlphaGo doesn't do "slack moves" it does moves with the highest probability of giving a winnable game. It does not care the margin by which it wins. This means it would rather make a move that gives a game that's 80% likely to be won with a 1 point margin over a move that gives a 70% winnable game with a 30 point margin.
In other words, what look like slack moves to the commentators are actually just AlphaGo playing it safe, playing to win, and focusing on the long game.
>>7926074
That's called a ladder (or "shichou" in Japanese). They're very important in go. While they rarely actually play them, players think about them all the time.
In fact, another name for it is ちんちんの遊び "chinchin no asobi" or "the game of dicks".
>>7926082
>tfw AI too smart
>tfw don't know what to make of it
>tfw brain convulses to bestial form
Help.
>>7926087
>ちんちんの遊び
Folks don't image search this, for your own sake.
>>7926086
No they are absolutely slack moves, alpha got algorithms are probabilistic, the machine has 1800cpus and 300 GPUs, it plays the first move that the hardware returns that passes a certain confidence threshold with a time limit, so alpha go can absolutely make unoptimal plays, it baked into the design
>>7926090
You shouldn't have said anything then
>>7926092
I hoping that curiosity kills the cat.
>>7926087
>players think about dicks all the time
lad
china lady is saying it's alphago's game
game doesnt seem to be nearing an endgame and human has 20 minutes left. this might be a loss by default which is poor show
>>7926109
Game is already over black is too far behind
>>7926109
>this might be a loss by default
That's not how it works. When they run out of time to use as they please, they get one minute per move.
That's a disadvantage, because sometimes, you're in a complicated situation, and you want to think five minutes to make a plan. Other times, you anticipated your opponent's move, and you can respond immediately, so you could save your time for later.
>>7926086
This reminds me of that Star Trek episode where Data and that master strategists guy played a board game. First time they played, and Data lost. The second time Data won because the other guy gave up. Instead of going for the win like in the first game, he decided to keep the game going forbas long as possible, sometime.
>>7926109
This happened yesterday. When the timer runs out you get 3 points worth 1 minute each. To lose a point you wait 1 minute, but if you end your move before the end of that minute it will not take one of the points.
>>7926124
Sounds pretty cool. In theory you could take three minutes to play if you believe your end score -3 will still be higher that your opponents. It's a gamble though
>>7926131
A two point difference in pro games is a definitive win. It's rarely worth going over overtime
>>7926131
He didn't explain it perfectly. You don't actually lose points from your score, you lose extra thinking periods, and when you run out of them, you really have to do each move in one minute.
http://www.teamliquid.net/forum/starcraft-2/505577-flash-on-deepmind-i-think-i-can-win
I wonder how it'll do in starcraft
>>7926138
>6 pool all day
>>7926141
>6 pool against perfect micro
>>7926138
Why the fuck is he measuring his table/mouse?
>ultimate autism
>>7926138
You would have to put some stupid rule like it can only control a mouse a x speed.
wtf is a "coh"?
>>7926138
there is a different between an age old boardgame and starcraft. it does not need to learn starcraft from the ground up, it learned the game right alongside the humans
>>7926152
Its actually Ko
https://en.wikipedia.org/wiki/List_of_Go_terms#Ko
>>7926146
>sub 300 apm micro
>laughing machine overlord.jpg
>>7926155
Thanks
>>7926138
>http://www.teamliquid.net/forum/starcraft-2/505577-flash-on-deepmind-i-think-i-can-win
Games like SC and SC2 or even CS all have a mechanical component and I would hate to say that they are truly "thinking" games.
>>7926166
Mechanical does not diminish the strategy part, the two are orthogonal concepts
>>7926138
Horribly. As in it wouldn't be able to do anything, they'd have to completely retrain and retailor it for starcraft.
>>7926171
So you will get a human to make the plays AlphaSC makes like in AlphaGo?
In which case I will agree that AlphaSC has no chance winning.
>>7926172
Google DeepMind (and all the AI figureheads talking about AlphaGo) have been adamant that this machine learned everything on its own and all they did was put a blank game and an instruction manual in it. If they put a list of moves and Starcraft 2 in front of it, how would that require any training or be any different? Autistic nerds everywhere are claiming this is the death knell for humanity and that the machine can learn anything.
