AI beats the game Go

[Jub]

When I said "A quarter" I might have been sugercoating it.
The exact quote by the head of the department was "get rid of most of the QA division" .
What you're describing is jus classifying a bug as belonging to a predefined category + a script with the answer/class.
So - what's your current job ?

I'm between work at the moment, the new management they hired at the tech support firm just made the job unworkable compared to what it was when I was hired. I'm trying to figure out how I might afford going back to school while paying rent, bills, and eating well. If not it's back into retail for me.

[Purple]

If you're so shortsighted that you cannot grasp the basic problem with trying to keep special rights for yourself, while explicitly not caring if anyone else gets them... you're not worth talking to.

Listen, of course I would prefer a world where nobody is screwed over. That goes without saying. Which is why I did not say it. I sort of come from the position that you will make the tacit assumption that I am not the second coming of Hitler.

All I am saying is that if that ain't an option and I had a choice between a world where everybody is screwed and one where everybody but me is screwed I would pick #2 without a moments thought. It's called having a self preservation instinct.

[Adam Reynolds]

Do you really not see the logical problem in that position? If anyone else had their way, they would do exactly the same thing and technological progress would die.

[Simon_Jester]

Do you really not see the logical problem in that position? If anyone else had their way, they would do exactly the same thing and technological progress would die.

That's (at best) a practical problem, not a logical one.

The logical problem with his position is that "screw everyone else over if you want, but leave me out of it" is a rule that cannot be applied consistently. It leads to Prisoner's Dilemma scenarios in which everyone chooses the 'defect' option and everyone suffers as a consequence.

Naked self-interest does nothing for one's position in a moral debate, except to disqualify one from participating.

If you're so shortsighted that you cannot grasp the basic problem with trying to keep special rights for yourself, while explicitly not caring if anyone else gets them... you're not worth talking to.

Listen, of course I would prefer a world where nobody is screwed over. That goes without saying. Which is why I did not say it. I sort of come from the position that you will make the tacit assumption that I am not the second coming of Hitler.

You're clearly not; Hitler had charisma and the ability to convince his associates he was trustworthy.

All I am saying is that if that ain't an option and I had a choice between a world where everybody is screwed and one where everybody but me is screwed I would pick #2 without a moments thought. It's called having a self preservation instinct.

The way you say this reveals thought processes which make you not worth talking to- see my previous post.

[Purple]

The way you say this reveals thought processes which make you not worth talking to- see my previous post.

So you would chose #1? You would willingly choose to suffer just because you can't save the world?

[Adam Reynolds]

The logical problem with his position is that "screw everyone else over if you want, but leave me out of it" is a rule that cannot be applied consistently. It leads to Prisoner's Dilemma scenarios in which everyone chooses the 'defect' option and everyone suffers as a consequence.

That is sort of what I meant.

[cosmicalstorm]

AGI is near according to this LW poster. The data set used was about the size of a 30yr olds life exp?

jacob_cannell comments on [Link] AlphaGo: Mastering the ancient game of Go with Machine Learning - Less Wrong Discussion

13ESRogs27 January 2016 09:04PM

You are viewing a comment permalink. View the original post to see all comments and the full post content.

Article Navigation

Comments (112)

CC Licenced

Comments (112)

Sort By: Best

You are viewing a single comment's thread. Show more comments above.

jacob_cannell30 January 2016 12:22:49AM * 9 points[-]

This is a big deal, and it is another sign that AGI is near.

Intelligence boils down to inference. Go is an interesting case because good play for both humans and bots like AlphaGo requires two specialized types of inference operating over very different timescales:

rapid combinatoric inference over move sequences during a game(planning). AlphaGo uses MCT search for this, whereas the human brain uses a complex network of modules involving the basal ganglia, hippocampus, and PFC.slow deep inference over a huge amount of experience to develop strong pattern recognition and intuitions (deep learning). AlphaGo uses deep supervised and reinforcement learning via SGD over a CNN for this. The human brain uses the cortex.

Machines have been strong in planning/search style inference for a while. It is only recently that the slower learning component (2nd order inference over circuit/program structure) is starting to approach and surpass human level.

Critics like to point out that DL requires tons of data, but so does the human brain. A more accurate comparison requires quantifying the dataset human pro go players train on.

A 30 year old asian pro will have perhaps 40,000 hours of playing experience (20 years * 50 * 40 hrs/week). The average game duration is perhaps an hour and consists of 200 moves. In addition, pros (and even fans) study published games. Reading a game takes less time, perhaps as little as 5 minutes or so.

So we can estimate very roughly that a top pro will have absorbed between 100,000 games to 1 million games, and between 20 to 200 million individual positions (around 200 moves per game) .

AlphaGo was trained on the KGS dataset: 160,00 games and 29 million positions. So it did not train on significantly more data than a human pro. The data quantities are actually very similar.

Furthermore, the human's dataset is perhaps of better quality for a pro, as they will be familiar with mainly pro level games, whereas the AlphaGo dataset is mostly amateur level.

The main difference is speed. The human brain's 'clockrate' or equivalent is about 100 hz, whereas AlphaGo's various CNNs can run at roughly 1000hz during training on a single machine, and perhaps 10,000 hz equivalent distributed across hundreds of machines. 40,000 hours - a lifetime of experience - can be compressed 100x or more into just a couple of weeks for a machine. This is the key lesson here.

The classification CNN trained on KGS was run for 340 million steps, which is about 10 iterations per unique position in the database.

The ANNs that AlphaGo uses are much much smaller than a human brain, but the brain has to do a huge number of other tasks, and also has to solve complex vision and motor problems just to play the game. AlphaGO's ANNs get to focus purely on Go.

A few hundred TitanX's can muster up perhaps a petaflop of compute. The high end estimate of the brain is 10 petaflops (100 trillion synapses * 100 hz max firing rate). The more realistic estimate is 100 teraflops (100 trillion synapes * 1 hz avg firing rate), and the lower end is 1/10 that or less.

So why is this a big deal? Because it suggests that training a DL AI to master more economically key tasks, such as becoming an expert level programmer, could be much closer than people think.

The techniques used here are nowhere near their optimal form yet in terms of efficiency. When Deep Blue beat Kasparov in 1996, it required a specialized supercomputer and a huge team. 10 years later chess bots written by individual programmers running on modest PC's soared past Deep Blue - thanks to more efficient algorithms and implementations.

