On March 11 2016 04:37 WaveofShadow wrote: And once again, what about this is revolutionary, wei? If my lack of understanding is the issue here then by all means explain to me why should I care about this, specifically?
Computers making decisions without complete information is one of the first steps on the road to true machine learning and, at a distance, AI.
Most machine logic at this point in time is based on parameters. If x then y. There's pattern recognition to be sure, in things like your Google searches, but that's really just taking quantifiable data and finding results based on recency/quantity/etc.
For a machine to be able to think about multiple possible outcomes, and make a decision is huge. It's not picking the one that has historically had the best chance of success, like you see with chess openings. It's not picking it based off of some logic that says "if he moves here, then play here", as a simple Go AI might do on a video game. It's also not randomly putting down pieces until a clearer picture comes out. It's analyzing the board, thinking multiple moves in advance, and choosing a strategy that has a chance of success in the future, without choosing it off of a script.
It bases this move not on some logic that it has been told by the programmers, but based on the concepts of the game and its own experience with previous matches. It, in essence, learns, in the way a person playing Go over time might learn. Easier to visualize here:
Granted, it only does Go right now, but a more complicated version of this system might be taught how to do any number of things without needing to program. Actual AI. The implications are, frankly, staggering. Compared to Deep Blue, Chess is a much more solvable game than Go, just by nature of what it is. You're not talking a double handful of opening turns, you're talking thousands of variations.
And it made moves not that it learned by watching professionals. It wasn't copying moves, which would not be as impressive because all that would mean is it was able to copy the highest chance of success. It was making them up, creating higher odds of success with untested but thought out moves. If you don't think that's fascinating then I have no idea what to tell you.
Excellent explanation, Req. Thank you.
Basically, a program that got trounced by a vastly inferior player 5 months ago, just internally practiced vs itself with almost no outside input, and is now dominating the worlds best at GO.
That alone is insane enough but then remember that GO is a game that isn't solveable, the fastest supercomputers would take around 10^140 years to calculate all go variations, and yet in 5 months it's become better at INTUITIVELY understanding positions than the best human player in a game we've been studying for thousands of years.
When you think of things that humans can do better than robots, go is one of the few things people thought would stay on that list for at least a fair bit longer, and now it's been completely crushed in 5 months...
It was kind of the last objectively evaluate-able thing left that humankind was actually the best at. And now that's gone
We're better at sports
A car can drive faster than a sprinter, and I'm sure if you added a humvee into a game of american football, it would be able to ram through most blockers... I can't think of any "sports" where a properly engineered robot wouldn't just completely destroy people unless you created a bunch of anti robot rules to completely optimse the game for someone with an exact humans physiology, and then it's just a matter of engineering, not AI.
If the thing making a human being superior is purely a side effect of our physical makeup, and an AI could do any mental part way faster/better, then it's literally only a matter of time before a cyborg, and eventually an android that's superior is created.
Computers have so far always needed operators to make objective macro decisions, but if AI reached a point where any objective decision making can be made better by an AI than by a human, then a future where the only thing giving a human any power anywhere was the fact nobody had bothered to make a suitable AI yet is suddenly a very real possibility. Deep mind has shown that given a framework, a network can optimise one of the most incalculable and cerebral challenges known to man.
Up until two days ago, if someone said that an AI could do anything objective a human brain could, better. You were able to say "what about go?"
What can you say now?
If that's not dramatic enough for you look at it this way:
I think most people still believe that there's "something about" humans, or about life, that separates us from machines. But this is a very real and very fast step towards the argument that we're just a bag of relatively inefficient chemical reactions, and all the tools to build an objectively superior being are already coming into existence.
This is the sort of thing that I mean. You're going on about how amazing robots are, while ignoring how amazing humans are.
Deepblue or alphago or whatever it's name is beat this guy at Go. That's really cool and a technological marvel to be sure.
But that doesn't mean that it's anywhere close to having the intuitive or cognitive power of a human. It's still a machine that doesn't do anything but play Go. The human player does a lot more throughout the day than just play Go. Even while he's playing Go he's doing thousands of other things. I imagine if you figured out what percentage of his brains processing power is actually devoted to playing the game, and only let the computer operate at that level, the human player would win easily.
It's not to say that AI advancement isn't awesome or that they can't theoretically become more intelligent than humans. It's just that pretending that we are already there, or even remotely close to getting there, is ludicrous. There's still a long way to go before we have computers that are anywhere close to having the abilities of a human.
