Interesting article! However, Deep Blue, Watson, and AlphaGo are very different from one another. I don't think anyone deemed beating humans at chess or jeopardy impossible at the time Deep Blue and Watson were built. On AlphaGo the point about generalizing AI is valid, yet I think the author doesn't fully appreciate the novelty of approach described in the AlphaGo paper. Their work advances the field, has more general utility than Deep Blue's chess or Watson's jeopardy programs. AlphaGo paper specifically represents an advance in machine learning algorithms for games in general. As I understand, Watson's new NLP algorithm is PRISMATIC [1]. PRISMATIC is a rule-based NLP system and AlphaGo is more statistical inference/neural networks. Even if AlphaGo's 'policy-network- value-network' framework is not too generalizable, the philosophical implication is that we can build AIs that can mimic 'human intuition'. Jeopardy and chess have lesser components of 'human intuition' than Go. They are apples and oranges. So, I wonder if the author errs in bringing Deep Blue and Watson-Jeopardy into the picture.
In my opinion while Watson-Jeopardy and Deep Blue were 'over-fitting' for Jeopardy and chess respectively, AlphaGo algorithms are more general and over-fits for the larger category of 'games'.
Traditional AI suffers from a love of solving parlor tricks. Solving tick tack toe, checkers, chess, poker, Jeopardy(tm), are parlor tricks. It seems important because frankly humans just suck at parlor trick type problems and other forms of intelligence like say a cat brain don't even get parlor tricks. SO then we found that computers were really good at parlor tricks and it seemed like we were really onto something here. But nope.
On the other hand playing Go is not really a parlor trick, it's actually hard simple symbolic logic totally fails to grasp the problem at the first level.
If a problem is too difficult to solve, then we solve simpler related problems first. In doing so, we typically gain insight into the more challenging problem. For example, before Calculus was discovered, areas and volumes were all computed in ad hoc ways, via "parlor tricks". Only by generalizing the insights gleamed from some of these "tricks" did we stumble upon Calculus.
Problems such as chess certainly did not seem like "parlor tricks" at the time they were proposed. In fact, many thought of chess play as being an idyllic example of human intelligence. Just because we do not understand how to solve Go today doesn't mean that it won't be a trivial "parlor trick" in ten years.
Well that's the crux of the problem - the corollary to your question is "can you provide a firm definition of intelligence?" Nobody can yet, so it's all speculation and subjective opinion.
The reason that I personally consider all these things parlor tricks (including, hypothetically, complete mastery of Go) is that I see no path from these particular types of systems to general intelligence. A human can take in arbitrary sensory data and make all sorts of conclusions and associations with it. Does this particular system have the capability to get to a point where it can see an apple falling and posit a theory of gravity? Will it ever be able to read subtle cues in facial/oral/bodily expression and combine them with all sorts of other data, instantaneously, to achieve compelling real-time social interaction? Will this system ever invent the game of Go, or anything else, because it felt like it? No, it has absolutely no framework to do any of those things, or countless other things humans can do. It's a machine built with a single purpose in mind, and it can only serve that purpose. It's a glorified function call. I don't think this type of machine will just wake up one day after digging deeper and deeper into these "hard" tasks. We need breadth, not depth.
Parlor trick is something that looks interesting only as long as you make incorrect assumptions about how it's being done. For example, a card that seems to appear out of nowhere while in reality it is held behind the hand between little and index finger. Or a program that seems to engage in planning when in reality it simply iterates over all possible solutions.
If AlphaGo eventually beats humans at go I think it says more about the game (or the way humans play the game) than about AI in general. Its strength comes from mcts with better heuristics. Previous ConvNet work posted here https://news.ycombinator.com/item?id=10558831 was already impressive in generating human like moves and combining it with search was an obvious next step. Unbeknownst to all Google's DeepMind team had already succeeded in doing so at that time.
> If AlphaGo eventually beats humans at go I think it says more about the game (or the way humans play the game) than about AI in general.
AI problems are always mysterious until they're solved. People say this sort of thing every time an AI achieves some task which had previously been deemed impossible.
> AI problems are always mysterious until they're solved.
There's truth there but this line is over-used. Playing chess better than a human was never deemed impossible and the ultimate solution doesn't seem to have any crazy insights that would stun a researcher from the 1960s. (You can read the source for Stockfish which is state of the art and open source). The improvements came in the form of more and more horsepower through hardware, more and more efficient innovations in the tree search very specific to the game of chess (bit board representation, move ordering), and simplifying evaluation and tweaking parameters according to unsupervised learning. Correct me if I'm wrong?
At the end of the day, chess (and go) are discrete games with perfect information played one turn at a time according to a tree of simple and trivially determined possible moves with clear criteria for winning. I don't see why we'd put this on a pedestal as the example of something generally considered uniquely human, so we'd better expand our imaginations of what we can achieve with AI. As the parent said, the mistake may have been in supposing there wasn't a solvable evaluation function for Go positions when in fact, through some more human ingenuity, there is.
So what happens if (when?) we come up with a system that does everything a human can do, better than a human, but doesn't contain any 'crazy insights', just a bunch of incremental improvements on what we have now?