Vote upVote down

---

http://lesswrong.com/r/discussion/lw/n8 ... of_go/d2o5

[Starglider]

Less Wrong has a lot of posters that are good at chaining simple calculations but bad at appreciating qualitative arguments. As a pattern processing exercise, the problem of guiding a Go move search algorithm is relatively flat and simple, albeit needing a larger number of learned patterns than most NNs to date can handle for good performance. The result can indeed be optimised down to something more tractable via techniques similar to classic NN rules extraction (back in the day, resolving trained NNs back to symbolic rulesets; it's a bit more complicated for modern large NNs but there's a conceptually similar simplification to more efficient statistical techniques). However this does not mean that AGI is imminent, because the layering and modularisation challenges remain and progress here is slow and not so amenable to more hardware and local algorithmic tweaks. It's definitely relevant progress, and we are certainly moving down the road to AGI (several at once actually, albeit at different rates), but there are several major hurdles still to clear.

[Ziggy Stardust]

It's been a while since I've read about neural networks, Starglider (and when I have read about them, it has been more about their applications in the field of biostatistics), but IIRC one of the outstanding issues with them as a modeling technique is that, even though they are remarkably effective, in many ways they are still a bit of a black box. That is, you can design a neural network model that faithfully replicates some probabilistic or biological process in a consistent and powerful way, but it's not always clear exactly what mechanisms within the model are responsible for its behavior, making it difficult to deconstruct (which tends to slow the research down, because it is harder to develop the next model when you aren't entirely sure which aspect of the last model you should be focusing on). Am I remembering correctly, or am I under a misapprehension about NN?

[Starglider]

It's been a while since I've read about neural networks, Starglider (and when I have read about them, it has been more about their applications in the field of biostatistics), but IIRC one of the outstanding issues with them as a modeling technique is that, even though they are remarkably effective, in many ways they are still a bit of a black box. That is, you can design a neural network model that faithfully replicates some probabilistic or biological process in a consistent and powerful way, but it's not always clear exactly what mechanisms within the model are responsible for its behavior, making it difficult to deconstruct (which tends to slow the research down, because it is harder to develop the next model when you aren't entirely sure which aspect of the last model you should be focusing on). Am I remembering correctly, or am I under a misapprehension about NN?

Well, that's contraversial. When NNs were first introduced they were certainly opaque, and the leading edge of NN research always produces nets that are too complex for the corresponding generation of analysis tools. However the waters here are substantially muddied by numerous NN researchers actually being proud of the opacity of their designs, treating it as mysterious emergent magic that meant that an NN solution was inherently superior even if it wasn't outperforming a symbolic alternative. Supposedly this is because not understanding how it works guarantees that you do actually have a learning algorithm which is in theory more generalisable than a non-connectionist approach which tends to have more hand-written non-learned complexity, although if you take a close look at how carefully NN researchers select and preprocess training data (in domains outside something at clean as Go) this is pretty dubious assumption.

Anyway, after the second round of NN success (three decades after the first perceptron flash in the pan) the Symbolic Empire Struck Back (back then they were the ones with all the funding, the situation is reversed now) with the initial burst of NN Rules Extraction papers, starting in the late 80s and prevalent in the first half of the 90s. Assorted techniques were developed to simplify NNs back down to compact predicates and/or continuous functions. The initial goal was to use NNs as a rule learning system for production rule based systems, but in the process it demonstrated that NNs weren't do anything nearly as clever as what their promoters thought they were. Of course connectionism has come a long way since then, while symbolic AI was hamstring by people (a) not using probability correctly, (b) not adopting sensible solutions to symbol grounding, (c) insufficient effort to integrate with statistical ML and continous modelling techniques and of course (d) reliance on simple production rules and SOAR-type chunking rather than automated algorithm design. Despite that, the thorn in the side of the people prematurely hyping NN continued through the early 2000s, as some of the more pragmatic ML people developed techniques like SVMs (as used by nearly every online product recommendation system) that replicated or exceeded NN performance with much simpler algorithms. NN rules extraction and more generally NN function extraction has continued in the background up to the current data and if you google for it you will find a quiet but steady stream of papers on the topic.

This is not to say that NN can't work or that it isn't one of the leading approaches to AGI, but it does illustrate that (a) it's not magic, (b) it often does less clever things than you thought, (c) you are trading off hardware efficiency and ability to understand what the system is doing for a more general learning ability that can crack many kinds of problem that more specific approaches have yet to be developed for. So in that sense it is throwing hardware at the problem to leapfrog researchers working on more efficient techniques, although scaling NNs certainly has needed some algorithmic improvements not just bigger supercomputers.

All that said, I would note that NNs on their own aren't actually a complete tabula rasa solution no matter what the hardware. The brain illustrates the need for modularisation, layering and structural complexity at many levels of organisation that simple NNs just don't have; Edelman's modularised NN systems were perhaps the first good example of what we're going to need to do to get NNs to move towards AGI. The neuromorphic (brain simulation) people know this very well, some startups get it too e.g. Numenta's attempt, but quite a few other startups pretend it's just a scaling issue. To make NNs produce AGI from scratch without hand-designing (or copying from humans) a lot of extra complexity, you would have to use an evolutionary approach. Evolved NNs (evolving the topological properties and the neuron function itself) are certainly a subfield that has received a fair bit of attention (from the artificial life crowd), but it's incredibly inefficient because you're compounding the evaluation inefficiency of NNs on top of the search inefficiency of genetic algorithms. It is effectively equivalent to simulating the entire evolutionary history of life from insects up to human beings. I'd be surprised if a pure 'emeregence' based approach like that ever worked, but you never know, someone might invent sufficiently large quantum computers for it.

[Channel72]

Yeah - to add to what Starglider is saying, the basic idea (historically, at least) behind neural networks was simply classifiers - i.e., a data structure that can be fed some training data, and then can be fed never-before-seen data and be able to tell you "okay, this data can be classified as a [X]". In its simplest form, an NN (a simple perceptron) was basically just a linear function that draws a line between various points on a graph, in order to separate things on each side of the line into separate classes. Of course, it was immediately realized back in the 60s or whatever that this didn't even work for a simple XOR function, and so eventually some calculus was applied to create multi-layered NNs that use back propagation so that the firing of neurons in the left-most layer propagates to the firing of neurons in the output layer (the right-most layer).

In general, an NN is just a simply NxM matrix, where each column is a "layer", and the neurons in each layer have a firing function (usually a sigmoidal function) which fires if some certain input meets a certain threshold, and the result of the firing pattern in one layer is the input for the next layer, until the output layer is reached, after which you have your classifications.