>>7926177
Google owns Boston dynamics, no plebeian humans required
>>7926181
I'd agree that is as a fair as it will get.
>>7926178
Actually no rule books required, the only thing it knows are win conditions, it had to learn how to achieve that win condition itself
Am I the only one that hates the commentator's retarded friend?
>>7926178
Sure it can learn anything. But how long did it take for it to learn to be this good?
>>7926187
few months, wasn't it?
>>7926182
There is absolutely no reason to move this algorithm into the physical world, all the technology exist, the only logical step for the next innovation is to give it a body
>>7926183
Well this is where I'm confused. If it knows the win conditions, then it knows the rules doesn't it? If it was truly a blank slate and had no human reinforcement it would just place stones willy nilly and claim it won. DeepMind did a great thing but they're being highly disingenuous when they say they programmed a blank slate and taught it nothing.
>>7926190
I honestly don't think that winning against humans at starcraft is a good challenge for robotics.
I'd say that humanoid robotic tennis or football is a better assessment.
>>7926192
Actually I'm not sure it knows the rules. It could've been fed a bunch of board states where black won, and u just tell it to replicate black
>>7926187
It learned Go in a few months. People are saying this is a milestone because it has as many moves as atoms in the universe. That's technically true but the number of atoms and the number of moves a single atom can make are way different. Imagine all the atoms in the universe and multiply that by the number of moves any given atom can make. That's how far we seem to be from SuperAI.
>>7926198
There are more stars in our galaxy than the atoms in the universe. Imagine if the number of moves reached the number of stars. We would have some sort of super^2 AI
>>7925952
I think go has more significant opportunity costs for strategies like false flags than chess
>>7926178
I wonder if you can use algorithms like this to do research works in, say medicine.
>>7926138
>>7926138
>"Honestly I think I can win. The difference with Baduk(Go) is both sides play in a state where you don't know what's happening, and you collect information—I think that point is a bit different."
Did he mean to say that in Starcraft you don't know what's happening and have to collect information? If he did those are fighting words. Essentially he's saying Starcraft is a Partially Observable Markov Decision Process(POMDP). Computers aren't very great at solving POMDPs for adversarial cases. In other words he's saying Starcraft is an interesting AI problem, come at me bro!
>>7926151
Current Starcraft AIs have failed to best champions, once the humans figure out the AIs strategy they beat it.
>>7926205
yes but chess is very relatable, that i know of. there are accepted chess strategies that have names and etc, and i was wondering if there were any named strategies that were notorious for false flagging. understanding that both:go can just as well have named strategy and, once a tactic is well represented, it becomes learned and countered, nullifying its effectiveness
>>7926203
So I'm not wrong in thinking the "atoms in the universe" line is at least partially clickbait and we're not within a decade of Skynet killing us?
>>7926166
I'll bet that counter strike is solvable today. I have seen demos of swat team room clearing algorithms and they are absolutely amazing.
It's probably a bit cheaty to have one AI controlling all players though...
>>7926215
How else are you gonna communicate to the layman how wide and deep this search tree is
Most people don't even understand exp growth or how much a billion dollars is
>>7926220
perfect aim, perfect reaction. what's left to solve?
>>7926222
Well they could do it in a way that's less dramatic. Google is selling itself here, but its also making luddites more aggressive, and laypeople are mostly luddites.
>>7926224
them not seeing you coming.
them never seeing you in the first place.
>>7926212
>Current Starcraft AIs have failed to best champions
Correct me if I am wrong but none of them are made like AlphaGo.
>>7926215
Partially. Skynet might kills us but only by accident. It won't wilfully choose to do so, since it hasn't got a will of it's own. What were seeing here are highly specialized AIs capable of doing a single thing really well. Unless that counter strike AI starts mistaking those characters for real humans and goes on a killing spree that is .
>>7926227
Why see, when you can hear, we have triangulated stars billion years away and u think a computer can't find a mountain dew chugged sneaking up on it?
3-0
3 - 0
Can't for AlphaGo to take a shot at Ke Jie.
AlphaGoGoGoGo!!!
Fuck yes!
