2:40 "learns how to play in a human like manner" ... "We show it 100,000 game replays ... then it plays itself 30 million times." Yes, that's how I learn things. I look at 100,000 examples of how to do it. Then I try it myself 30 million times. Works like a champ. :P
Hey.... I recognise that place! It's Four Quarters in Peckham! :D .... :/ Did the BBC report mention where they were filming? I don't recall the name of the place being mentioned.
It does not matter how you draw your go board. There is no difference between playing the game of Go on grid or squares as long as the coordinates are the same. Think about it.
+Maquintoshuelo You are correct. Even tho any board size is possible for Go (including 18x18), standard coordinate system is 19x19, which they in animation probably wanted to use. But I mean, in principle, having 19x19 square board and playing stones inside the boxes means the same game as playing on cross-sections of 19x19 lines (which is more traditional). Same thing, you can play chess on 8x8 grid (unusual although possible). Therefore no big deal with the "low quality, wrong animation".
This is definitely breakthrough technology. But, we have an epic journey ahead still. The reality check however, is going to show up in real world applications of this.
It is funny that they imagine all sorts of uses and mention many of them. But they carefully don't mention economy, the one area that really needs greed free analysis.
+trefod There's a short series of videos on Computerphile where the host talks about the dangers of naive artificial intelligence. It's one of those scenarios where we really can't afford to make mistakes.
+Salafrance I've seen it, but I think the danger in this case is not the computer. It is those people that benefit greatly from not having oversigt of any kind.
Wow BBC this video is wrong in alot of ways my god first and foremost alphago hardware require much more then a single labtop running in asynchronous or distributed mode. Actual specs 48 CPU & 8 gpu
I just came here to comment on the board, Im happy to see that others have done that before me. Btw there was an earler video from bbc which was supposed to explain go, and that was a failure as well.
Bill Hutchinson No engine plays that move nowadays and the team claimed it was thinking the queen move first then at the last moment changed to the bishop move.
Bill Hutchinson Well IBM acknowledged it on television in front of millions of people. I don't think you even know about the controversy. IBM won't deny any of this.
So I know nothing about computers, but can somebody tell me why quantum computing couldn't process all of the possibilities, just like DeepBlue, but on a much grander scale?
Because quantum computing is still in it's infancy, it consists mainly of theory and speculation. The only working prototypes that exist to this date can't do much more than simple arithmetics. Also, the fact that there are more possible legal Go positions than atoms in the universe doesn't help.
Healthcare and science, healthcare and science healthcare and science and self driving cars.... No no military or market or media manipulation applications for us just healthcare and science.
+UnicyclDev A program that can learn to play games by itself isn't a breaktrhough? This is a bigger breakthourgh than there has ever been on the field of AI.
Real Ai is all about new situations. And that's the problem. Just look at the double swing setup. There is no algorithm that can predict which way a double swing will go. AlphaGo fakes intuition. Real intuition doesn't look back to decide. Let's run this 5 game sets for a couple of times and you'll see the real go master learn to beat AlphaGo by evolution of it's intuition. And they know that at AlphaGo btw :)
Lol, 500 years + then we can talk "around the corner". This is just probabilistic computation + "evolutionary" selection of data, driven by the logic of the data arrangement. True AI would be a machine which not only solves a problem, but which can *reflect* upon that process and itself. This is mechanistic and algorithmic computation, a program running as programmed, including the probabilistic prospecting part. Without sentience, AI is but more and more complex computation-programs, without a shred of actual "understanding", as in a meta-understanding of what is happening. Not even close.
