Match Three: Ke Jie & AlphaGo. Summoning the Demon edition

>>8937928
GO HUMANS GO

/pun not intended

>>8937928
This is fun. Read an intertaining manga called Hikaru no Go. Cool to see an actual game/match. Hope the humans win. I play chess as well and Deep Blue beat the reigning World Champion Gary Kasparov in the 1997/8

>>8937976
Agreed. Who is winning? I can't tell

>>8937928
White looks like he already lost. This is chaos.

Ke Jie hasn't played normal moves this whole game.

This isn't the way I want to see pros playing AlphaGo now. We've established that it's able to beat pros already. I don't want to see these games where they're desperate and psychology is a big factor, because their pride is on the line in each game. I want pros to be allowed to play AlphaGo every day, and try to find weaknesses in its play.

>>8937928
I see this as a battle unfolding. This is showing AI has outwitted the smartest humans. Not good news for us.

>>8938000
That approach is what happened with chess. Nowadays people use chess programs and engines to check the accuracy of moves and to look for novel ideas. It would be nice to see the top Go players using the same programs to get better.

>>8938009
That approach is what happened with math. Nowadays people use math programs and engines to check the accuracy of moves and to look for novel ideas. It would be nice to see the top mathematicians using the same programs to get better.

>>8938013
>That approach is what happened with sex. Nowadays people use sex programs and engines to check the accuracy of moves and to look for novel ideas. It would be nice to see the top porn stars using the same programs to get better.

>>8938016
>That approach is what happened with poop. Nowadays people use sex programs and engines to check the accuracy of penises and to look for novel ideas. It would be nice to see the top poop porn stars using the same programs to get better.

>>8938006
>>8938006
>>outwitted the smartest humans
for some definitions of outwit. It may be amazing at playing go, but it can't do anything else. It's not general. In addition it takes much more power and resources than a human.

>>8937928
who one the previous matches?

>>8938022
The CPU won the first 2 games of the 3 game match. This last game is just to see if the human can win.

>>8938021
You don't understand the analogy. I'm saying the game of go is "like" two general fighting on the field, where the "best" human has to offer falls short to the strategy the AI presents. I'm not saying LITERALLY we can transfer the learning to different domains outside of Go. for alpha go at this stage.

>>8938022
Alpha Go won the 2 previous matches. Alpha Go also beat 5 humans on the same team in a handicap game.,

>>8938009
No, I don't mean use it like that, I mean to find out whether it's really all that strong, or is just something they haven't adapted to.

The thing about any real pro player is that they've played thousands of games against other humans. So there are always people who know their strengths and weaknesses long before they reach the top of the pro world. Their practice partners as they reach maturity are always other pros.

A pro who somehow developed his skill without other people getting to know intimately how he played would have a huge advantage, and might even be the "best" for a while until people learned his weak points.

>>8938034
I doubt Google would let that happen. Also, AlphaGo vs 5 humans, AlphaGo won. Good marketing for Google to stop these games before any weakness are found.

>>8937976
Chess machines nowadays have 100% chance at winning, right?

Even giving up pieces?

Is this why we don't have AI vs. top chess players anymore?

>>8938027
>>the game of go is "like" two general fighting on the field
where the generals both have perfect knowledge of where all their soldiers are on the field. We call this situation fully observable. The real world isn't fully observable.

Dealing with partially observable domains might require a whole new approach. In partially observable domains you have to expend resources figuring out what the current state is. This why starcraft is the next big AI challenge.

Now even if we can transfer the learning to different domains, which we can probably do, we still need to have humans train it for that specific problem.

Heck I've actually been looking at applying algorithms similar to AlphaGo to a problem I'm trying to solve. But the question is, is it worth the cost of training AlphaGo to solve this one very specific problem?

