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

Veeky Forums let's watch AI kill humanity.

youtube.com/watch?v=ru0E7N0-kFE

Other urls found in this thread:

youtube.com/watch?v=xZYQafx3pIQ
twitter.com/demishassabis/status/867584056095002624
realclearscience.com/blog/2015/08/the_normal_man_who_was_missing_a_brain.html
en.wikipedia.org/wiki/Neuroplasticity
erikvanderwerf.tengen.nl/5x5/5x5solved.html
erikvanderwerf.tengen.nl/pubdown/SolvingGoICGA2009.pdf
nwo.nl/onderzoek-en-resultaten/onderzoeksprojecten/i/03/21003.html
blog.sina.com.cn/s/blog_53a2e03d0102vyt5.html
youtube.com/watch?v=JNrXgpSEEIE
youtube.com/watch?v=1aMt7ulL6EI
youtu.be/dsMKJKTOte0?t=3345
twitter.com/SFWRedditGifs

GO HUMANS GO

/pun not intended

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

Agreed. Who is winning? I can't tell

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.

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

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.

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.

>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.

>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.

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

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.

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

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.

youtube.com/watch?v=xZYQafx3pIQ

>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.

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

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.

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

>free will

>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.

twitter.com/demishassabis/status/867584056095002624

Already happened. Google the history of the last few days

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.

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

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

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.

>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

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?

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

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

AlphaGo style general AI is more dangerous than nukes.

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.

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

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.

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

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

>>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.

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

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.

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

is he reading from a teleprompter??

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.

>> 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.
realclearscience.com/blog/2015/08/the_normal_man_who_was_missing_a_brain.html

en.wikipedia.org/wiki/Neuroplasticity

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.

LMAO

2:16:25

>Congratulations to AlphaGo
>>Right yes, Right
>or is it too early?
>>but you know this um, this whole event, it was not really about who wins, right?
>Riiiiiight
>>Hmmmm
>No matter who wins, --
>>Aah Nnnnnn--
>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

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.

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.

>AIs are weaponized autists
At last I truly see

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.
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
erikvanderwerf.tengen.nl/pubdown/SolvingGoICGA2009.pdf

6x6 may also be solved.
And should be possible
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.
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

>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.

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.

You're forgetting that AlphaGo also learns.

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

>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. "

youtube.com/watch?v=JNrXgpSEEIE

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

I agree it was more comfy.

youtube.com/watch?v=1aMt7ulL6EI

Around 7 minute mark they speak about that move.

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.

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

is there such a thing as AI vs AI tournaments?

>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.

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.

>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

>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

>energy is a limiting constraint
topkek

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

>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.

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

>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".

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

>>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.

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

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

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.

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

youtu.be/dsMKJKTOte0?t=3345

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

>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.

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

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

...

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.

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.

>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.

>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

>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.