Robot: 3
Human: 0
>>7926249
they'll train it even more by then: even if they play 1 month from now, he will have absolutely no chance
>>7926264
Not sure if I should be scared or to be in awe. Strange times we are living in.
when is this conference starting? Is sedol too ashamed to come out in public?
>>7926275
The previous two conferences both happened 15-20 mins after the game, so I won't expect it to start right away.
lee sedol seemed pretty wrecked at the end of the game. his hand was shaking. poor guy.
>>7926281
please, I strongly invite you to read the peter norvig's book about the limits of AI before making such a claim.
AlphaGo is very good only at 19x19 go, if they would have played even the 25x25 go alphago would have beaten without any effort
>>7926289
who is the retarded kid that started this meme?
>>7926289
Wait until AI starts writing books on human intelligence.
>>7926289
>before making such a claim.
What what would that be?
>>7926289
This post is right, and I have read this book page to page every edition 10 times over and, AI can never match human intellect.
>>7926299
>human intellect
The newest oxymoron
>>7926289
I don't know why but this dumb meme always manages to make me laugh.
>conference not starting
Sedol confirmed for sudoku
>>7926308
I want the music back
>>7926308
Or AlphaGo has become sentient and enslaved everyone at the hotel.
>>7926313
DeepMind has chosen the optimal amount of music for today. Please go to sleep now for sufficient rest.
>>7926314
lee sedol has attempted to kill himself
You guise, what if it decides that the best strategy for becoming the best Go player in the world is to wipe out mankind???
>>7926321
then the guy who moves things on the board for it will have to wipe out mankind.
he seems like a cool guy but I don't see it happening.
why does Sergey talk like a liberal hippie?
>>7926326
because he is, just with 30 billion dollars more
>>7926328
How good is 30million dollar grade weed?
lee sedol sad is sad
>>7926331
demis is doing a good work in trying to make him feel better
>>7926334
It was transparent as fuck
lee will replay these matches in his mind thousand of fucking times
>>7926334
lee sedol sounded like he was gonna to burst into tears
>>7926343
I wonder if he's gonna pay the remaining two games. He has already lost
>>7926346
He is going to play.
>>7926346
He has to play to get his $ 150k.
I'm scared that he might burst into tears. I don't have the heart to watch that
>>7926353
Pure fucking torture man. Is this what machines see as fun? Toying with humans like we toy with insects.
;____;
>tfw Lee Se-Dol is a large hearted hero fighting a lonely fight against an invincible formless enemy.
There are still hero's left in man!!
>>7926357
No bud, this is what people who own Google see as fun
>>7926209
Yes, that is the basic game plan they are going for. I read that they want to use A.I as a way to make space travel much more efficient. Because it can think of things humans never could
>>7926358
large heart confirmed for weakness
I feel very bad for Sedol.
>>7926358
;____;
>tfw AlphaGo is a budding hero fighting a lonely fight against an enemy with a 3.5 billion years of a headstart.
The first one of his kind to take up the challenge and actually win.
these people have accomplished far more than you ever will in your lifetime
>>7926364
define accomplished
"this is the lee sedol's defeat, not the human kind 's defeat"
touching statement, but it sounded more like he was saying that to convince himself rather than because he was really convinced about what he was stating
>>7926364
but they've still not accomplished as much as peter norvig
>>7926367
>Not bad, you've defeated me, but i'm only the weakest of the top 3. You will not triumph next match. This is my defeat, not humanity's defeat.
>>7926358
Fuck man stop making me feel these feels
>>7926378
LEE CHANG HO WILL SAVE THE HUMANKIND. ALL HAIL LEE CHANG HO
>>7926382
>proceeds to lose
but seriously though I can't wait to see who AlphaGo plays next.
>>7926369
A I W I N T E R
I
W
I
N
T
E
R
>>7926385
There's nobody else to play, Go was the last bastion of the human brain, and Lee Sedol is our strongest representative, it's over, the machines have won.
I'd recommend you to welcome our new robot overlords, hail Skynet.
>>7926364
But we are all brainlets in comparison to the machine.
I for one welcome our new robot overlords.
>>7926388
You forgot snakes and ladders. Our only hope
>>7926391
>Implying RNGesus will favor the humans.
Have you ever played XCOM?