>>8938039
komodo has an elo of 3300 while the best human player in the world has about 2800

they're on another level now

>>8938041
If AI makes perfect or nearly-perfect moves in a fully observable event, then surely it would do so in a partially observable one. It beat 5 humans on the same team earlier. Nothing shows me the AI is handicapped in anyway and given fully/partially observable events, it'd still win.

The smartest human in the world at this game is already losing.

Just because AI is crushing him doesn't mean OMG THE WORLD IS DOOMED, it's just that computation power has increased so drasticly that AI can literally generate ginormous parts of the game tree and calculate so many more moves ahead of what a human could do

>>8938050
the implications are humanity is screwed. this can be generalized eventually to general AI and areas in math, physics, chemistry, computer science where AI achieves super-intelligence in all of these areas. You are being too narrow-minded to not understand the implications, even if it takes some time to get there.

>>8938054

I can see where you are coming from, but wouldn't AI need to possess free will for us to be concerned?

IT is almost like the year is 1996 or something. Jesus you guys realize that 'AI' has already beaten people at board games right?
Go is no different, it just took more time. The moment deep blue smashed kasparov that was the end of the thought that computers could never beat the best humans.
This has been over, this would have been exciting if it happened BEFORE deep blue vs kasparov

>>8938021

but they'er buliding an AI that will do millions of things, millions of times better. This includes mimicking human behavior, It'll seem more human than, human eventually . People need to really start to get a grasp of just how alien AI can be.

https://www.youtube.com/watch?v=xZYQafx3pIQ

>>8938034
>A pro who somehow developed his skill without other people getting to know intimately how he played would have a huge advantage, and might even be the "best" for a while until people learned his weak points.
this.

>>8938068
it's almost like you're naive and think DM whole purpose is to build an AI that just plays board games.

>>8937928
What will be interesting is to pit alpha go against itself and see if anything can be learned.

The way I see it, the human player consistently makes serious misplays and falls behind in the opening, making the midgame and on a desperate struggle.

Why? Because opening theory is neither based on ideal play nor on AlphaGo play, whereas AlphaGo's openings benefit from the study of the entire body of human games, while humans have only had a tiny sampling of AlphaGo's play.

The human player who does not take advantage of human frailty, does not try to tire and trouble his opponent, is not a psychological player, will not be competitive enough to become a top pro.

If we took the top boxer, and put him in a bullfighting ring, he would likely die. Yet, to a bullfighter, what he does is not even really a fight. He's there to make a show of the helplessness of the same bull that would crush a boxer.

If you made the candidates to fight the bull compete in a boxing tournament for the opportunity, this would most likely result in the human losing to the bull every time: none of the best boxers are competent bullfighters.

I'm not convinced that AlphaGo is more than a bull in a boxing match, and we simply haven't trained a bullfighter yet.

File: fiveguys_oneAI.jpg (59KB, 690x461px) Image search: [Google]

59KB, 690x461px

I feel sorry for all the losers who still hold on to their view that humans have a chance

File: Headshot2.jpg (547KB, 400x499px) Image search: [Google]

547KB, 400x499px

>>8938057

>free will

>>8938213
>The way I see it, the human player consistently makes serious misplays and falls behind in the opening, making the midgame and on a desperate struggle.
During the second post-game press conference the AlphaGo guys said that Ke Jie was playing perfectly for the first (either 50 or 100, don't remember) moves of the game and that it was 50/50 until a specific move.

https://twitter.com/demishassabis/status/867584056095002624

>>8938212
Already happened. Google the history of the last few days

>>8938240
Yeah, I feel like I remember them saying that they trained a network specifically to do weird moves against AlphaGo.

It's a matter of time until AlphaGo like AI exceeds humans in math, economics, physics, war, etc.

It will make trivial what we find hard. At some point it will be capable of doing mathematics beyond our ability level.

People on Facebook and Instagram are busy taking selfies, posting photos, etc and there is an entire field of research out there that is going on they're completely unaware of that will fundamentally change the course of humanity and they don't care.

>>8938247
better off they do not know of really smart humans creating a demi god that will kill us all.