+Hypatia No its not "a program running as programmed, including the probabilistic prospecting part",. also it taught itself how to play go, so it definitely understands whats going on. maybe you should go read a bit more about neural networks. The only hardcoded parts are the monte carlo search which looks at the different variations of the game after a possible move, but the main problem is how do you determine which move to look the variations for, because you cant possibly look at the variations for every single possible move, this is where the neural nets comes in. the NN part decides which move they should waste computational power on and look at the variations for that move. AlphaGo can choose the best possible moves 57% of the time which they then run the monte carlo algorithm on that move(s) to see the different variations. So its not exactly a hardcoded program with some probability equations sticked at the end. Of course this is not the AI as we think of it buts its pretty damn close, any AI we make in the future is gonna be based of neural nets so its a step in the right direction. Given alpha go's state it should be unbeatable, its loss in game 4 was due to luck by the human player, apart from that, I expect it to be a pretty damn good AI for playing Go.
+Hypatia 500 years is ridiculously too much, come on, just doing the math, in just 30 years we will have thousands of more computing power in a single pc. At that time big enterprises like google can show us the real thing. And if you just need something like sentience, it can happen sooner. It just will not be as clever as you (note that there are people for example that have a condition and have problem visualizing, imagining, but they still live a normal life)
+Hypatia in at least one of his presentations (I think it was 'The Future of Artificial Intelligence'), Demis Hassabis explicitly states that this is not '+ "evolutionary" selection of data'. Instead, it's a neural net system capable of reinforcement learning, which appears to be modelled on something like the dopaminergic mechanism in a natural brain. I'd have made the same assumption had I not watched the presentations, but at one point he mentions the infeasibility of creating a fitness function for Go, contrasting it with a relatively straightforward function for the game of Chess. Watch the video - it's *really* interesting.
Let us look at this without becoming so emotional. They were excited in 1997. We are excited in 2016. But, while computers became faster and with much more memory, they didn't become much smarter. I think that if we look at this objectively there has been less advances in AI than the media would like to make believe us. To the masses self-driving cars, translation machines, computers "understanding what they see" might sound as novel research areas, but they are not. It is at least since the 80' now that we hear about these things with lots of funds spent in neural network learning, genetic algorithms, etc., but they did not meet the expectations. Just consider how machine translation works after 30 years of efforts spent in it from IBM to Google. Even if AlphaGo will win, that would not make much a difference for applications in the real world. I would be much more cautious before giving in to unjustified expectations of a bright future dominated by AI. Or will we become excited again in 2035?
+Pinco Pallino Yes, machines in 2016 are infantile. But the more informed I get during a deep read on this topic the more I realize, they are reaching benchmarks faster than many even predicted. I'd review the subject more if I were you. Yes I'm saying your arguing from ignorance. Yes things are moving very quickly and across many branches of every category of Genetics, Robotics(AI/AGI), Nanotechnology. What does we are developing a refinement and scale-up of atomic-level manufacturing mean to you? Its a fantasy, as flight was once?
0:30 - 0:37 Right away an association as "synonymous" of two radically different questions. And at 5:36, again, this fool who understands nothing of science or AlphaGo enters the stage to lecture us on how we are supposed to *utterly falsely* interpret this. Yes, certainly, AlphaGo constitutes pivotal progress in automatic learning and consequent decision making, but as I mention in my own video on "artificial intelligence" *nothing in AlphaGo indicates any "thinking" or "artificial intelligence". 1. Everything intelligent about AlphaGo, that is its **_way_** of learning, was put in there by humans (the computer only supplies the calculating power - same story since Turing and it will never change) 2. no matter how much these systems progress, they will never develop inteligence in the sense of a mind - they will never **_know_** the solutions they find*.
+Soteriologe The algorithm they're using, no matter how it's interpreted, is essentially just error optimization. Which is obviously outside our traditional definition of "intelligence". The issue is, we understand so little about the brain at a foundational level that we cannot determine an appropriate definition of what could be considered intelligence anyway. Though a few interesting theories, such as predictive processing, are popping up recently suggesting that our brain may essentially just be an error optimization machine. More than likely these theories will be proven wrong, or something more fashionable will come along, but it seems like a decent theory to me. The point I'm trying to get across is don't say never, when you (or anyone on the planet) don't understand how intelligence even emerges.