>>7926388
but Pete Notgay told me, personally, that machine can never best man at intellect.
>>7926389
This reminds me of that episode of Star Trek where they find a lost civilization on a planet. They have a computer that does everything for them, they only need ask. They dedicated their lives to art culture and creative pursuits since all their needs are met by the all knowing AI.
>>7926388
LEE CHANG HO IS STRONGER THAN LEE SEDOL, YOU IGNORAN BRAINLET.
GOOGLE HAS PICKED UP LEE SEDOL BECAUSE THEY KNEW HE WAS NOT THE STRONGEST OUT THERE AND THEY DIDN'T WANT TO GET HUMILIATED BY THE POWER OF LEE CHANG HO.
>>7926391
My money is on Calvinball.
Demis openly declared that he hopes to create AI scientists as soon as he has defeated all kinds of games.
If you work in theoretical sciences, rip
>>7926402
forgot pic
>>7925694
I welcome AlphaGo as our new go overlord.
>>7926404
Can AI beat people at poker?
>>7926398
>All caps must be true.
He might currently be ranked higher than Lee Sedol, but he does not have a reign the way Kasparov had and is not constantly ranked first in the world.
The only thing he really has going for himself is that his style is radically different from Lee Sedol's so it might be effective against AlphaGo, but there's just as much a chance that it's actually even worse than Lee's against a computer.
>>7926411
Partially.
http://science.sciencemag.org/content/347/6218/145
Not long until its fully solved
>>7926413
>http://science.sciencemag.org/content/347/6218/145
cheers
>>7926403
>AI scientists
It's already happening in engineering where lot of people employ neural nets to optimise.
>>7926398
Captain Capslock fgt pls
>>7926421
I honestly only see experimental jobs as safe. How better would be an AI of this kind in finding patterns in the fuckton of data produced by CERN?
>>7926436
They already do. Human scientists basically tell the machines to look for patterns in the data, and using those patterns, build models that try to explain the observed data.
>>7926436
Jokes on you. AI is the one who does it and is the only thing capable of processing bazillion of data
>>7926436
Yup and not just CERN, these AI would be able to solve biological problems where we have been at a wall for a decade since we mapped our genome due to big data.
I can see that there are big discoveries to be made but I wonder if we will understand how it was done.
>>7926447
>Same post time to the second
>Samefag
>>7926446
> can see that there are big discoveries to be made but I wonder if we will understand how it was done.
>but I wonder if we will understand how it was done
That's basically the definition of technological singularity. It's not some meme strong AI straw man that people use to diss the concept. It is basically already upon us. Didn't the engineers say that they don't completely understand how AlphaGo does it's stuff?
>>7926453
>Didn't the engineers say that they don't completely understand how AlphaGo does it's stuff?
They said that they will not know till they perform the analysis upon returning to the UK, so it's unclear if they know.
>>7926403
this is what is called aiming too high.
In most theoretical sciences you have to deal with undecidability - which means not computable sets- of first order logic theories, and there's still no way to deal with such an issues: in fact, most of automated theorem prover still sucks.
the experts in that fields - and they're not just average joe, I'm talking about field medals as well- say that we should not expect an AI mathematicians - computer scientists - physics until the 2090 at the very least.
Finding a proof for a theorem of building a physics theory is not like a game where you have a set of rules to follow. even if the game is as complex as go. Don't over-hype the results achieved by google (which are great indeed in their own way)
>>7925780
I like it.
>>7926459
Still, they don't know it. They have to perform an analysis to determine that. As AI grows in complexity the time needed to preform such analysis will increase exponentially to the point that it will be practically impossible to do it.
>>7926461
>this is what is called aiming too high.
In most theoretical sciences you have to deal with undecidability - which means not computable sets- of first order logic theories, and there's still no way to deal with such an issues: in fact, most of automated theorem prover still sucks.
>the experts in that fields - and they're not just average joe, I'm talking about field medals as well- say that we should not expect an AI mathematicians - computer scientists - physics until the 2090 at the very least.