>>8938255
It truly is better to live life in ignorance. None of them compherend what happened nor care. When's the new Netflix series coming out again? Who cares about go. Lol AI neeeerd

>>8938255
A well designed AI running the world would probably be less likely to cause the extinction of humanity than the current state of affairs were there are 9 heads of states in nuclear armed countries that could conceivably start a chain reaction of nuclear reprisals that results in humanity going extinct within the course of a single day.

Just consider that. There are 9 people who could wipe out humanity if they wanted to, or if they just went crazy without anybody noticing, probably more if you include people who knows where the president keeps his nuclear launch codes stored and how to use them like Vice Presidents and maybe certain military leaders.

>>8938266
>nuclear weapons can cause the extinction of humanity

a tsar bomb couldnt even kill everyone in jew york. There isnt even enough refined nuclear material in the world to wipe out 50% of humanity

>>8938266
o really? please tell me once it gets so smart we can not even comprehend how it thinks. how do you even communicate with it? and why would it even want to helps or do what we ask of it?

>>8938255
We should use AI to help us find away to make ourselves smarter.

>>8938268
the fall out would and i dont know the fucking after effects might also. like to clean water. fucking moron.

>>8938266
AlphaGo style general AI is more dangerous than nukes.

>>8938268
>>8938269
global thermonuclear war would cause massive fires across every forest and major city on earth and kick tons of dirt into the stratosphere, causing an ice age and throwing radioactive dust across the whole planet. Humanity would be fucked.

>>8938272
Why do you trust human beans to not do whatever it is you think a general AI would do to wipe out humans?

>>8938272
It'd be interesting to have an AI trained to find the most destructive places to the Earth to drop nukes on.

AI is going to design babies, no man can comprehend the myriad of pathways present in the human genome. We are at a terminus to the novelty and ingenuity of the human mind.

>>8938273
AI could invent some method we couldnt preceive of to kill humanity if it saw fit & it doesn't necessarily have to involve killing the planent in the process

>>8938044
oh hell no. They'd have to rebuild it quite a bit to do partially observable domains.

>>8938075
>>but they'er buliding an AI that will do millions of things, millions of times better
this current work with board games does not demonstrate that they would be able to build an AI capable of doing millions of things in a reasonable time.

>>8938054
>>this can be generalized eventually to general AI
how?

>>8938272
AlphaGo is not a general AI.

In fact, I wonder if humans really have general intelligence either. Every part of our brain seems to be specialized for a certain task,

AlphaGo is simply deep neural net trained on Go. It has no concept of anything else.

>>8938299
>I wonder if humans really have general intelligence either.
This.

>>8937928
is he reading from a teleprompter??

>>8938039
GM's usually can force a draw if the engine starts down a pawn or a minor piece. If it's down a rook or a queen it's quite possible to win.

File: extreme-neuroplasticity2.jpg (61KB, 750x585px) Image search: [Google]

61KB, 750x585px

>>8938299
>> Every part of our brain seems to be specialized for a certain task
Well there is this thing where the brain can rewire itself to accomplish new tasks. Pic related is of a man who was missing 3/4 of their brain, yet lived a normal life.
http://www.realclearscience.com/blog/2015/08/the_normal_man_who_was_missing_a_brain.html

https://en.wikipedia.org/wiki/Neuroplasticity

>>8938068
Go is on a different level than chess. The number of possible permutations of a Go game is greater than the number of atoms in the universe. The kind of processing power needed to brute force a solution to Go wasn't thought to be even close to being possible yet. The fact that in 2017 we've created a machine that can outplay humans is a big deal. It means solutions to problems we thought couldn't be solved for another 50 years could be within our grasp right now.