+Soteriologe I completely disagree with you on this. Your 'never' approach to logic is via history alone rather broken. These machine learning systems are infants in a vast world. So were we all. A baby has to learn how to eat, crawl, talk, walk, swim, jump and eventually muse over artificial intelligence. And they come with a brain far beyond the point we are at right now technologically. These limited capacities are but a tiny beginning of something incredible. Never say never, as its an instant fail to recognize that this, is just a beginning. Early attempts at flight, were disasters, and even cute and silly looking. We do not laugh about it now do we? Yes, we even went to the moon and now we have machines working on mars and beyond. Next.
+mastertheillusion How I arrive at the "never" is not something I discussed here. That doesn't mean I have no reason for that. One of them is I tend to count myself as a Lucasian regarding Gödel's incompleteness theorems. Another is that this silly propaganda - _totally designed to obfuscate the befuddlement of the scientific community regarding "intelligence" and mixing it purely rhetorically which subjective terminology which is completely unwarranted_ - has been repeated for decades, along with the ever same promises that we are ALWAYS almost there; with promises receding ever further into the future. And of course, while this racket is sold with health concerns, this is mostly going to benefit the ever increasing domination of the people by an ever smaller class of dominators - robots controlling the population. But there are still other reasons that you don't quite sound like you're worth debating, for instance once someone is prepared to assume _as obvious_ that not only certain ways of mental functioning and capacity, but basic qualia themselves would be an emergent property. *But look: I am all for decisive advances in these areas as I stated in my OP, IF (!) these advances really would serve the good of the community and help destroy the dominator culture, but what I am against is this propaganda as well as to philosophically falsely call it "artificial intelligence". Just call it artificial "expert systems" or "decision making" and it's fine with me.*
+Soteriologe good point, but in some sense, aren't we all merely machines that did well at error optimization? After all, in the real world, making too many mistakes could adversely affect one's ability to pass on their set of genes and so forth. Still, I find this development an exciting milestone on the path toward true AI. This is to date one of the best systems built by humans that can exhibit early patterns of learning. Baby steps in a way, but eventually all babies grow up to become intelligent adults (with the exception of Trump..)
A lot of hype... AI still has a huge way to go. Anybody with a small machine learning background knows that this is another "super specialized" system and that's it.
AlphaGo can't even put a stone on the grid, according to BBC.
At 5:10, Go stones are played on the grid lines, not inside the boxes.
+cisy exactly, and my OCD went apeshit.
+cisy I would put the blame on the animator. They didn't know how the game works and were given inadequate instructions.
Yeah I died a little inside...
It’s like a 5 second google search for the animator.
Xaxaxa
One day an AI will watch this
It already does. How do you think RUclips auto-flags videos and detects things like which game you're playing, etc.
Τοmorrow Al is carefuuly watching
AI will watch this and comment on why it's algo recommended it..
2:40 "learns how to play in a human like manner" ... "We show it 100,000 game replays ... then it plays itself 30 million times." Yes, that's how I learn things. I look at 100,000 examples of how to do it. Then I try it myself 30 million times. Works like a champ. :P
>placing the stones on the squares and not on the intersections
+Patralgan just bbc things
I'm still triggered
why is everyone so mean about this video??
What is this stone placement at 5:10 ? Where is the respect ?
용산고 손
손
ᄋ ᄋ 손
Hey.... I recognise that place! It's Four Quarters in Peckham! :D
....
:/
Did the BBC report mention where they were filming? I don't recall the name of the place being mentioned.
the scientist had moderate tones, but the narrator add a bit of hype. I wonder if its intended to promote the event or he actually meant it?
It does not matter how you draw your go board. There is no difference between playing the game of Go on grid or squares as long as the coordinates are the same. Think about it.
+Štěpán Doan The difference is just one row and one column. 18x18
+Maquintoshuelo You are correct. Even tho any board size is possible for Go (including 18x18), standard coordinate system is 19x19, which they in animation probably wanted to use.