Go is a step in the right direction, but at the end of the day it's still something that is the sort of thing computers are naturally good at, a game with simple rules and finite possibilities. We think it's amazing because it's the kind of thing that is difficult for us.
When are computers going to be good at things that are difficult for them but come naturally to humans? Like understanding language, or creating artwork. Etc etc.
You're right, AlphaGo is meant to do one thing. But because it is technology, imagine putting thousands of these together, each tasked with doing one specific thing. Get enough of these things together and you've got yourself, effectively, a brain.
Of course I'm oversimplifying it, and of course there is a level of subjectivity that humans will probably always have. It's not a huge deal because of what it is now, it's a huge deal because of...
On March 11 2016 04:37 WaveofShadow wrote: And once again, what about this is revolutionary, wei? If my lack of understanding is the issue here then by all means explain to me why should I care about this, specifically?
Computers making decisions without complete information is one of the first steps on the road to true machine learning and, at a distance, AI.
Most machine logic at this point in time is based on parameters. If x then y. There's pattern recognition to be sure, in things like your Google searches, but that's really just taking quantifiable data and finding results based on recency/quantity/etc.
For a machine to be able to think about multiple possible outcomes, and make a decision is huge. It's not picking the one that has historically had the best chance of success, like you see with chess openings. It's not picking it based off of some logic that says "if he moves here, then play here", as a simple Go AI might do on a video game. It's also not randomly putting down pieces until a clearer picture comes out. It's analyzing the board, thinking multiple moves in advance, and choosing a strategy that has a chance of success in the future, without choosing it off of a script.
It bases this move not on some logic that it has been told by the programmers, but based on the concepts of the game and its own experience with previous matches. It, in essence, learns, in the way a person playing Go over time might learn. Easier to visualize here:
Granted, it only does Go right now, but a more complicated version of this system might be taught how to do any number of things without needing to program. Actual AI. The implications are, frankly, staggering. Compared to Deep Blue, Chess is a much more solvable game than Go, just by nature of what it is. You're not talking a double handful of opening turns, you're talking thousands of variations.
And it made moves not that it learned by watching professionals. It wasn't copying moves, which would not be as impressive because all that would mean is it was able to copy the highest chance of success. It was making them up, creating higher odds of success with untested but thought out moves. If you don't think that's fascinating then I have no idea what to tell you.
Excellent explanation, Req. Thank you.
Basically, a program that got trounced by a vastly inferior player 5 months ago, just internally practiced vs itself with almost no outside input, and is now dominating the worlds best at GO.
That alone is insane enough but then remember that GO is a game that isn't solveable, the fastest supercomputers would take around 10^140 years to calculate all go variations, and yet in 5 months it's become better at INTUITIVELY understanding positions than the best human player in a game we've been studying for thousands of years.
When you think of things that humans can do better than robots, go is one of the few things people thought would stay on that list for at least a fair bit longer, and now it's been completely crushed in 5 months...
It was kind of the last objectively evaluate-able thing left that humankind was actually the best at. And now that's gone
We're better at sports
Give it until we've built functioning androids.
And then we can have robot football like all those awful FOX kicks too and from commercials.
On March 11 2016 04:37 WaveofShadow wrote: And once again, what about this is revolutionary, wei? If my lack of understanding is the issue here then by all means explain to me why should I care about this, specifically?
Computers making decisions without complete information is one of the first steps on the road to true machine learning and, at a distance, AI.
Most machine logic at this point in time is based on parameters. If x then y. There's pattern recognition to be sure, in things like your Google searches, but that's really just taking quantifiable data and finding results based on recency/quantity/etc.
For a machine to be able to think about multiple possible outcomes, and make a decision is huge. It's not picking the one that has historically had the best chance of success, like you see with chess openings. It's not picking it based off of some logic that says "if he moves here, then play here", as a simple Go AI might do on a video game. It's also not randomly putting down pieces until a clearer picture comes out. It's analyzing the board, thinking multiple moves in advance, and choosing a strategy that has a chance of success in the future, without choosing it off of a script.