It's pretty straightforward, and at least as I've seen, is used pretty much as a black box. Programmers who make use of neural nets are usually trying to classify some kind of data, so they start out by using plain intuition or common sense to come up with a "feature vector", which basically just means a list of "features" that can help with the classification. Like, if you wanted to classify some object as a "car", your feature vector might include things like "has 4 wheels", "has a muffler", etc. Obviously, real feature vectors can include hundreds or thousands of features, and tend to be normalized before being fed to the neural net. The network is then "trained" with a large-ish set of training data (truth data) - meaning you feed the neural network training data, and explicitly tell it that "these things are definitely all of class [X]". After the neural net is trained, it will (hopefully) be able to classify never-before-seen data correctly most of the time.

In practice, there's a lot of trial and error. I've done this a lot with document classification - using neural nets to try and determine if a certain document is say, a news article vs an entertainment blog, etc. Basically it involves just a lot of intuition, trial and error, and creativity to come up with a feature vector that doesn't suck. There's also a lot of "black magic" involved in coming up with a training set that's good enough, but not too specific, or else the NN may "overfit" and fail to generalize well when it sees never-before-seen data (mostly because it considered irrelevant details/noise in the training data as something that was important.)

All in all, correct use of neural nets seems to be something more of an art than an exact science, because it's pretty impossible to "debug" a neural net after the fact. If your NN doesn't do what you want, there's nothing really to do but try a different feature vector/training set, and keep experimenting until you have an NN which generalizes very well with never-before-seen data.

[Starglider]

That is the simplest kind of NN yes. There are lot of more complicated flavours of varying popularity, e.g. spiking neural networks, which have had a lot of hype (sooo sleek, sooo biomorphic, sooo computationally inefficient) but negligible practical applications to date. But the brain is a spiking neural network (with a lot of non-obvious side channels), so funding for that particular concept is immune to poor short-term results.

[K. A. Pital]

...

I need not reply to your post in pieces because I only really need to say one thing. Or rather ask one thing. Why do you find it in any way strange, surprising or potentially negative that I would, like any sensible human being despise anything that has the potential to make my life more difficult?

Purple, you are either trolling, or failing the ethics equivalent of the Turing Test.

If you're trolling, you're not worth talking to.

If you're so shortsighted that you cannot grasp the basic problem with trying to keep special rights for yourself, while explicitly not caring if anyone else gets them... you're not worth talking to.

Uh, fact of the matter is, a lot of people behave like that. It is a very common attitude. Not in my backyard, etc. People rarely value others or their well-being as much or more than their own. If that is an ethics Turing test, most people would fail it as they're not immediately sacrificing their lives for the embetterment of someone else's position, and quite likely are making someone's position incredibly, horribly bad. Maybe even that of thousands or millions of others in case of high-level decision planners (who often still enjoy an enormous life standard regardless of the fact).

I know Purple looks like an ethics troll sometimes - and maybe he is - but the question is still valid. Why is his behaviour strange, but not that of other people?

There are certainly hordes of low-to-mediocre skilled programmers who can and should be eliminated by relatively narrow automated software engineering tools.

The good old Starglider...

It is impressed upon the workpeople, as a great consolation, first, that their sufferings are only temporary (“a temporary inconvenience"), secondly, that machinery acquires the mastery over the whole of a given field of production, only by degrees, so that the extent and intensity of its destructive effect is diminished. The first consolation neutralises the second. When machinery seizes on an industry by degrees, it produces chronic misery among the operatives who compete with it. Where the transition is rapid, the effect is acute and felt by great masses. History discloses no tragedy more horrible than the gradual extinction of the English hand-loom weavers, an extinction that was spread over several decades, and finally sealed in 1838. Many of them died of starvation, many with families vegetated for a long time on 2½ d. a day. On the other hand, the English cotton machinery produced an acute effect in India. The Governor General reported 1834-35:

“The misery hardly finds a parallel in the history of commerce. The bones of the cotton-weavers are bleaching the plains of India.”

No doubt, in turning them out of this “temporal” world, the machinery caused them no more than “a temporary inconvenience.” For the rest, since machinery is continually seizing upon new fields of production, its temporary effect is really permanent.

I have no doubt the elimination of mediocre programmers is at hand by the likes of the industrious Geek Galts.

[Simon_Jester]

Stas, it's not so much that he IS prepared to sell out literally everyone else except himself.

It's that he doesn't grasp why this is something to at least be embarrassed about... He's like "yes, what's the problem with sacrificing everyone else as long as I'm okay? Sure, that's somehow less good, but I don't really care about them."

Which means:
1) He doesn't have the grace to be satisfying to talk to because he has no restraint or capacity to not say the shitty stupid things popping into his head
2) He's so philosophically and historically illiterate, he doesn't grasp that when you sell others to a process that devours them, you are incredibly likely to end up next on the chopping block.

[K. A. Pital]

True, what he says sounds stupid - and is stupid. I just think that it at least a form of honest stupidity rather than merely a play. I may be wrong, though.

[Lord Revan]

True, what he says sounds stupid - and is stupid. I just think that it at least a form of honest stupidity rather than merely a play. I may be wrong, though.

While I think Purple isn't really quite as sociopathic as he seems due to having "friends" who enhance that sort behaviour instead of reducing it, I think it's not all a play either and he is truly quite socially blind.

[Purple]

Stas, it's not so much that he IS prepared to sell out literally everyone else except himself.

It's that he doesn't grasp why this is something to at least be embarrassed about... He's like "yes, what's the problem with sacrificing everyone else as long as I'm okay? Sure, that's somehow less good, but I don't really care about them."

Which means:
1) He doesn't have the grace to be satisfying to talk to because he has no restraint or capacity to not say the shitty stupid things popping into his head
2) He's so philosophically and historically illiterate, he doesn't grasp that when you sell others to a process that devours them, you are incredibly likely to end up next on the chopping block.

Or, and this is something you have completely failed to consider I am simply being honest. As KA said most people are in fact, when it comes down to it going to do the same. I have simply accepted this fact and thus I feel absolutely no need to pretend to be a special snowflake and hide behind fake morality. I prefer to tell you to your face that I will betray you than to promise you everlasting loyalty knowing full well that I am going to stab you in the back if it comes down to it. It's as simple as that.

And yes, I absolutely do know that selfishness and betrayal and cycles there of tend to cycle back and devour everyone. This is not some sort of hidden secret. It's a logical and scientific fact. But I also know that this is just how humans are. It's a simple fact that humans are not intellectual creatures of pure logic. We are when it comes down to it creatures that struggle to use logic in order to temper emotion and instinct. You are like this, I am like this, we are all like this. And thus we are going to go down that road to self destruction if the situation is right. Because our most basic and fundamental instincts tell us to screw the other guy and save your self.