But I mean, in principle, having 19x19 square board and playing stones inside the boxes means the same game as playing on cross-sections of 19x19 lines (which is more traditional). Same thing, you can play chess on 8x8 grid (unusual although possible). Therefore no big deal with the "low quality, wrong animation".
4:1 to AlphaGo that's overkill!
+Jeffrey Tse I was wondering what the score was.
This is definitely breakthrough technology. But, we have an epic journey ahead still.
The reality check however, is going to show up in real world applications of this.
Curse you, Alphago!
- A student who suffers from this video because it is a school listening test
lmao
we have a high quality a.i. war machine, general. This is a major technological breakthrough, so what should we do with it?
general: make it play go
용산고인사람이 잇을거시여..
유비 유비 ㅠㅠ
유비 유비 저요
It is funny that they imagine all sorts of uses and mention many of them. But they carefully don't mention economy, the one area that really needs greed free analysis.
+trefod There's a short series of videos on Computerphile where the host talks about the dangers of naive artificial intelligence. It's one of those scenarios where we really can't afford to make mistakes.
+Salafrance I've seen it, but I think the danger in this case is not the computer. It is those people that benefit greatly from not having oversigt of any kind.
Wow BBC this video is wrong in alot of ways my god first and foremost alphago hardware require much more then a single labtop running in asynchronous or distributed mode. Actual specs 48 CPU & 8 gpu
I just came here to comment on the board, Im happy to see that others have done that before me. Btw there was an earler video from bbc which was supposed to explain go, and that was a failure as well.
this guy worked on theme park with peter molyneux
Whats the song at 2:19?
Apache (Jump On It) by The Sugarhill Gang
What is the song at the beginning? Does anyone know?
They didn't master chess in the 90's. That game was clearly rigged and they had five grandmasters standing in the back room.
Source?
Bill Hutchinson No engine plays that move nowadays and the team claimed it was thinking the queen move first then at the last moment changed to the bishop move.
Source?
Bill Hutchinson Well IBM acknowledged it on television in front of millions of people. I don't think you even know about the controversy. IBM won't deny any of this.
That's why I'm asking for a source. I'd like to look into it.
Alphago didn't play the atari games, that was another AI!
Go is an atari game.
"atoms in the universe"?
이거 듣기평가다 시불알
Novak Djokovic is better than Roger.
So I know nothing about computers, but can somebody tell me why quantum computing couldn't process all of the possibilities, just like DeepBlue, but on a much grander scale?
Because quantum computing is still in it's infancy, it consists mainly of theory and speculation. The only working prototypes that exist to this date can't do much more than simple arithmetics. Also, the fact that there are more possible legal Go positions than atoms in the universe doesn't help.
there are more possibilities than there are atoms in the universe.
Healthcare and science, healthcare and science healthcare and science and self driving cars.... No no military or market or media manipulation applications for us just healthcare and science.
Alpha go isnt even close to being an AGI wtf is wrong with the bbc?
hyped over speculation. AlphaGo is a amazing in its own right, but it isn't a historical turning point in AI. BBC sounds like Buzzfeed in this video.
+UnicyclDev A program that can learn to play games by itself isn't a breaktrhough? This is a bigger breakthourgh than there has ever been on the field of AI.
So it is now brute force but with a *.txt.
Same thing.
Real Ai is all about new situations. And that's the problem. Just look at the double swing setup. There is no algorithm that can predict which way a double swing will go. AlphaGo fakes intuition. Real intuition doesn't look back to decide. Let's run this 5 game sets for a couple of times and you'll see the real go master learn to beat AlphaGo by evolution of it's intuition. And they know that at AlphaGo btw :)
When AI understand that humans are morons they will use us as slaves.
imagine, you are smarter then pig right, would you let the pig to command you :D
Meruem could beat it.
Comp Chomp did it better.