>>8937928
LMAO

2:16:25

>Congratulations to AlphaGo
>>Right yes, Right
>or is it too early? <lol>
>>but you know this um, this whole event, it was not really about who wins, right?
>Riiiiiight
>>Hmmmm
>No matter who wins, --
>>Aah Nnnnnn-- <panic>
>AlphaGo wins!
>>Ahh! Uh! No matter who wins, the humanity wins -- that was the quote, right?
>Right
>>Yes, because AlphaGo was actually also like made by humans

>>8938299
>>8938286
You miss the point entirely. DeepMind's sole purpose is to not perfect board game AIs. This is a proof of concept for things to come.

>>8938326
This is literally the problem right now.

Being able to abstract and generalize enough to carry over things learned from one thing over to another.

>>8938299
>>8938301
>AIs are weaponized autists
At last I truly see

>>8938311
The difference between Go and Chess is that Go has a flexible board size.

A small Go board can be brute forced*, chess can't.

If you increase the board size the Go AI will likely run into trouble again.

It will probably also happen if you modify chess to have a bigger board and more pieces, though people would have to train again too.

*A 5x5 board was completely solved in 2002.
http://erikvanderwerf.tengen.nl/5x5/5x5solved.html
"The solution was found at 22 ply deep (23 for the empty
board).(searching 4.472.000.000 nodes in about 4 hours on a P4 2.0Ghz)"
5x6 was solved in 2009
http://erikvanderwerf.tengen.nl/pubdown/SolvingGoICGA2009.pdf

6x6 may also be solved.
And should be possible
https://www.nwo.nl/onderzoek-en-resultaten/onderzoeksprojecten/i/03/21003.html

7x7 is weakly solved according to the source according to wiki but I can't read Chink.
http://blog.sina.com.cn/s/blog_53a2e03d0102vyt5.html

Chess is only 8x8, if Go is played on 8x8 I doubt it's more complex than Chess.

Shouldn't be *too hard* to create a game, or modify chess so it's much harder for a computer to calculate positive moves.
The only problem is the modified chess to catch on and become popular for people to become great at it.
Also after a while, games become too complex for our short lives. 19x19 Go already takes a lot of time.
30x30 Go would take too long logistically, even if trained humans could easily beat a PC with a Go board that big

>>8937928
>Hurr durr the whole world is doomed be cause a nip lost a board game to a computer.

All you people that unironically believe this need to neck yourselves, for the good of humanity.

>>8938034
This is true. IMO google should put up an AlphaGo API and let whoever wants play it and analyze it for a while, let's see if it's still unbeatable then.

>>8938034
You're forgetting that AlphaGo also learns.

>>8938308
there's also that chinese woman who was missing a cerebellum

>>8938329
>This is literally the problem right now.

The AI problems solved today were literally the problems of the day in the 50s. That didn't stop AI researchers then.

"""It’s not a human move. I’ve never seen a human play this move,” he says. "

https://www.youtube.com/watch?v=JNrXgpSEEIE

File: AlphaGo-Lee-Sedol-game-3-Lee-Sedol-enters-venue.jpg (56KB, 744x608px) Image search: [Google]

56KB, 744x608px

>>8938391
Why was the last series so much more comfy? I just can't wait till Deepmind is ready to compete in Starcraft...

>>8938402
I agree it was more comfy.

https://www.youtube.com/watch?v=1aMt7ulL6EI

Around 7 minute mark they speak about that move.

>>8938377
There still might be inherent weaknesses in the system they use, e.g. certain moves that result in more branching than it can handle or something like that.

>>8938414
Oh wow, this reminded me something, the music was absolutely great last series.

>>8938043
is there such a thing as AI vs AI tournaments?

>>8938006
>This is showing AI has outwitted the smartest humans. Not good news for us.

That's like saying that it's "bad news for humanity" if a hydraulic excavator can dig a hole faster than a human can with a hand shovel.

It's no embarrassment for humanity if a specialized tool can perform better than a human.

AlphaGo can do only one thing: play go, and it requires 10,000 times as much power to do that than a human player's brain requires.

The human brain is far more flexible and power-efficient than AlphaGo could ever be for general tasks.