This by the way is why I think any and all philosophy that tries to define morality systems is bullshit. It tries to be prescriptive and tell people how to behave all the while completely ignoring how they actually will behave. When a real crisis arises you can take all your moral codes and stuff them where the sun don't shine because humans will be humans. And all your nice moral frameworks are going to collapse when faced with that reality.


PS. Excuse me if any of my language today sounds aggressive or anything. Today is very stressful for me.

[Esquire]

In that case, how do you explain all the billions of people all throughout history who've sacrificed their own best interest in order to help care for others? I don't even disagree with you, logically, but your definitions are bizarrely narrow. "Man is a political animal," after all. We get measurable chemically-induced pleasure and satisfaction out of doing nice things for others. Or, at least, most of us do; it's perfectly all right not to experience the world in the same way, but stop pretending you understand humanity in general.

No doubt you're about to point out that this isn't morality, it's biology. So what? Philosophy predates science, and explains the human experience instead of the underlying mechanisms. Just because there's a neurotransmitter that makes "treat others as you wish to be treated" work doesn't mean it doesn't. You might also say that the above argument is invalid because it's simply a better-understood self-interest at work. This is precisely my point.

[Purple]

In that case, how do you explain all the billions of people all throughout history who've sacrificed their own best interest in order to help care for others? I don't even disagree with you, logically, but your definitions are bizarrely narrow. "Man is a political animal," after all. We get measurable chemically-induced pleasure and satisfaction out of doing nice things for others. Or, at least, most of us do; it's perfectly all right not to experience the world in the same way, but stop pretending you understand humanity in general.

There is a lot to say here. But I don't have too much time so I won't go super deep into the detail of my thoughts right now.

1. Compare the number of people who sacrificed their "best interests" to those that sacrificed to the point of death or having their lives destroyed. It is my position that #1 is healthy behavior for us as a social species and #2 is not. And my initial argument was entirely pitted against #2.

I would, to paraphrase my self if faced with the option of having #2 for everyone in the world or #2 for everyone in the world but me pick the option that includes "but me".

2. Notice that when someone performs an act of self sacrifice as extreme as #2 and dies he is no longer capable of producing children. Evolution thus selects against #2.

3. I find people who deny this and try to pretend they would commit #2 to be either hypocritical or unaware of their own nature due to a lack of experience.

4. I find the western world in general to have a massive lack of experience with such issues as realistically it has been several generations since the last mass starvation, war or any other serious crisis that posed a serious threat to it. Really the last proper situation that fits that description for Europe is WW2 and for the united states is the civil war.* This all leads western people to be inherently predisposed to having their heads up their own behind about their superior morality without ever having to face a proper test of it.

* Barring various small scale local disasters like hurricane Catrina that only effect a miniscule part of the population and thus do not change the mainstream.

No doubt you're about to point out that this isn't morality, it's biology. So what? Philosophy predates science, and explains the human experience instead of the underlying mechanisms. Just because there's a neurotransmitter that makes "treat others as you wish to be treated" work doesn't mean it doesn't. You might also say that the above argument is invalid because it's simply a better-understood self-interest at work. This is precisely my point.

Basically my point on philosophy is that I fundamentally dislike any attempt at forming a morality system that is based on prescription rather than description. And I do this for the same reason why I would dislike a similar approach to physics, chemistry or any other scientific discipline. If you base your work on your opinion of how things should be as opposed to observations of how things are the best you can hope for is fictional literature.

Now I don't mind fiction my self. And I actually enjoy reading philosophy my self. It's fun. But what I do mind is when people start taking it seriously and treat it as science. And to treat any codified morality system based on prescription as something to be observed or enforced is very much akin treating a book of fiction as the one true word of god that describes how the world came to be, men were born and the universe functions. So when ever anyone calls me out on my "morality" citing some vague ethical system of his own design or that he borrowed from others I really can only roll my eyes.

This is why I have infinitely more respect for psychology and other sciences that try and actually describe and analyze human behavior and give us something functional to work with.

[Esquire]

I'd like to apologize; I was tired and misunderstood your argument slightly. I thought it was a lot less nuanced and well-reasoned than it actually is.

1. Compare the number of people who sacrificed their "best interests" to those that sacrificed to the point of death or having their lives destroyed. It is my position that #1 is healthy behavior for us as a social species and #2 is not. And my initial argument was entirely pitted against #2.

I would, to paraphrase my self if faced with the option of having #2 for everyone in the world or #2 for everyone in the world but me pick the option that includes "but me".

2. Notice that when someone performs an act of self sacrifice as extreme as #2 and dies he is no longer capable of producing children. Evolution thus selects against #2.

These two lines of reasoning seem fairly closely tied together to me, so I'll address them together. I actually don't disagree with either point, although I'd want to modify your second slightly - if someone sacrifices themselves before they've reproduced, then they're out of the gene pool. Nearly all examples of self-sacrifice seem to be parents sacrificing themselves for their children, or others sacrificing themselves for people close enough to be included in the ape-brain's percieved 'family' - soldiers and their squadmates, for example. There's a spectrum of altruism, if you like, and it will quite naturally be much more commonplace and powerful the close someone is to the altruist.

3. I find people who deny this and try to pretend they would commit #2 to be either hypocritical or unaware of their own nature due to a lack of experience.

Again, since you're talking about self-sacrifice rather than general altruism, no argument here... except with the caveats mentioned above, since people fairly regularly do - or did, anyway, in more dangerous times - sacrifice themselves for others. Closely related or dearly loved others, overwhelmingly, but others nonetheless.

4. I find the western world in general to have a massive lack of experience with such issues as realistically it has been several generations since the last mass starvation, war or any other serious crisis that posed a serious threat to it. Really the last proper situation that fits that description for Europe is WW2 and for the united states is the civil war.* This all leads western people to be inherently predisposed to having their heads up their own behind about their superior morality without ever having to face a proper test of it.

Absolutely, I've said the same myself quite often.

Basically my point on philosophy is that I fundamentally dislike any attempt at forming a morality system that is based on prescription rather than description. And I do this for the same reason why I would dislike a similar approach to physics, chemistry or any other scientific discipline. If you base your work on your opinion of how things should be as opposed to observations of how things are the best you can hope for is fictional literature.