Lol, 500 years + then we can talk "around the corner". This is just probabilistic computation + "evolutionary" selection of data, driven by the logic of the data arrangement. True AI would be a machine which not only solves a problem, but which can *reflect* upon that process and itself. This is mechanistic and algorithmic computation, a program running as programmed, including the probabilistic prospecting part. Without sentience, AI is but more and more complex computation-programs, without a shred of actual "understanding", as in a meta-understanding of what is happening. Not even close.
+Hypatia Wrong, AlphaGO already reflects its actions. AlphaGo is a neural network, not a decision tree.
+Hypatia No its not "a program running as programmed, including the probabilistic prospecting part",. also it taught itself how to play go, so it definitely understands whats going on. maybe you should go read a bit more about neural networks. The only hardcoded parts are the monte carlo search which looks at the different variations of the game after a possible move, but the main problem is how do you determine which move to look the variations for, because you cant possibly look at the variations for every single possible move, this is where the neural nets comes in. the NN part decides which move they should waste computational power on and look at the variations for that move. AlphaGo can choose the best possible moves 57% of the time which they then run the monte carlo algorithm on that move(s) to see the different variations. So its not exactly a hardcoded program with some probability equations sticked at the end. Of course this is not the AI as we think of it buts its pretty damn close, any AI we make in the future is gonna be based of neural nets so its a step in the right direction. Given alpha go's state it should be unbeatable, its loss in game 4 was due to luck by the human player, apart from that, I expect it to be a pretty damn good AI for playing Go.
+Hypatia 500 years is ridiculously too much, come on, just doing the math, in just 30 years we will have thousands of more computing power in a single pc. At that time big enterprises like google can show us the real thing.
And if you just need something like sentience, it can happen sooner. It just will not be as clever as you (note that there are people for example that have a condition and have problem visualizing, imagining, but they still live a normal life)
ritsukasa Lets hope so:)
+Hypatia in at least one of his presentations (I think it was 'The Future of Artificial Intelligence'), Demis Hassabis explicitly states that this is not '+ "evolutionary" selection of data'.
Instead, it's a neural net system capable of reinforcement learning, which appears to be modelled on something like the dopaminergic mechanism in a natural brain.
I'd have made the same assumption had I not watched the presentations, but at one point he mentions the infeasibility of creating a fitness function for Go, contrasting it with a relatively straightforward function for the game of Chess.
Watch the video - it's *really* interesting.
Skynet is born!
WTF is that board? All credibility destroyed in a second.
Let us look at this without becoming so emotional. They were excited in 1997. We are excited in 2016. But, while computers became faster and with much more memory, they didn't become much smarter. I think that if we look at this objectively there has been less advances in AI than the media would like to make believe us. To the masses self-driving cars, translation machines, computers "understanding what they see" might sound as novel research areas, but they are not. It is at least since the 80' now that we hear about these things with lots of funds spent in neural network learning, genetic algorithms, etc., but they did not meet the expectations. Just consider how machine translation works after 30 years of efforts spent in it from IBM to Google. Even if AlphaGo will win, that would not make much a difference for applications in the real world. I would be much more cautious before giving in to unjustified expectations of a bright future dominated by AI. Or will we become excited again in 2035?
+Pinco Pallino Yes, machines in 2016 are infantile. But the more informed I get during a deep read on this topic the more I realize, they are reaching benchmarks faster than many even predicted. I'd review the subject more if I were you.
Yes I'm saying your arguing from ignorance. Yes things are moving very quickly and across many branches of every category of Genetics, Robotics(AI/AGI), Nanotechnology. What does we are developing a refinement and scale-up of atomic-level manufacturing mean to you? Its a fantasy, as flight was once?
So, tell me... besides microelectronics and chips, what did nanotechnology change in our daily life?
SHAMWOW DUH!
Your conception of AI is years behind. Do some catching up.