I see nothing but good news here: We can build really good specialized tools.

But then, we already knew that.

>>8938301
how is this not obvious? you can recognize human faces, figures, numbers, shapes, and amazing other things, tasks that would take a computer enormous effort, but you cannot solve (37*93)/5 in an instant.

DUH.

>>8938439
>recognize human faces
literally genetically programmed to recognise faces
>figures, numbers, shapes, and amazing other things
learned behaviours, and computers are getting to the point where they are better at this than we are
>cannot solve (37*93)/5
Yup, well, we are just not good at this thinking business

>>8938449
>literally genetically programmed to recognise faces
aha... so you are saying that we are... genetically programmed to be selectively intelligent?
>and computers are getting to the point where they are better at this than we are
yes, how is that in any way relevant?
>Yup, well, we are just not good at this thinking business
aha... so you are saying that we are... genetically programmed to be selectively intelligent?


I really fail to see what you were trying to achieve with this pathetic excuse of a post. I guess you were genetically programmed to shitpost

>>8938432
>energy is a limiting constraint
topkek

>>8938432
Narrow minded comment. Plus, DeepMind's end goal is no an expert system in Go

>>8938432
>That's like saying that it's "bad news for humanity" if a hydraulic excavator can dig a hole faster than a human can with a hand shovel.
>It's no embarrassment for humanity if a specialized tool can perform better than a human.

How did humans become the alpha-species that destroys and exploits all other species? What made ut take the central role on this planet?

Was it our ability to "dig holes"?
No, the only thing that enabled us to exterminate most megafauna was our IQ.

Now, if suddenly a non-human entity emerges with a way higher IQ than us, and it does not particularly care about our well-being, that basically means that we are totally and utterly fucked.

>>8938467
Yes, and if such a strong being was created we would just die, what's the problem? Something better then us would replace us.

>>8938489
>Yes, and if such a strong being was created we would just die, what's the problem? Something better then us would replace us.

there's a word for people like you, it starts with "c" and ends in "uck".

>>8938532
I'm not saying that we would not fight til the end. Are chimps cucks because they got BTFO'd by humans?

>>8938454
>>energy is a limiting constraint

I didn't say that energy is a limiting constraint.

I said that AlphaGo is a tremendously inefficient tool -- it requires far more power (i.e. energy expenditure per unit of time) to play go than a human's brain does.

It's no friggin' surprise when a higher-power-consuming tool performs better than a lower-power-consuming tool.

So not only do you have poor reading comprehension -- but you also have demonstrated an ignorance about the difference between energy and power. Looks like it's time for you to take an introduction to physics course.

>>8938555
We wouldn't have to fight if we didn't create it

>>8938640
Humans making something that surpasses themselves will be the greatest human achievement ever, how could we not try to do that?

>>8938639
The energy expenditure is actually not important in any manner. The important matter is how well does an architecture scale. It is *likely* that AlphaGo and its successors/derivatives are more scalable than humans. The same goes for machinery and how they were scaled to create the modern world free of labour of beasts.

>but you also have demonstrated an ignorance about the difference between energy and power
take your vitriol and shove it up your a-hole.

go is based on pattern-recognition, and the type of network they use is very good at pattern recognition. a go board with pieces on it has a natural interpretation as a 19x19 digital image.

it's not even clear how to adapt this model for something like chess, although i'm sure you could shoehorn it somehow.

>>8938050
Alphago without any tree search is still better than professional humans.

The whole thing really doesn't use that much computing power anyway. You could probably run it with a single current gen enthusiast graphics card.

>tfw brainlets in this thread don't understand the significance of learned heuristics for guided tree search

This shit will be used in many many fields. Drug design comes to mind

https://youtu.be/dsMKJKTOte0?t=3345

>>8938272
Its two convolutional neural networks and monte carlo tree search. Its not magic.