Now I don't mind fiction my self. And I actually enjoy reading philosophy my self. It's fun. But what I do mind is when people start taking it seriously and treat it as science. And to treat any codified morality system based on prescription as something to be observed or enforced is very much akin treating a book of fiction as the one true word of god that describes how the world came to be, men were born and the universe functions. So when ever anyone calls me out on my "morality" citing some vague ethical system of his own design or that he borrowed from others I really can only roll my eyes.

This is why I have infinitely more respect for psychology and other sciences that try and actually describe and analyze human behavior and give us something functional to work with.

I haven't read a lot of modern philosophy - I think it's narrow-minded and incredibly dull for the most part - but I've got a fairly good grounding in the ancients up through the mid-19th century. Let's take Aristotle as an example; his ethics basically goes "I've noticed that most people want to be happy, and mostly aren't. Here's some general thoughts as to why that is and how you might go about being happier, maybe it'll help." Even if individual sentences are prescriptive, the project is descriptive, or at least deductive. His advice is, essentially, to live moderately and continually try to improve oneself. Does that fall under your idea of prescriptive ethics? Left to my own devices I'd say it isn't, or at least it's much less prescriptive than a religious morality.

[Purple]

I'd like to apologize; I was tired and misunderstood your argument slightly. I thought it was a lot less nuanced and well-reasoned than it actually is.

The issue is probably in my writing. These arguments usually start with me posting something and having no intention to start an argument and than people jump on it and I feel the need to defend my self and explain things. And than things get stuck as people misinterpret what I say and I entangle my self trying to correct them and my lack of mastery of the finer nuances of the English language under stress jumps out.

These two lines of reasoning seem fairly closely tied together to me, so I'll address them together. I actually don't disagree with either point, although I'd want to modify your second slightly - if someone sacrifices themselves before they've reproduced, then they're out of the gene pool. Nearly all examples of self-sacrifice seem to be parents sacrificing themselves for their children, or others sacrificing themselves for people close enough to be included in the ape-brain's percieved 'family' - soldiers and their squadmates, for example. There's a spectrum of altruism, if you like, and it will quite naturally be much more commonplace and powerful the close someone is to the altruist.

True. I absolutely agree with you on the idea that there is a spectrum of value along which we position other people and that the spectrum of our willingness to sacrifice something is directly related to it.

The issue I have really is that people take this and extrapolate from it into some sort of abstract set of morality guidelines that they than apply to perfect strangers. And than they accuse me of being evil for putting my own welfare before that of total strangers. And abstract total strangers at that.

The whole thing really started with me saying words to the effect of: If you want to automate jobs away at least skip mine. And than people called me evil for it.

I haven't read a lot of modern philosophy - I think it's narrow-minded and incredibly dull for the most part - but I've got a fairly good grounding in the ancients up through the mid-19th century. Let's take Aristotle as an example; his ethics basically goes "I've noticed that most people want to be happy, and mostly aren't. Here's some general thoughts as to why that is and how you might go about being happier, maybe it'll help." Even if individual sentences are prescriptive, the project is descriptive, or at least deductive. His advice is, essentially, to live moderately and continually try to improve oneself. Does that fall under your idea of prescriptive ethics? Left to my own devices I'd say it isn't, or at least it's much less prescriptive than a religious morality.

Basically I draw the line at the point where the ethical system becomes an end unto it self. It's difficult to explain in words but it is very easy to notice. In essence its the moment you notice the system starts having a feedback loop.

Everything is fine for as long as the system is designed based on the following algorithm: Observation of humans -> conclusion
The point where it becomes problematic is when that changes to: Observation of humans + data from the system -> conclusion

At that point any new rule is no longer based on simply the observations but also other concerns such as consistency which detract from the accuracy of the results. And over time these other concerns act like a spider web entangling ever more the work until such a point where the system it self becomes completely detached from its original purpose to describe human behavior.


The funny thing is that I do not notice this too much in actual official works as much as in the people interpreting them. It is as if the people who read this sort of thing and try to extrapolate from them only take away the words said and not the purpose of the words. Like most people who discuss these things on the internet, in real life, in politics or anywhere else. So when they try and draw new conclusions to add to the system they completely neglect observation in favor of consistency. That's where you get stupidity of the sort of that guy who in this thread said that me not wanting to sacrifice my existence for total strangers equated to failing a "morality litmus test".

[cosmicalstorm]

Some updates on the argument that vast AI progress can be made in very short time.

http://www.milesbrundage.com/blog-posts ... i-progress

AlphaGo and AI Progress

2/27/2016

1 Comment

Introduction

AlphaGo’s victory over Fan Hui has gotten a lot of press attention, and relevant experts in AI and Go have generally agreed that it is a significant milestone. For example, Jon Diamond, President of the British Go Association, called the victory a “large, sudden jump in strength,” and AI researchers Francesca Rossi, Stuart Russell, and Bart Selman called it “important,” “impressive,” and “significant,” respectively.

How large/sudden and important/impressive/significant was AlphaGo’s victory? Here, I’ll try to at least partially answer this by putting it in a larger context of recent computer Go history, AI progress in general, and technological forecasting. In short, it’s an impressive achievement, but considering it in this larger context should cause us to at least slightly decrease our assessment of its size/suddenness/significance in isolation. Still, it is an enlightening episode in AI history in other ways, and merits some additional commentary/analysis beyond the brief snippets of praise in the news so far. So in addition to comparing the reality to the hype, I’ll try to distill some general lessons from AlphaGo’s first victory about the pace/nature of AI progress and how we should think about its upcoming match against Lee Sedol.

What happened

AlphaGo, a system designed by a team of 15-20 people[1] at Google DeepMind, beat Fan Hui, three-time European Go champion, in 5 out of 5 formal games of Go. Hui also won 2 out of 5 informal games with less time per move (for more interesting details often unreported in press accounts, see also the relevant Nature paper). The program is stronger at Go than all previous Go engines (more on the question of how much stronger below).

How it was done

AlphaGo was developed by a relatively large team (compared to those associated with other computer Go programs), using significant computing resources (more on this below). The program combines neural networks and Monte Carlo tree search (MCTS) in a novel way, and was trained in multiple phases involving both supervised learning and self-play. Notably from the perspective of evaluating its relation to AI progress, it was not trained end-to-end (though according to Demis Hassabis at AAAI 2016, they may try to do this in the future). It also used some hand-crafted features for the MCTS component (another point often missed by observers). The claimed contributions of the relevant paper are the ideas of value and policy networks, and the way they are integrated with MCTS. Data in the paper indicate that the system was stronger with these elements than without them.