It bases this move not on some logic that it has been told by the programmers, but based on the concepts of the game and its own experience with previous matches. It, in essence, learns, in the way a person playing Go over time might learn. Easier to visualize here:
Granted, it only does Go right now, but a more complicated version of this system might be taught how to do any number of things without needing to program. Actual AI. The implications are, frankly, staggering. Compared to Deep Blue, Chess is a much more solvable game than Go, just by nature of what it is. You're not talking a double handful of opening turns, you're talking thousands of variations.
And it made moves not that it learned by watching professionals. It wasn't copying moves, which would not be as impressive because all that would mean is it was able to copy the highest chance of success. It was making them up, creating higher odds of success with untested but thought out moves. If you don't think that's fascinating then I have no idea what to tell you.
Excellent explanation, Req. Thank you.
Basically, a program that got trounced by a vastly inferior player 5 months ago, just internally practiced vs itself with almost no outside input, and is now dominating the worlds best at GO.
That alone is insane enough but then remember that GO is a game that isn't solveable, the fastest supercomputers would take around 10^140 years to calculate all go variations, and yet in 5 months it's become better at INTUITIVELY understanding positions than the best human player in a game we've been studying for thousands of years.
When you think of things that humans can do better than robots, go is one of the few things people thought would stay on that list for at least a fair bit longer, and now it's been completely crushed in 5 months...
It was kind of the last objectively evaluate-able thing left that humankind was actually the best at. And now that's gone
We're better at sports
A car can drive faster than a sprinter, and I'm sure if you added a humvee into a game of american football, it would be able to ram through most blockers... I can't think of any "sports" where a properly engineered robot wouldn't just completely destroy people unless you created a bunch of anti robot rules to completely optimse the game for someone with an exact humans physiology, and then it's just a matter of engineering, not AI.
If the thing making a human being superior is purely a side effect of our physical makeup, and an AI could do any mental part way faster/better, then it's literally only a matter of time before a cyborg, and eventually an android that's superior is created.
Computers have so far always needed operators to make objective macro decisions, but if AI reached a point where any objective decision making can be made better by an AI than by a human, then a future where the only thing giving a human any power anywhere was the fact nobody had bothered to make a suitable AI yet is suddenly a very real possibility. Deep mind has shown that given a framework, a network can optimise one of the most incalculable and cerebral challenges known to man.
Up until two days ago, if someone said that an AI could do anything objective a human brain could, better. You were able to say "what about go?"
What can you say now?
If that's not dramatic enough for you look at it this way:
I think most people still believe that there's "something about" humans, or about life, that separates us from machines. But this is a very real and very fast step towards the argument that we're just a bag of relatively inefficient chemical reactions, and all the tools to build an objectively superior being are already coming into existence.
But that doesn't mean that it's anywhere close to having the intuitive or cognitive power of a human. It's still a machine that doesn't do anything but play Go. The human player does a lot more throughout the day than just play Go. Even while he's playing Go he's doing thousands of other things. I imagine if you figured out what percentage of his brains processing power is actually devoted to playing the game, and only let the computer operate at that level, the human player would win easily.
You're missing the whole point if you think it's because a computer is good at GO.
This isn't a computer that people taugh to play GO, it's a computer which taught ITSELF to play GO.
They created a computer capable of learning, pointed it in GO's direction, and it taught itself how to beat the world champion at GO.
Things like processing power, or multitasking are irrelevant, those are only going to increase exponentially into the future as they always have. The point here is that a computer program TAUGHT ITSELF how to be be better than the world champion at one of the only remaining problem types where we thought humans had an advantage, because they're "unsolveable".
It's not taking human knowledge and applying more processing power, it's built it's own way of playing from the ground up, in a game which can't be won based off calculating speed, and is now beating the best humans.
These kind of problems are the very ones which made an actual AI seem like fantasy 10 or 20 years ago, and deep mind is a proof of concept that neural networking can solve them WAY WAY more easily than almost anyone thought possible.
When talking about future tech things like "processing power" or "multitasking" are irrelevant because they scale. It's the roadblocks that are interesting, and the type of thinking GO requires was one of the most fundamental of those blocks.
No, I get that. I absolutely get that. It's amazing.
But the thing is, Go is still fundamentally not as complicated as lots of other things that we do every day, but take for granted, because they come naturally to us. Just the simple act of having a conversation (for example) is much more complex, complicated and involved than playing Go. We think conversations are easy because we are designed to be good at them. Computers are designed to be good at things like games.