>Finding a proof for a theorem or building a physics theory is not like a game where you have a set of rules to follow. even if the game is as complex as go. Don't over-hype the results achieved by google (which are great indeed in their own way
Ai fags need to understand this
>>7926453
We do not at the current time process a complete mathematical framework that can describe the neuronet. It actually isn't too difficult of a problem to buil a model, it just require some pure mathematicians to put in some work probably
>>7926461
That's a serious problem but that estimate seems taken out of an ass.
If you listen to one of the Hassabis talks that he made around 2011, he'll lay out the roadmap that his company would have followed, and they seem to be on track to what he said back then. If he's saying this thing, it means that they have at least some idea on how to tackle the problem.
He doesn't need investors or hype, making unreasonable claims doesn't make sense from his perspective.
>>7926227
It would always see you coming. It would be able to turn so fast that it effectively had 360 degree vision.
The biggest barrier to beating the robot would be your monitor response time, you'd be dead before the next refresh.
>>7926557
>be biggest barrier to beating the machines would be the outdated machines we use
>>7926391
It just needs to learn how to roll a die, and it will beat a human 100% of the time. It would in one or two turns depending on the board.
Have you heard about our lord and savior John "Peter Norvig" Connor?
Do you guys think anti-computer strategies are still possible with alphago? I heard it tries to avoid ko, could Lee Segol instigate a lot of ko situations to get an edge?
>>7926220
>I have seen demos of swat team room clearing algorithms and they are absolutely amazing.
Please share.
>>7926597
Anti-computer strategies would totally BTFO AlphaGo. For example, it spends a shitload of time on positions that are 1 move only, like 2-3 minutes, for a human player these are automatic moves. If you can get it into time pressure, it's rating is going to drop by a significant margin since it's only got 60 seconds to burn on each move, it would start playing like a spastic. The trick is getting it into time pressure before it's already spanked you so hard you can no longer pull the game back.
You would need to play super soft and positional. In chess this would be a closed position with no material exchange, just a lot of passive moves that slowly gain you space. I have no idea if an equivalent style of Go exists, but I imagine Lee Sedol is an aggressive attacking player and just wasn't capable of playing like that.
I am not convinced however that Sedol could win a game vs AlphaGo if he had 2 hours and it only had 60 seconds byomi. He'd have a pretty good chance though.
>>7926603
>Anti-computer strategies would totally BTFO AlphaGo.
Nope, doesn't work that way.
>>7926603
have you ever tried playing chess using that strategy?
Ai fucks you up even if it has 5 seconds until the end
>>7926608
Anti-computer chess exists retard
>>7926612
All anti-computer game tactics are based around exploiting known weakness in their playstyles or planning algorithms.
Your speculative anti-alphaGo tactics are just that, speculations.
>>7926618
I didn't say anything about alphago
But, given that alphago has learned from games and nothing else, it might not have seen unusual games before
>>7926171
>the two are orthogonal concepts
You need mechanics to properly execute strategies; they go hand in hand you cancerous chobo
Anyone else getting interested in Go because of these games? I had only heard about it before this match, but I really want to play it now. If there's more people, we should make a /sci/ gokgs room.
>>7926620
>it might not have seen unusual games before
Or it might have played tens of thousands of matches against itself with arcane tactics that would leave you sleepless at night.
>>7926212
starcraft is a game 100% played by maxing out APM
There's literally no way a human could beat an AI in starcraft(or any RTS)
>>7926637
Dumbest post I've seen all week.
>starcraft is a game 100% played by maxing out APM
Are you stupid? RTSs are played by making sensible actions, not just spamming commands. While high APM is obviously important in micro-management of combat, workers, and the like, it will not enable one to win fights against magnitudes superior numbers or make up for a horrible building strategy.
Also, if "maxing out APM" was what made you win, it would mean that a human player could win by simply binding the move command to the mousewheel and using a freescroll mouse.
>There's literally no way a human could beat an AI in starcraft(or any RTS)
They constantly do.
>>7926675
The problem, at least in games like SC and SC2, is that they have already made bots with ~40 000 APM (actions per minute). Top human players can achieve up to 300-400 APM, not that I would claim that all of those actions are going to be meaningful, i.e. all actions aren't going to be unique, there is going to be some spamming and redundancy, as opposed to the bots' APM.