ㅋㅋ
0:30 - 0:37 Right away an association as "synonymous" of two radically different questions. And at 5:36, again, this fool who understands nothing of science or AlphaGo enters the stage to lecture us on how we are supposed to *utterly falsely* interpret this. Yes, certainly, AlphaGo constitutes pivotal progress in automatic learning and consequent decision making, but as I mention in my own video on "artificial intelligence" *nothing in AlphaGo indicates any "thinking" or "artificial intelligence". 1. Everything intelligent about AlphaGo, that is its **_way_** of learning, was put in there by humans (the computer only supplies the calculating power - same story since Turing and it will never change) 2. no matter how much these systems progress, they will never develop inteligence in the sense of a mind - they will never **_know_** the solutions they find*.
+Soteriologe
The algorithm they're using, no matter how it's interpreted, is essentially just error optimization. Which is obviously outside our traditional definition of "intelligence".
The issue is, we understand so little about the brain at a foundational level that we cannot determine an appropriate definition of what could be considered intelligence anyway. Though a few interesting theories, such as predictive processing, are popping up recently suggesting that our brain may essentially just be an error optimization machine.
More than likely these theories will be proven wrong, or something more fashionable will come along, but it seems like a decent theory to me.
The point I'm trying to get across is don't say never, when you (or anyone on the planet) don't understand how intelligence even emerges.
+Brad Butcher
But I agree, that guy has no idea what the fuck he's talking about
+Soteriologe I completely disagree with you on this. Your 'never' approach to logic is via history alone rather broken.
These machine learning systems are infants in a vast world. So were we all. A baby has to learn how to eat, crawl, talk, walk, swim, jump and eventually muse over artificial intelligence. And they come with a brain far beyond the point we are at right now technologically.
These limited capacities are but a tiny beginning of something incredible. Never say never, as its an instant fail to recognize that this, is just a beginning.
Early attempts at flight, were disasters, and even cute and silly looking.
We do not laugh about it now do we?
Yes, we even went to the moon and now we have machines working on mars and beyond.
Next.
+mastertheillusion How I arrive at the "never" is not something I discussed here. That doesn't mean I have no reason for that. One of them is I tend to count myself as a Lucasian regarding Gödel's incompleteness theorems. Another is that this silly propaganda - _totally designed to obfuscate the befuddlement of the scientific community regarding "intelligence" and mixing it purely rhetorically which subjective terminology which is completely unwarranted_ - has been repeated for decades, along with the ever same promises that we are ALWAYS almost there; with promises receding ever further into the future. And of course, while this racket is sold with health concerns, this is mostly going to benefit the ever increasing domination of the people by an ever smaller class of dominators - robots controlling the population. But there are still other reasons that you don't quite sound like you're worth debating, for instance once someone is prepared to assume _as obvious_ that not only certain ways of mental functioning and capacity, but basic qualia themselves would be an emergent property. *But look: I am all for decisive advances in these areas as I stated in my OP, IF (!) these advances really would serve the good of the community and help destroy the dominator culture, but what I am against is this propaganda as well as to philosophically falsely call it "artificial intelligence". Just call it artificial "expert systems" or "decision making" and it's fine with me.*
+Soteriologe good point, but in some sense, aren't we all merely machines that did well at error optimization? After all, in the real world, making too many mistakes could adversely affect one's ability to pass on their set of genes and so forth. Still, I find this development an exciting milestone on the path toward true AI. This is to date one of the best systems built by humans that can exhibit early patterns of learning. Baby steps in a way, but eventually all babies grow up to become intelligent adults (with the exception of Trump..)
A lot of hype... AI still has a huge way to go. Anybody with a small machine learning background knows that this is another "super specialized" system and that's it.
Yes, maybe we should give up on it.
Brainbuster That is actually not such a bad idea according to many experts. Now fuck off cry baby
pinochska
Cool, man.
Brainbuster Thanks :)
+pinochska I believe they trained it to play quite a wide variety of games, so, seems like it is not that specialized at all!
Hate ai
Then don't eat food, breathe air, drive a car, go to school, or take medicine ever..... because ai, now and forever will improve all of that.