It's a big deal and a big step sure, but it doesn't mean that computers are anywhere near smarter than humans or rendering them obsolete or whatever Hollywood scenario blah blah. There's still a huge, huuuuuge ways to go before computers are doing the sorts of things that are as complex as the things that we take for granted because we do them so easily.
Basically, when you say Go is one of the "only remaining problem types where humans have an advantage" you're just straight wrong.
On March 11 2016 04:37 WaveofShadow wrote: And once again, what about this is revolutionary, wei? If my lack of understanding is the issue here then by all means explain to me why should I care about this, specifically?
Computers making decisions without complete information is one of the first steps on the road to true machine learning and, at a distance, AI.
Most machine logic at this point in time is based on parameters. If x then y. There's pattern recognition to be sure, in things like your Google searches, but that's really just taking quantifiable data and finding results based on recency/quantity/etc.
For a machine to be able to think about multiple possible outcomes, and make a decision is huge. It's not picking the one that has historically had the best chance of success, like you see with chess openings. It's not picking it based off of some logic that says "if he moves here, then play here", as a simple Go AI might do on a video game. It's also not randomly putting down pieces until a clearer picture comes out. It's analyzing the board, thinking multiple moves in advance, and choosing a strategy that has a chance of success in the future, without choosing it off of a script.
It bases this move not on some logic that it has been told by the programmers, but based on the concepts of the game and its own experience with previous matches. It, in essence, learns, in the way a person playing Go over time might learn. Easier to visualize here:
Granted, it only does Go right now, but a more complicated version of this system might be taught how to do any number of things without needing to program. Actual AI. The implications are, frankly, staggering. Compared to Deep Blue, Chess is a much more solvable game than Go, just by nature of what it is. You're not talking a double handful of opening turns, you're talking thousands of variations.
And it made moves not that it learned by watching professionals. It wasn't copying moves, which would not be as impressive because all that would mean is it was able to copy the highest chance of success. It was making them up, creating higher odds of success with untested but thought out moves. If you don't think that's fascinating then I have no idea what to tell you.
Excellent explanation, Req. Thank you.
Basically, a program that got trounced by a vastly inferior player 5 months ago, just internally practiced vs itself with almost no outside input, and is now dominating the worlds best at GO.
That alone is insane enough but then remember that GO is a game that isn't solveable, the fastest supercomputers would take around 10^140 years to calculate all go variations, and yet in 5 months it's become better at INTUITIVELY understanding positions than the best human player in a game we've been studying for thousands of years.
When you think of things that humans can do better than robots, go is one of the few things people thought would stay on that list for at least a fair bit longer, and now it's been completely crushed in 5 months...
It was kind of the last objectively evaluate-able thing left that humankind was actually the best at. And now that's gone
We're better at sports
A car can drive faster than a sprinter, and I'm sure if you added a humvee into a game of american football, it would be able to ram through most blockers... I can't think of any "sports" where a properly engineered robot wouldn't just completely destroy people unless you created a bunch of anti robot rules to completely optimse the game for someone with an exact humans physiology, and then it's just a matter of engineering, not AI.
If the thing making a human being superior is purely a side effect of our physical makeup, and an AI could do any mental part way faster/better, then it's literally only a matter of time before a cyborg, and eventually an android that's superior is created.
Computers have so far always needed operators to make objective macro decisions, but if AI reached a point where any objective decision making can be made better by an AI than by a human, then a future where the only thing giving a human any power anywhere was the fact nobody had bothered to make a suitable AI yet is suddenly a very real possibility. Deep mind has shown that given a framework, a network can optimise one of the most incalculable and cerebral challenges known to man.
Up until two days ago, if someone said that an AI could do anything objective a human brain could, better. You were able to say "what about go?"
What can you say now?
If that's not dramatic enough for you look at it this way:
I think most people still believe that there's "something about" humans, or about life, that separates us from machines. But this is a very real and very fast step towards the argument that we're just a bag of relatively inefficient chemical reactions, and all the tools to build an objectively superior being are already coming into existence.
But that doesn't mean that it's anywhere close to having the intuitive or cognitive power of a human. It's still a machine that doesn't do anything but play Go. The human player does a lot more throughout the day than just play Go. Even while he's playing Go he's doing thousands of other things. I imagine if you figured out what percentage of his brains processing power is actually devoted to playing the game, and only let the computer operate at that level, the human player would win easily.
You're missing the whole point if you think it's because a computer is good at GO.