Overall AI performance vs. algorithm-specific progress

Among other insights that can be gleaned from a careful study of the AlphaGo Nature paper, one is particularly relevant for assessing the broader significance of this result: the critical role that hardware played in improving AlphaGo’s performance. Consider the figures below, which I’ll try to contextualize.
Picture
This figure shows the estimated Elo rating and rank of a few different computer Go programs and Fan Hui. Elo ratings indicate the expected probability of defeating higher/lower ranking opponents – so, e.g. a player with 200 points more than her opponent is expected to win about three quarters of the time. Already, we can note some interesting things. Ignoring the pink bars (which indicate performance with the advantage of extra stones), we can see that AlphaGo, distributed or otherwise, is significantly stronger than Crazy Stone and Zen, previously among the best Go programs. AlphaGo is in the low professional range (“p” on the right hand side) and the others are in the high amateur range (“d” for “dan” on the right hand side). Also, we can see that while distributed AlphaGo is just barely above the range of estimated skill levels for Fan Hui, non-distributed AlphaGo is not (distributed AlphaGo is the one that actually played against Fan Hui). It looks like Fan Hui may have won at least some, if not all, games against non-distributed AlphaGo.

I’ll say more about the differences between these two, and other AlphaGo variants, below, but for now, note one thing that’s missing from this figure: very recent Go programs. In the weeks and months leading up to AlphaGo’s victory, there was significant activity and enthusiasm (though by much smaller terms, e.g. 1-2 at Facebook) in the Go community about two Go engines – darkforest (and its variants, with the best being darkfmcts3) made by researchers at Facebook, and Zen19X, a new and experimental version of the highly ranked Zen program. Note that in January of this year, Zen19X was briefly ranked in the 7d range on the KGS Server (used for human and computer Go), reportedly due to the incorporation of neural networks. Darkfmcts3 achieved a solid 5d ranking, a 2-3 dan improvement over where it was just a few months earlier, and the researchers behind it indicated in papers that there were various readily available ways to improve it. Indeed, in the most recent KGS Computer Go tournament, according to the most recent version of their paper on these programs, Tian and Zhu said that they would have won against a Zen variant if not for a glitch (contra Hassabis who said darkfmcts3 lost to Zen - he may not have read the relevant footnote!). Computer Go, to summarize, was already seeing a lot of progress via the incorporation of deep learning prior to AlphaGo, and this would slightly reduce the delta in the figure above (which was probably produced a few months ago), but not eliminate it entirely.

So, back to the hardware issue. Silver and Huang et al. at DeepMind evaluated many variants of AlphaGo, summarized as AlphaGo and AlphaGo Distributed in the figure above. But this does not give a complete picture of the variation driven by hardware differences, which the next figure (also from the paper) sheds light on.

Picture
This figure shows the estimated Elo rating of several variants of AlphaGo. The 11 light blue bars are from “single machine” variants, and the dark blue ones involve distributing AlphaGo across multiple machines. But what is this machine exactly? The “threads” indicated here are search threads, and by looking in a later figure in the paper, we can find that the least computationally intensive AlphaGo version (the shortest bar shown here) used 48 CPUs and 1 GPU. For reference, Crazy Stone does not use any GPUs, and uses slightly fewer CPUs. After a brief search into the clusters currently used for different Go programs, I was unable to find any using more than 36 or so CPUs. Facebook’s darkfmcts3 is the only version I know of that definitely uses GPUs, and it uses 64 GPUs in the biggest version and 8 CPUs (so, more GPUs than single machine AlphaGo, but fewer CPUs). The single machine AlphaGo bar used in the prior bar, which indicated a large delta over prior programs, was based on the 40 search thread/48 CPU/8 GPU variant. If it were to show the 1 GPU/48CPU version, it would be only slightly higher than Crazy Stone and Zen - and possibly not even higher than the very latest Zen19X version, which may have improved since January.

Perhaps the best comparison to evaluate AlphaGo against would be darkfmcts3 on equivalent hardware, but they use different configurations of CPUs/GPUs and darkfmcts3 is currently offline following AlphaGo’s victory. It would also be interesting to try scaling up Crazy Stone or Zen19X to a cluster comparable to AlphaGo Distributed, to further parse the relative gains in hardware-adjusted performance discussed earlier. In short, it’s not clear how much of a gain in performance there was over earlier Go programs for equivalent hardware – probably some, but certainly not as great as between earlier Go programs on small clusters and AlphaGo on the massive cluster ultimately used, which we turn to next.

AlphaGo Distributed, in its largest variant, used 280 GPUs and 1920 CPUs. This is significantly more computational power than any prior reported Go program used, and a lot of hardware in absolute terms. The size of this cluster is noteworthy for two reasons. First, it calls into question the extent of the hardware-adjusted algorithmic progress that AlphaGo represents, and relatedly, the importance of the value and policy networks. If, as I’ve suggested in a recent AAAI workshop paper, “Modeling Progress in AI,” we should keep track of multiple states of the art in AI as opposed to a singular state of the art, then comparing AlphaGo Distributed to, e.g. CrazyStone, is to compare two distinct states of the art – performance given small computational power (and a small team, for that matter) and performance given massive computational power and the efforts of over a dozen of the best AI researchers in the world.

Second, it is notable that hardware alone enabled AlphaGo to span a very large range of skill levels (in human terms) – at the lowest reported level, around an Elo score of 2200, up to well over 3000, which in human terms is the difference between amateur and pro level skills. This may suggest (an issue I’ll return to again below) that in the space of possible skill levels, humans occupy a fairly small band. It seems possible that if this project had been carried out, say, 10 or 20 years from now, the skill level gap traversed thanks to hardware could have been from amateur to superhuman (beyond pro level) in one leap, with the same algorithmic foundation. Moreover, 10 or 20 years ago, even with the same algorithms, it would likely not have been possible to develop a superhuman Go agent using this set of algorithms. Perhaps it was only around now that the AlphaGo project made sense to undertake, given progress in hardware (though other developments in recent years also made a different, like neural network improvements and MCTS).

Additionally, as also discussed briefly in “Modeling Progress in AI,” we should take into account the relationship between AI performance and the data used for training when assessing the rate of progress. AlphaGo used a large game dataset from the KGS servers – I have not yet looked carefully at what data other comparable AIs have used to train on in the past, but it seems possible that this dataset, too, helped enable AlphaGo’s performance. Hassabis at AAAI indicated DeepMind’s intent to try to train AlphaGo entirely with self-play. This would be more impressive, but until that happens, we may not know how much of AlphaGo’s performance depended on the availability of this dataset, which DeepMind gathered on its own from the KGS servers.