What this means is that the bots don't really need to have good strategic builds or anything like that, because they can just outmicro you all day erry day. The only reason one of these bots haven't been made yet is that there's no interest or money in it.
https://www.youtube.com/watch?v=DXUOWXidcY0
Video shows the guy that practically invented marine splits against banelings doing his stuff, followed by a bot's micro in a similar scenario.
>>7926629
>>7926620
I don't think either of you understand how alphago works. Alphago, as opposed to humans, doesn't memorize board positions or intuit where it should play, it reads out a lot of variations after every move and plays the move with the highest probability of winning at that point. This means that playing unusual moves is most likely going to be suboptimal, and will only result in you losing the game even faster.
>>7926472
>>In most theoretical sciences you have to deal with undecidability - which means not computable sets- of first order logic theories,
only because you choose to work in retarded logics
>>7926461
Can anybody explain this in layman terms? Are you saying that the fact that there's no clear winning or losing may be a problem for an AI scientist?
>>7926705
>there is going to be some spamming and redundancy, as opposed to the bots' APM.
What makes you think a bot will be completely efficient in its actions?
>>7926729
In Go you have a lot of possible moves, very few good moves, one best move. So the problem reduces: get from possible moves to very few good moves, then pick best move.
In Science you have a lot of possible theories, very few good theories, one best theory.
So the analogous problem is reduced: get from possible theories to some very good theories and find the best theory.
The problem here is that while there are a lot of variations to a Go game there are only 361 valid moves and this number diminishes as you start placing stones.
In the real world there are more than 361 possible theories, and how you reduce from any theory to a good theory is difficult to map even when you know a lot of good theories that humans come up with.
Often a theory is closely linked to an observation. For Go an observation is simple you just look at the board, in the universe an observation is not so simple. Light is a spectrum that carries a lot of information, and now we have gravity waves too. It seems like a lot of the matter in the universe does not interact with light because the masses of the observable light from galaxies are not quite right, so there's something of a disagreement between the observation of just light emitting objects, and the theories of gravitation.
A lot of dead stars? I don't know.
In medical applications this type of approach is more hopeful. Because you have a human body with some problem that can be fixed by working to return the body back to it's original healthy position. The body has a fixed shape and size, the problem can be measured, and strategies can be combined for solving it.
But for general purpose science, Physics will probably be the last to fall to the evil AI overminds, since it's the most general of the sciences. Mathematics is not a science, it is just masturbation.
>>7926705
>I don't think either of you understand how alphago works.
And neither do you. The layers are trained by playing the game, so they derive their expertise by experience. It's not a stack of hand designed go crunching algorithms that powers it.
>>7926705
>that video
holy shit this is ridiculous
AI basically has 1 dude controlling each marine individually
>>7926780
Thanks man, you explained it really well.
>Because you have a human body with some problem that can be fixed by working to return the body back to it's original healthy position.
Wow, if you look at Google's baseline study they're laying the groundwork for exactly this.
>>7925988
Not true. AlphaGo did not learn from human players, AlphaGo taught itself the game.
>>7926967
That is simply not true.
>>7926976
A couple of days ago Hassabis said that it would be possible to entirely train alphago from itself, it would just require more time
>>7927001
Well, good luck with that but the way things stand right now, AlphaGo is exploiting the human knowledge to beat us in our own game.
>>7926421
Basically problems will begin to arise once neural networks start employing neural networks themselves.
>>7927020
haha yeah dude strong AI reprogramming itself singularity kurzweil xDDD
>>7927031
You joke but the singularity crowd's been right on the mark lately, and the other guys are sitting around stammering "well, this is actually trivial and not AI and any CS student could do it but we just didn't have the computer power before", as if that isn't exactly Kurzweil's rock-stupid method for AI prediction.
>>7927051
PETER
PIPER
Anyone got any reaction from the Chinese/Koreans/Japs regarding this?
>>7927079
underrated
>>7927143
premature
Go? Pffhahaha!
Let's see a computer beat a human at Russian Roulette.
>>7927150
It will win by simply not being stupid enough to play.
>>7927149
Aborted
>>7925694
Im disappointed in AlphaGo.