This isn't a computer that people taugh to play GO, it's a computer which taught ITSELF to play GO.
They created a computer capable of learning, pointed it in GO's direction, and it taught itself how to beat the world champion at GO.
Things like processing power, or multitasking are irrelevant, those are only going to increase exponentially into the future as they always have. The point here is that a computer program TAUGHT ITSELF how to be be better than the world champion at one of the only remaining problem types where we thought humans had an advantage, because they're "unsolveable".
It's not taking human knowledge and applying more processing power, it's built it's own way of playing from the ground up, in a game which can't be won based off calculating speed, and is now beating the best humans.
These kind of problems are the very ones which made an actual AI seem like fantasy 10 or 20 years ago, and deep mind is a proof of concept that neural networking can solve them WAY WAY more easily than almost anyone thought possible.
When talking about future tech things like "processing power" or "multitasking" are irrelevant because they scale. It's the roadblocks that are interesting, and the type of thinking GO requires was one of the most fundamental of those blocks.
It is pretty awesome what has been accomplished but I think people are slightly mis-attributing what this AI has done/is doing. It's 'learning' in the strict sense but it is still learning very much like a computer by brute forcing a task with mathematics, it's just arriving there by running millions of iterations of the exercise instead of having a problem tree created to 'solve.' AlphaGo is still doing what previous computers have done(making the statistically highest value move towards winning at any point) it's just now making it by using its 'memory' which is a large step in the direction of human thought and amazing but still very computer like.
I have tons of thoughts on this subject that I feel I can't articulate as well as I'd like through text(at least, it would take forever) but I think I am decidedly not in either of the camps that appear o have formed here. This is spectacular, but I don't think it is as groundbreaking as some are making it out to be(especially since it's actually a decent time in the making and was actually being raced towards by many different teams).
Yeah, I agree with Red. It's still, fundamentally, taking a math problem with a solution and figuring out how to solve it. It's just doing it in a way where it can solve it by making assumptions since it can't know everything. That's amazing, but it's not amazing on the scale of the sorts of amazing things humans do.
So, while it's a big step, it's not "oh my god the robots are taking over". No, they're really not.
On March 11 2016 11:58 Ketara wrote: But the thing is, Go is still fundamentally not as complicated as lots of other things that we do every day, but take for granted, because they come naturally to us. Just the simple act of having a conversation (for example) is much more complex, complicated and involved than playing Go.
Sure, but the difference between a solvably complex problem and an unsolvably complex problem is a lot less than the difference between an unsolvably complex problem and a somewhat more complicated unsolvably complex problem.
On March 11 2016 11:58 Ketara wrote:We think conversations are easy because we are designed to be good at them. Computers are designed to be good at things like games.
Well, no. We're not *designed* to be good at anything. What we do is adapt to the the various sensory inputs we're given and learn to accomplish tasks and become better at those tasks.
The milestone here is that we went from having an AI that's designed to do a particular thing to one that can do just that--learn to do things.
On March 11 2016 11:58 Ketara wrote: It's a big deal and a big step sure, but it doesn't mean that computers are anywhere near smarter than humans or rendering them obsolete or whatever Hollywood scenario blah blah.
Literally nobody is saying that, lol.
On March 11 2016 11:58 Ketara wrote: Basically, when you say Go is one of the "only remaining problem types where humans have an advantage" you're just straight wrong.
He isn't, you're just not clearly understanding what he means by "problem types".
It may not be there yet, but this is a big step. True learning is one of the first hurdles towards actual AI, and proving that they can do it this well is a major stride in that direction.
Of course, with all things science and technology, the end point might be a decade away, it might be 50 years away. But it's a markable advancement, and that's not nothing.
The whole joke with 42 or whatever in hitchhikers guide is that the most complex thing is asking the question rather than the answer. Not in the doctor kind of way but in the Irobot kind of way to how to advance.
if the standard for AI is 'setting goals like humans' then that's just sneaking some 'soul' into the intelligence talk. it's not like humans are amazingly sophisticated goal pickers. modeling a human emotional system isn't theoretically hard. it's just that when people see the causal machinery they stop projecting 'intelligence' onto the machine.
replicating the 'human experience' is really not what AI is about.
People are absolutely saying that computers are smarter than people and robots are going to take over the world Yango.
I'm just saying that even though this is a justifiably big deal, it's not like we are going to have ghost in the shell major kusanagi type androids running around in the next 10 years. That might happen some day, but I'd be skeptical of it happening even within our lifetime.