Finally, in addition to adjusting for hardware and data, we should also adjust for effort in assessing how significant an AI milestone is. With Deep Blue, for example, significant domain expertise was used to develop the AI that beat Gary Kasparov, rather than a system learning from scratch and thus demonstrating domain-general intelligence. Hassabis at AAAI and elsewhere has argued that AlphaGo represents more general progress in AI than did Deep Blue, and that the techniques used were general purpose. However, the very development of the policy and value network ideas for this project, as well as the specific training regimen used (a sequence of supervised learning and self-play, rather than end-to-end learning), was itself informed by the domain-specific expertise of researchers like David Silver and Aja Huang, who have substantial computer Go and Go expertise. While AlphaGo ultimately exceeded their skill levels, the search for algorithms in this case was informed by this specific domain (and, as mentioned earlier, part of the algorithm encoded domain-specific knowledge – namely, the MCTS component). Also, the team was large –15-20 people, significantly more than prior Go engines that I’m aware of, and more comparable to large projects like Deep Blue or Watson in terms of effort than anything else in computer Go history. So, if we should reasonably expect a large team of some of the smartest, most expert people in a given area working on a problem to yield progress on that problem, then the scale of this effort suggests we should slightly update downwards our impression of the significance of the AlphaGo milestone. This is in contrast to what we should have thought if, e.g. DeepMind had simply taken their existing DQN algorithm, applied it to Go, and achieved the same result. At the same time, innovations inspired by a specific domain may have broad relevance, and value/policy networks may be a case of this. It's still a bit early to say.

In conclusion, while it may turn out that value and policy networks represent significant progress towards more general and powerful AI systems, we cannot necessarily infer that just from AlphaGo having performed well, without first adjusting for hardware, data, and effort. Also, regardless of whether we see the algorithmic innovations as particularly significant, we should still interpret these results as signs of the scalability of deep reinforcement learning to larger hardware and more data, as well as the tractability of previously-seen-as-difficult problems in the face of substantial AI expert effort, which themselves are important facts about the world to be aware of.

Expert judgment and forecasting in AI and Go

In the wake of AlphaGo’s victory against Fan Hui, much was made of the purported suddenness of this victory relative to expected computer Go progress. In particular, people at DeepMind and elsewhere have made comments to the effect that experts didn’t think this would happen for another decade or more. One person who said such a thing is Remi Coulom, designer of CrazyStone, in a piece in Wired magazine. However, I’m aware of no rigorous effort to elicit expert opinion on the future of computer Go, and it was hardly unanimous that this milestone was that long off. I and others, well before AlphaGo’s victory was announced, said on Twitter and elsewhere that Coulom’s pessimism wasn’t justified. Alex Champandard noted that at a gathering of game AI experts a year or so ago, it was generally agreed that Go AI progress could be accelerated by a concerted effort by Google or others. At AAAI last year, I also asked Michael Bowling, who knows a thing or two about game AI milestones (having developed the AI that essentially solved limit heads-up Texas Hold Em), how long it would take before superhuman Go AI existed, and he gave it a maximum of five years. So, again, this victory being sudden was not unanimously agreed upon, and claims that it was long off are arguably based on cherry-picked and unscientific expert polls.

Still, it did in fact surprise some people, including AI experts, and people like Remi Coulom are hardly ignorant of Go AI. So, if this was a surprise to experts, should that itself be surprising? No. Expert opinion on the future of AI has long been known to be unreliable. I survey some relevant literatures on this issue in “Modeling Progress in AI,” but briefly, we already knew that model-based forecasts beat intuitive judgments, that quantitative technology forecasts generally beat qualitative ones, and various other things that should have led us to not take specific gut feelings (as opposed to formal models/extrapolations thereof) about the future of Go AI that seriously. And among the few actual empirical extrapolations that were made of this, they weren’t that far off.

Hiroshi Yamashita extrapolated the trend of computer Go progress as of 2011 into the future and predicted a crossover point to superhuman Go in 4 years, which was one year off. In recent years, there was a slowdown in the trend (based on highest KGS rank achieved) that probably would have lead Yamashita or others to adjust their calculations if they had redone them, say, a year ago, but in the weeks leading up to AlphaGo’s victory, again, there was another burst of rapid computer Go progress. I haven’t done a close look at what such forecasts would have looked like at various points in time, but I doubt they would have suggested 10 years or more to a crossover point, especially taking into account developments in the last year. Perhaps AlphaGo’s victory was a few years ahead of schedule based on reported performance, but it should always have been possible to anticipate some improvement beyond the (small team/data/hardware-based) trend based on significant new effort, data, and hardware being thrown at the problem. Whether AlphaGo deviated from the appropriately-adjusted trend isn’t obvious, especially since there isn’t really much effort going into rigorously modeling such trends today. Until that changes and there are regular forecasts made of possible ranges of future progress in different domains given different effort/data/hardware levels, “breakthroughs” may seem more surprising than they really should be.

Lessons re: the nature/pace of AI progress in general

The above suggested that we should be at least slightly downgrade our extent of surprise/impressedness regarding the AlphaGo victory. However, I still think it is an impressive achievement, even if wasn’t sudden or shocking. Rather, it is yet another sign of all that has already been achieved in AI, and the power of various methods that are being used.

Neural networks play a key role in AlphaGo. That they are applicable to Go isn’t all that surprising, since they’re broadly applicable – a neural network can in principle represent any computable function. But AlphaGo is another sign that they can not only in principle learn to do a wide range of things, but can do so relatively efficiently, i.e. in a human-relevant amount of time, with the hardware that currently exists, on tasks that are often considered to require significant human intelligence. Moreover, they are able to not just do things commonly (and sometimes dismissively) referred to as “pattern recognition” but also represent high level strategies, like those required to excel at Go. This scalability of neural networks (not just to larger data/computational power but to different domains of cognition) is indicated by not just AlphaGo but various other recent AI results. Indeed, even without MCTS, AlphaGo outperformed all existing systems with MCTS, one of the most interesting findings here and one that has been omitted in some analyses of AlphaGo's victory. AlphaGo is not alone in showing the potential of neural networks to do things generally agreed upon as being "cognitive" - another very recent paper showed neural networks being applied to other planning tasks.

It’s too soon to say whether AlphaGo can be trained just with self-play, or how much of its performance can be traced to the specific training regimen used. But the hardware scaling studies shown in the paper give us additional reason to think that AI can, with sufficient hardware and data, extend significantly beyond human performance. We already knew this from recent ImageNet computer vision results, where human level performance in some benchmarks has been exceeded, along with some measures of speech recognition and many other results. But AlphaGo is an important reminder that “human-level” is not a magical stopping point for intelligence, and that many existing AI techniques are highly scalable, perhaps especially the growing range of techniques researchers at DeepMind and elsewhere have branded as “deep reinforcement learning.”