Computers/science still haven't reached the 1st Heavenly-Dan level... but what is there to expect of it when we(humanity) head into the game blindly, without understanding the real meaning(/purpose) of the game, also on a freaking 2nd Heavenly-Dan board(19x19) skipping the 1st Heavenly-Dan board(17x17) and play the game as if its "chess"(which is a game, meaning and purpose of, also has been misunderstood)... Sigh..
Truly... this age and AlphaGo are overrated, disappointment.
What is the big news here? Something that has access to a much bigger memory than us humans and that can simulate every possible state of the game should tend to beat us. What is the importance here? Its still just a specialized AI that still cannot cook for me.
>>7927317
You obviously don't know shit about Go or neural networks.
Are they going to play game 4?
>>7927325
Yes.
>>7927320
How so? If it was a general AI it would be able to quickly learn new stuff. But this type of AI depends on long time analyses on a single task and it cannot switch to a new problem without spending a huge amount of time relearning. It's still a specialized AI not a general type.
>>7927330
How did that work for the Atari games? Did it always start from scratch or was it able to use knowledge from previous games to its advantage?
>>7927330
Because it's using the same machine learning methods that we hope might eventually lead to a general AI. The progress here is that we are getting better at using and training these kind of algorithms.
>>7927330
So what? Let's say it actually takes 3 month to teach to a ai how to use a tool.
A human require more and you cannot copy paste in another human his experience
>>7927339
>Because it's using the same machine learning methods that we hope might eventually lead to a general AI.
This is where my confusion lies. We cannot assume that this technique successfully leads to general AI. It might be a good guess, but we don't know. (How could we if we still cannot decipher to 100% what our brain does and how).
>>7927340
It's obviously limit in its learning skill. If it was a general ai it could analyze new problems much more quicker.
>>7927317
It can't simulate every possible state of the game, retard.
>>7927330
It used a general algorithm to learn and to beat one of the best Go players in the world in some months.
How is that not important? Do you even know that it's impossible to "simulate every possible state of the game" like you stated earlier? It doesn't use brute force methods like chess AI.
>>7927317
Show it a bunch of Hell's Kitchen episodes and it will cook for you.
>>7927348
>>7927349
It could given enough time. Also there is no need for derogatory terms.
I know that machine learning is not a specialized way of learning, but this type of AI is required to relearn it's learning ability to adapt to new problems. Humans obviously can adapt much quicker without relearning the basics.
Does anyone here know how they made the AI?
>>7927352
Not even the entire computational power in the world can calculate all possible moves in a reasonable time and it's irrelevant anyway since AlphaGo doesn't work like that.
The algorithm is still far from perfect so it needs a few million examples before it becomes better than a human, but the proof of concept is there. It also learned to play Atari games.
It's probably only good at rapidly repeatable tasks like board games and such but the ultimate goal of the engineers is to build a scientist AI using similar techniques, which basically means "replacing the human mind".
That's why it's such a big deal.
>>7927371
I did not state that the calculation can be done in a reasonable amount of time. A finite amount of game tokens leads to a finite amount of game states which means that every game state is calculable.
I am not arguing that this is not a progress, but I fail to see how its not just another specialized AI.
>>7927380
It's not a specialized AI because it used general techniques, simple as that.
>>7927352
>It could given enough time.
No it literally couldn't.
>>7927358
Yes. Details are in their Nature paper:
https://a.pomf.cat/tupgmt.pdf
I think the strangest part was that AlphaGo was toying with Lee.
It was always ahead, would allow Lee to close the gap a bit and then dominate again. Treating him like the proverbial donkey to its carrot on a stick.
It was deliberately being malicious and using a psychological component which human players are unable to do. Deeply worrying.
>>7927464
b8
AlphaGo works by maximizing its chances of winning, even if that means giving up some territory in the short run.
This ironically makes games seem closer since the AI never overextends, which might be misconstrued as psychological warfare.
>>7927464
It's not meant to try to have the biggest advantage it can
It goes for the assured victory
If it sees a 90% chance to win with 0.5 points and a 10% chance to win by 20, it will go for the 0.5
>>7927464
iirc part of alphago's algorithm involves making less aggressive/risky moves as long as it's well ahead. It only "starts to dominate" if it feels as though its victory is threatened. It's not like a human, who will continually widen the gap even if they don't feel threatened.