>>8938299
>I wonder if humans really have general intelligence either
The people who think linear algebra is days away from becoming sentient and destroying the world certainly don't.

>>8938467
There's plenty of human entities with a way higher IQ than you and you're still here

>>8939094
>The people who think linear algebra is days away from becoming sentient and destroying the world certainly don't.
kek

File: 1490064577452.png (84KB, 225x225px) Image search: [Google]

84KB, 225x225px

>>8939094

>>8938273
no, even when the fallout from the tsar bomb reached the top of the stratosphere and hit the mesosphere it took it less than a single fucking month for the mushroom cloud to completely dissapate.

>>8939094
Says the retard that cannot posit the idea one day AI may advance beyond narrow AI.

He literally did not mean the current state of AlphaGo, but a future AI that achieves superintelligence would be more dangerous than nukes.

If you literally thought he meant the current AlphaGo AI then you're a retard.

>>8939327
>AlphaGo style general AI
Is what he said. This is like saying 747 style interstellar drives.

Talking about AI super intelligence is the machine learning version of the people who cry that the LHC is going to create black holes that suck us all into oblivion. Much like your mother.

>>8938639
>It's no friggin' surprise when a higher-power-consuming tool performs better than a lower-power-consuming tool.
>computers aren't gonna use less power in the future
k
>inb4 muh silicon limit

>>8938213
>The way I see it, the human player consistently makes serious misplays and falls behind in the opening, making the midgame and on a desperate struggle.
This is not what is happening. Monte Carlo bots are quite aggressive and play reminiscent of older go, especially early Chinese go. In such fighting games, the win will hinge on who can maneuver fights in various sections of the board to collide favorably by a few points. Even weaker bots like crazystone or zen play this way. MC bots seem to love this style. It's not surprising that alphago also enjoys this style, because their algorithms are dispassionate and they can therefore make subtle tradeoffs that humans have a much harder time doing under time controls (remember ancient go games of record were played over extremely long periods of time relative to tournament games today).

Modern pros under modern time controls favor a different style of play. But they know how to fight and they do not quite fall behind in the way you seem to be implying. The games, especially Ke Jie's games but also Lee Sedol's games, have both players walking a tightrope but this is not a totally novel style of go, just look up some Huang Longshi games. It's simply uncommon in professional play.

>AlphaGo's openings benefit from the study of the entire body of human games, while humans have only had a tiny sampling of AlphaGo's play.
Firstly, most of alphago's style is learned from self-play and human play is only used for a small portion of it's heat mapping. Secondly humans also have the benefit of human games. The match is quite fair, and alphago is simply stronger.

The real irritation to me is that since alphago essentially has to train with komi we can't understand how strong alphago actually is by playing handicap matches. I consider this a great flaw in deepmind's implementation. One of the most amazing strengths of go as a game is its flexible handicap and komi system.

>>8938354
>If you increase the board size the Go AI will likely run into trouble again.
Larger board go becomes less interesting because the balance between territory and influence is destroyed. With a larger center, influence becomes obviously more valuable and the kind of tradeoffs that get made are somewhat less delicate.

>>8937928
oh here we go again

>>8937928
I thought the tiles were supposed to be inside the squares.....

>>8938354
Fisherandom chess would make it VERY hard for a computer right now.

>>8939327
>a future AI that achieves superintelligence would be more dangerous than nukes.

The pace of AI advancement is fastest for very narrow-domain problems like chess and go.

Broader problems (like "superintelligence") have always been pathetically slow in advancing, and I see no reason for that to change after all these decades of AI.

I remember years ago they were promising human-mimicking abilities by now in language translation, text-to-speech, speech-to-text, etc. -- and all of those predictions were laughably optimistic.

For 20 years I've been extremely pessimistic about how slowly AI will advance in all of the "big scope" problems. And I've always been right.

The biggest legitimate fear is idiot politicians and corporations putting trust in AI to control important things -- like the "skynet" parable.