Personally, I'm super stoked for stuff like Google Now getting more intuitive. I love talking to my phone and I want it to be better at predicting what it can do for me.
On March 11 2016 04:37 WaveofShadow wrote: And once again, what about this is revolutionary, wei? If my lack of understanding is the issue here then by all means explain to me why should I care about this, specifically?
Computers making decisions without complete information is one of the first steps on the road to true machine learning and, at a distance, AI.
Most machine logic at this point in time is based on parameters. If x then y. There's pattern recognition to be sure, in things like your Google searches, but that's really just taking quantifiable data and finding results based on recency/quantity/etc.
For a machine to be able to think about multiple possible outcomes, and make a decision is huge. It's not picking the one that has historically had the best chance of success, like you see with chess openings. It's not picking it based off of some logic that says "if he moves here, then play here", as a simple Go AI might do on a video game. It's also not randomly putting down pieces until a clearer picture comes out. It's analyzing the board, thinking multiple moves in advance, and choosing a strategy that has a chance of success in the future, without choosing it off of a script.
It bases this move not on some logic that it has been told by the programmers, but based on the concepts of the game and its own experience with previous matches. It, in essence, learns, in the way a person playing Go over time might learn. Easier to visualize here:
Granted, it only does Go right now, but a more complicated version of this system might be taught how to do any number of things without needing to program. Actual AI. The implications are, frankly, staggering. Compared to Deep Blue, Chess is a much more solvable game than Go, just by nature of what it is. You're not talking a double handful of opening turns, you're talking thousands of variations.
And it made moves not that it learned by watching professionals. It wasn't copying moves, which would not be as impressive because all that would mean is it was able to copy the highest chance of success. It was making them up, creating higher odds of success with untested but thought out moves. If you don't think that's fascinating then I have no idea what to tell you.
Excellent explanation, Req. Thank you.
Basically, a program that got trounced by a vastly inferior player 5 months ago, just internally practiced vs itself with almost no outside input, and is now dominating the worlds best at GO.
That alone is insane enough but then remember that GO is a game that isn't solveable, the fastest supercomputers would take around 10^140 years to calculate all go variations, and yet in 5 months it's become better at INTUITIVELY understanding positions than the best human player in a game we've been studying for thousands of years.
When you think of things that humans can do better than robots, go is one of the few things people thought would stay on that list for at least a fair bit longer, and now it's been completely crushed in 5 months...
It was kind of the last objectively evaluate-able thing left that humankind was actually the best at. And now that's gone
We're better at sports
A car can drive faster than a sprinter, and I'm sure if you added a humvee into a game of american football, it would be able to ram through most blockers... I can't think of any "sports" where a properly engineered robot wouldn't just completely destroy people unless you created a bunch of anti robot rules to completely optimse the game for someone with an exact humans physiology, and then it's just a matter of engineering, not AI.
If the thing making a human being superior is purely a side effect of our physical makeup, and an AI could do any mental part way faster/better, then it's literally only a matter of time before a cyborg, and eventually an android that's superior is created.
Computers have so far always needed operators to make objective macro decisions, but if AI reached a point where any objective decision making can be made better by an AI than by a human, then a future where the only thing giving a human any power anywhere was the fact nobody had bothered to make a suitable AI yet is suddenly a very real possibility. Deep mind has shown that given a framework, a network can optimise one of the most incalculable and cerebral challenges known to man.
Up until two days ago, if someone said that an AI could do anything objective a human brain could, better. You were able to say "what about go?"
What can you say now?
If that's not dramatic enough for you look at it this way:
I think most people still believe that there's "something about" humans, or about life, that separates us from machines. But this is a very real and very fast step towards the argument that we're just a bag of relatively inefficient chemical reactions, and all the tools to build an objectively superior being are already coming into existence.
But that doesn't mean that it's anywhere close to having the intuitive or cognitive power of a human. It's still a machine that doesn't do anything but play Go. The human player does a lot more throughout the day than just play Go. Even while he's playing Go he's doing thousands of other things. I imagine if you figured out what percentage of his brains processing power is actually devoted to playing the game, and only let the computer operate at that level, the human player would win easily.
You're missing the whole point if you think it's because a computer is good at GO.
This isn't a computer that people taugh to play GO, it's a computer which taught ITSELF to play GO.