I’ve also looked in some detail at progress in Atari AI (perhaps a topic for a future blog post), which has led me to similar conclusion: there was only a very short period in time when Atari AI was roughly in the ballpark of human performance, namely around 2014/2015. Now, median human-adjusted performance across games is well above 100%, and the mean is much higher – around 600%. There is only a small number of games in which human-level performance has not yet been shown, and in those where it has, super-human performance has usually followed soon after.

In addition to lessons we may draw from AlphaGo's victory, there are also some questions raised: e.g. what areas of cognition are not amenable to substantial gains in performance achieved through huge computational resources, data, and expert effort? Theories of what's easy/hard to automate in the economy abound, but rarely do such theories look beyond the superficial question of where AI progress has already been, to the harder question of what we can say in a principled way about easy/hard cognitive problems in general. In addition, there's the empirical question of which domains there exist sufficient data/computational resources for (super)human level performance in already, or where there soon will be. For example, should we be surprised if Google soon announced that they have a highly linguistically competent personal assistant, trained in part from their massive datasets and with the latest deep (reinforcement) learning techniques? That's difficult to answer. These and other questions, including long-term AI safety, in my view, call for more rigorous modeling of AI progress across cognitive/economically-relevant domains.

The Lee Sedol match and other future updates

Picture
In the spirit of model-based extrapolation versus intuitive judgments, I made the above figure using the apparent relationship between CPUs and Elo scores in DeepMind’s scaling study. I extended the trend out to the rough equivalent of 5 minutes of calculation per move, closer to what will be the case in the Lee Sedol match, as opposed to 2 seconds per move as used in the scaling study. This assumes returns to hardware remain about the same at higher levels of skill (which may not be the case, but as indicated in the technology forecasting literature, naive models often beat no models!). This projection indicates that just scaling up hardware/giving AlphaGo more time to think may be sufficient to reach Lee Sedol-like performance (in the upper right, around 3500). However, this is hardly the approach DeepMind is banking on – in addition to more time for AlphaGo to compute the best move than in their scaling study, there will also be significant algorithmic improvements. Hassabis said at AAAI that they are working on improving AlphaGo in every way. Indeed, they’ve hired Fan Hui to help them. These and other considerations such as Hassabis’s apparent confidence (and he has access to relevant data, like current-AlphaGo’s performance against October-AlphaGo) suggest AlphaGo has a very good chance of beating Lee Sedol. If this happens, we should further update our confidence of the scalability of deep reinforcement learning, and perhaps of value/policy networks. If not, it may suggest some aspects of cognition are less amenable to deep reinforcement learning and hardware scaling than we thought. Likewise if self-play is ever shown to be sufficient to enable comparable performance, and/or if value/policy networks enable superhuman performance in other games, we should similarly increase our assessment of the scalability and generality of modern AI techniques.

One final note on the question of "general AI." As noted earlier, Hassabis emphasized the purported generality of value/policy networks over the purported narrowness of Deep Blue's design. While the truth is more complex than this dichotomy (remember, AlphaGo used some hand-crafted features for MCTS), there is still the point above about the generality of deep reinforcement learning. Since DeepMind's seminal 2013 paper on Atari has been applied to a wide range of tasks in real-world robotics as well as dialogue. As in the Atari case, there is reason to think that the methods discussed here are fairly general purpose, given the range of domains to which they have been successfully applied with minimal or no hand-tuning of the algorithms. However, in all the cases discussed here, progress so far has largely been toward demonstrating general approaches for building narrow systems rather than general approaches for building general systems. Progress toward the former does not entail substantial progress toward the latter. The latter, which requires transfer learning among other elements, has yet to have its Atari/AlphaGo moment, but is an important area to keep an eye on going forward, and may be especially relevant for economic/safety purposes. This suggests that an important element of rigorously modeling AI progress may be formalizing the idea of different levels of generality of operating AI systems (as opposed to the generality of the methods that produce them, though that is also important). This is something I'm interested in possibly investigating more in the future and I'd be curious to hear people's thoughts on it and the other issues raised above.


[1] The 15 number comes from a remark by David Silver in one of the videos on DeepMind’s website. The 20 number comes from the number of authors on the relvant Nature paper.

[SolarpunkFan]

Here's an update, the World record player of Go has been beaten two times so far by AlphaGo. It has to play three more matches.

http://www.slate.com/blogs/future_tense ... in_go.html

Make that, A.I. 2, human Go champion 0. Early Thursday morning Eastern time, Google’s AlphaGo software once again defeated 18-time world champion Lee Sedol in the second game of their five-game Go match in Seoul.

According to the Verge’s Sam Byford, reporting from the scene, the second game was “a gripping battle that saw Lee resign after hanging on in the final period of … overtime, which gave him fewer than 60 seconds to carry out each move.”

At least, human onlookers thought it was a gripping battle. To AlphaGo itself, however, it appears to have been something of a snoozefest. AlphaGo became confident of victory as early as halfway through the match, according to Demis Hassabis, head of Google’s DeepMind artificial intelligence division—at a time when the professional commentators couldn’t even tell who was ahead. Lee himself called it “a very clear loss on my part," in a statement provided by Google.

Possible caveat in there, but time will tell.

[SolarpunkFan]

Game three has been won by AlphaGo. Two more matches to go.

http://www.npr.org/sections/thetwo-way/ ... h-with-a-i

A Google A.I. program has beaten a master Go player — not once, not twice, but three times, clinching the best-of-five match between a computer and a human playing a notoriously complex game.

Lee Sedol, a 9-dan professional player considered to be one of the world's top Go players, expressed stunned resignation at the post-game press conference.

"I don't know what to say," he said through an interpreter. "I kind of felt powerless."

"I lost with quite a bit of, I guess, futility," he said later.

But he took personal responsibility for the match, held in his native South Korea — and didn't want people to conclude this meant a human could never best AlphaGo.

"Today's defeat was Lee Sedol's defeat," he said. "It was not the defeat of human beings."

[Zeropoint]

"Today's defeat was Lee Sedol's defeat," he said. "It was not the defeat of human beings."

Technically true, in that there may be some Go players out there who could beat AlphaGo. But then, John Henry beat the steam drill in their match.

The machines are improving rapidly, and human's aren't improving, at least not on the scale of a human lifetime.