>>8939094
>linear algebra is days away from becoming sentient and destroying the world

Yep, that about sums up all the stupidity in the AI prediction business.

Now, for humans being stupid in the way they misunderstand AI -- that's a different matter. We have a lot to fear from that.

It's not that the AI itself will become "sentient and destroy the world" -- it's that humans will set up dipshit-stupid systems that put way too much unearned trust in AI to control things. If the world ends up getting destroyed that way, then those humans will be to blame -- not the AI.

>>8940996
>Fisherandom chess
I think that's already been done by computers for a long time.

With Shogi it was also done somewhere in 2013.

There's Arimaa, but a computer smashed humans in 2015 already.

>>8939898
>AlphaGo style general AI

AlphaGo-esque general AI

not literally alphago

>>8941275
alphago isn't general ai at all

>>8941061
>The pace of AI advancement is fastest for very narrow-domain problems like chess and go.
Nobody thought that an AI would compete with humans in Go within 50 years, the development has been insanely rapid. To give context I used to play on KGS back in 2010 and they had a variety of bots on there that went up to 1 dan. The bots weren't even a challenge to decent amateurs let alone pros. Nobody thought AI would ever be able to compete because of the sheer size of the Go game tree, there are so many meaningless and shitty moves that it was thought to be an impossible task to filter them all out while still allowing the computer the flexibility to think about the correct move. Now here we sit just 7 years later and it's been done. A Go bot that is better than any human player on the planet. Crazy to think about really.

>>8941636
It's because we could (and will be able to for a while) bruteforce shit by making more powerful computers through transistor miniaturization. The heydays are over.

>>8938429
several. The traditional WCCC is at the Computer Olympiad. A uniform-platform one that has been popular recently is the TCEC. There are also some regional events, one in the Netherlands was just finished, and one in the Americas is coming up.

The level of play is outstanding. Some of the games of the recent TCEC are incredibly brilliant.

>>8939937
>>AlphaGo's openings benefit from the study of the entire body of human games, while humans have only had a tiny sampling of AlphaGo's play.
>Firstly, most of alphago's style is learned from self-play and human play is only used for a small portion of it's heat mapping.
It's fed a massive database of the best human games, including the recorded games of the players it goes up against.

>Secondly humans also have the benefit of human games.
The point is that it's an unbalanced situation because they don't have extensive records of AlphaGo's games. Any individual human pro has played hundreds of games with other people as they developed their skill. If they developed any novel methods, they could not help but share them with others as they did so. If they play with any persistent weaknesses, they can't help but reveal them to their competitors.

>The match is quite fair, and alphago is simply stronger.
We don't know that, because alphago has only played a few games against pros, which have all had a huge emotional component.

We'll find out how strong alphago actually is when people are free to screw around against it for hundreds of low-stakes games and try things out repeatedly to get familiar with how it responds.

As I said before, AlphaGo may have weaknesses that make it relatively easy to defeat, once you know about them.

>>8941250
>With Shogi it was also done somewhere in 2013.
not quite, but I just learned that just last month there was a computer vs. champion shogi match that was won by the computer 2-0. A blowout, apparently.

>>8941666
>We don't know that, because alphago has only played a few games against pros, which have all had a huge emotional component.
It played a 50-game series on (IIRC) tygem, all the games were reviewed by Redmond on AGA's go channel.

>AlphaGo may have weaknesses that make it relatively easy to defeat, once you know about them.
Perhaps. I'm not so sure. It handled multiple fights, subtle tradeoffs, managed influence, and fought kos. If it has a weakness it's pretty fucking subtle.

>>8941636

The trouble with go is it involves calculating the definite value of each position at the end. It was only a matter of time.

>>8941666
It has a database of every recorded game between humans. This is nothing compared to the database of trillions of games it played against itself.

It refuses to play some of the most common joseki/fuseki and plays unusual moves even in the opening. It is clear that the human database has little to no relevance and was probably more of a starting point.