They created a computer capable of learning, pointed it in GO's direction, and it taught itself how to beat the world champion at GO.
Things like processing power, or multitasking are irrelevant, those are only going to increase exponentially into the future as they always have. The point here is that a computer program TAUGHT ITSELF how to be be better than the world champion at one of the only remaining problem types where we thought humans had an advantage, because they're "unsolveable".
It's not taking human knowledge and applying more processing power, it's built it's own way of playing from the ground up, in a game which can't be won based off calculating speed, and is now beating the best humans.
These kind of problems are the very ones which made an actual AI seem like fantasy 10 or 20 years ago, and deep mind is a proof of concept that neural networking can solve them WAY WAY more easily than almost anyone thought possible.
When talking about future tech things like "processing power" or "multitasking" are irrelevant because they scale. It's the roadblocks that are interesting, and the type of thinking GO requires was one of the most fundamental of those blocks.
It is pretty awesome what has been accomplished but I think people are slightly mis-attributing what this AI has done/is doing. It's 'learning' in the strict sense but it is still learning very much like a computer by brute forcing a task with mathematics, it's just arriving there by running millions of iterations of the exercise instead of having a problem tree created to 'solve.' AlphaGo is still doing what previous computers have done(making the statistically highest value move towards winning at any point) it's just now making it by using its 'memory' which is a large step in the direction of human thought and amazing but still very computer like.
I don't think it's quite as simple as that though. It's easy to think "well it just worked out the best move", but you have to remember that this is GO and the scale is something completely different from almost any comparable problem.
Obviously we don't know the specifics of the hardware they're using, or the types of algorithms they're running, but for an idea of scale, the fastest supercomputer in the world (according to wikipedia) is the Tianhe-2, which operates at around 33.8petaflops, or 33e15 flops. If Deep mind was somehow running on Tianhe-2, using 100% of it's resources, for the entire 5 months since it got trounced last year, then it would have analysed through around 0.000+ Show Spoiler +
00002% of possible go positions. give or take a few orders of magnitude.
Not to mention even if it did just calculate that tiny subset of positions, we're already into yottabytes of data just to store it's results in tables, let alone accessing that data.
It managed to go through a minuscule subset of possible games, and from that create it's very own, completely new robust system of logic that lets it appraise brand new positions independently, and come up with optimal solutions, so while it did iterate to improve it's model, I think it's a bit of a disservice to imply that "all" it did was work out the best move by just checking more positions than a human could. It's mastered some rather abstract thinking, and most fascinatingly, is playing in some styles which appear odd to the human professionals watching the games.
Sorry if the post seems a bit "click baity" or anything, but a lot of the people I've talked to have seemed to think it's just a computer working out the best move by calculations, and it can be hard to explain why that's not the case since humans are generally pretty bad at comprehending numbers of this kind of size :p
It's like managing to understand the entire earth from looking at one hydrogen atom, If the earth was so large that our current earth occupies less space in this theoretical earth then that hydrogen atom does in ours, and then times that by 10^20.
On March 11 2016 12:22 Ketara wrote: People are absolutely saying that computers are smarter than people and robots are going to take over the world Yango.
I'm just saying that even though this is a justifiably big deal, it's not like we are going to have ghost in the shell major kusanagi type androids running around in the next 10 years. That might happen some day, but I'd be skeptical of it happening even within our lifetime.
Personally, I'm super stoked for stuff like Google Now getting more intuitive. I love talking to my phone and I want it to be better at predicting what it can do for me.
On March 11 2016 12:22 Ketara wrote: People are absolutely saying that computers are smarter than people and robots are going to take over the world Yango.
Ok fine, but that's not what anyone here or anyone with a brain that can process information more effectively than AlphaGo is saying.
In other news, does anybody have any experience with Malaysia?
Probably not, but could be helpful since I guess I'm moving to Kuala Lumpur. I was only there for like a week so, if anybody knows anybody or anything plz ty plz.
Whomever shared that Civ5 Killian Experience video a while back, thank you so so very much. His videos are a garunteed laugh for me when I need it. Or when I don't. Fuck it.
On March 11 2016 12:31 Ketara wrote: In other news, does anybody have any experience with Malaysia?
Probably not, but could be helpful since I guess I'm moving to Kuala Lumpur. I was only there for like a week so, if anybody knows anybody or anything plz ty plz.