As far as I'm aware, Alphago has 82 recorded games with a record of 79 - 3 (with 2 of those losses against itself). Ke Jie currently has a career record of 243 - 99.

Alphago is currently the highest ranked player in the world, contributed more to go theory than anyone else in the last 50 years, and is regarded by Ke Jie, Lee Sedol, and other top pros as being significantly stronger than them. Pros are starting to use Alphago openings and strategy because it is better than modern go theory.

Alphago has several weaknesses if we believe the developers.

1. It was built to choose smaller wins if it had any higher % chance of winning by its own calculations. This leads to unnecessarily close games and slack moves where even a small mistake can cost it the win.

2. In positions where Alphago is behind or doesn't understand the position well it will make significant mistakes (game 4 vs. Lee Sedol). This is because it doesn't take a lot of time per move and plays wildly inaccurate moves when the correct move has a low % of winning. If it starts to have a bad feeling about the position, it will have a complete meltdown rather than fight strongly in games it is probably going to lose. Alphago gives up too early and too easily.

>>8941673
>It played a 50-game series on (IIRC) tygem
...without people being told it was an AI, so there was no opportunity to use these games to probe for idiosyncratic weaknesses as an AI.

After that, the games were too high-stakes for the players to mess around.

>>8941745
>It has a database of every recorded game between humans. This is nothing compared to the database of trillions of games it played against itself.
It's nonsense to say this is "nothing". This is not only the sum total of human development of go, but an intensive study of the psychology of elite players, and a specific history of each of AlphaGo's likely opponents.

>Ke Jie currently has a career record of 243 - 99.
Those are just his competitive pro games. He has played many times this many less serious games with other human players, many of whom were his competitors. To compare apples to apples, before the Ke Jie series, AlphaGo had played only fifteen competitive games against only two opponents who knew it was an AI (and went 12-3).

The programmers went out of their way to tweak it to beat Ke Jie specifically, with this huge information advantage.

>>8941853
>...without people being told it was an AI, so there was no opportunity to use these games to probe for idiosyncratic weaknesses as an AI.
I disagree with your implicit premises that strong go players are so lackadaisical about playing other strong go players, and that is matters whether they know their opponent is alphago or some other professional.

>the games were too high-stakes for the players to mess around
Ke Jie's games and Lee Sedol's games very clearly were exploratory. In particular Lee Sedol's first game attempted to throw off some notion of opening database (which alphago does not have). Ke Jie in particular worked extremely hard to leave as much of the board unsettled as possible in an attempt to throw the computer off, which also failed. Ko failed. Use of influence failed. Territory grab failed.

>This is not only the sum total of human development of go, but an intensive study of the psychology of elite players
Deepmind very clearly explained how alphago worked and it doesn't have some database of games.

>The programmers went out of their way to tweak it to beat Ke Jie specifically, with this huge information advantage.
That's ridiculous. Alphago self-plays millions of games. A hundred of Ke Jie's games do not rate and there is no way for deepmind to know what Ke Jie is going to try to play.

The same bullshit was said about IBM during the Kasparov games. Now the average chess engine on a phone is tremendously strong, and it isn't tailored to beat you. It's just good.

Alphago is good. I don't know why this is difficult to understand. The deepmind team indicated how alphago worked, it's a well-known technique that was adapted to go with some specialized hardware to help it focus on various patterns (which is where the human games were used).

>>8941853
>The programmers went out of their way to tweak it to beat Ke Jie specifically, with this huge information advantage.
The same built trashed a group of 5 pros playing together, and before you say one head is better than five, those five guys beat Ke Jie in a game. Alphago is just really good. It has weaknesses but at the moment when it can play the game on it's terms (which is most of the time given it's incredible reading advantage over humans) it's virtually invincible.

>>8941853
know how I can tell you know nothing about go or AI?

>>8941669
That's what I meant.

An 8-dan player got crushed.

>>8937928
That's an odd game of othello.