Does that mean we aren't intelligent? Or does it mean that the system isn't intelligent but that we are "because we do it differently"? Or do we accept at that point that intelligence is composed of simple building blocks interacting in complex ways (which we already know, if we eschew Cartesian duelism)?
We would declare it intelligent? Certainly the people of today would call it intelligent. What I suspect you are claiming is that the people of that future would not call it intelligent, and this is the basis for arguing why that objection is not valid today. But that extrapolation to the future is just your speculation.
Or, in my view the most likely, that isn't in fact possible and the system we eventually arrive at that does that, _will be_ extremely different from what we have now.
Although I actually think that we'll never make a system that does _everything_ a human can do at all, simply because that would be silly.
And of course, there also has never been a human who can do everything that humans can do, so this bar is way too high anyway.
Perfect information isn't true, you don't exactly know the opponents next move. This broadens the search tree exponentially. Generally, with many hard problems, the size of the problem is a problem, when memory is limited.
Playing chess better than humans was certainly deemed impossible by some people (not the ones who wrote chess programs, of course). See for example Hubert Dreyfus:
While progress came not as quickly as AI researchers (or their universities' publicity departments) had hoped, computers can now do a lot of the things that he wrote about for example in "What computers can't do".
Dreyfus does not appear to have claimed that computers would never be able to play chess well. At least, not in that book.
He reacted with skepticism when Newell and Simon said in 1957 that a computer would be world chess champion by 1967 and, well, he was right to.
He said that the computational techniques in use for computer chess in the 1970s wouldn't be capable of producing a world-class player, and he was probably wrong about that -- largely, I guess, because he didn't foresee how big an impact a performance improvement of ~10000x could have.
If he actually claimed that playing chess better than humans was impossible, can you say where?
There is another way of saying the same thing. A lot of seemingly groundbreaking progress in AI usually happens when people (people, not machines) discover a clever way of mapping a new unsolved problem to another - well-solved - problem. That's a valid and insightful observation that you shouldn't dismiss with such an ease.
Why 'seemingly'? Why is that not 'actual' groundbreaking progress? The achievement of any level of AI will by necessity require the chaining together of processes which are not themselves 'intelligent'. It has to be bootstrapped somehow.
On the day they build a walking, talking AI who can converse fully with a human, write a symphony, design a building, feel and express emotions and all the rest of the things we define as being essentially in the domain of humain intelligence, everyone will say, "But of course, none of those things required intelligence at all. This is all an elaborate collection of illusions and ugly hacks."
And they'll be right, but I suspect that the brain is the same way.
Why 'seemingly'? Why is that not 'actual' groundbreaking progress?
Because mapping problems to pre-existing algorithms is bread and butter of computer science and software engineering. To be groundbreaking a work needs to change our understanding of the underlying issues. In a lot of popularized cases that does not happen.
i tend to agree with that. it does seem like there's some general instinct in people that other things (animals, "artificially" intelligent machines) could possess human-like intelligence and subjective experience. i mean, we both seem to be tacitly agreeing to that here. there's also obviously a general instinct that's completely the opposite, and i'd guess that's probably the more prevalent instinct (in the population at large and in many conflicted individuals).
i think the opinion that humans are less special than we once thought, especially on expansive time scales, will only become more widespread.
The fact that people in the past have used the line "that's not really intelligence" doesn't at all invalidate the point that is being made, which is that games like Go/Chess/etc. have little to do with the real world. The real world is exponentially harder than those games (imperfect info, imperfect sensors, infinite state space, large - if not infinite - possible moves, infinite time horizon).
If AlphaGo eventually beats humans at go I think it says more about the game (or the way humans play the game) than about AI in general.
I'd agree with the rest of your post but this statement seems to imply there's some "general level" that can't reached by the incremental approaches that yielded the results.
The progress that we see here is incremental progress in a defined area but it's also progress towards more general and easier to implement approach - it's yet "point and solve" but it might be a step toward "point and solve". Given little "general intelligence" is understood, no one can say now for certain we won't arrive at it through a series of these advances.
Let me expand a little more on what I meant and see if it makes more sense to you:
Much about Go/Weiqi is still unknown even to the top players. For example Ke Jie (the current highest ranked player) feels that the current komi (compensation given to white for playing second) is too generous and he prefers white (last year he won almost all his white games). This is why professional players seem very excited about AlphaGo. If AlphaGo can consistently beat top players it may teach us more about the game. It may discover or settle questions about josekis (pattern conventions). It may tell us what komi is fair. It could settle questions about how a particular board position should be valued. On the other hand since it uses patterns learned from human plays, this could also motivate new theories of play that could defeat past strategies. Or if no one can come up with ways to defeat AlphaGo it may indicate the valuation function it produces is approaching ideal.
But AlphaGo is very much about combining existing tools of ConvNet and MCTS, albeit in a highly innovative manner, to solve the search problem of Go. Its success or failure could teach us a lot about how amenable the Go game is to such an approach (and potentially advance theories about the Go game), more than I would say about how in general problems in AI can be solved. That is what is learned here is very specific because Go is a very visual game and/or humans play it in a very visual way to take advantage of the massive parallelism of our vision system to quickly narrow down choices.
> I'd agree with the rest of your post but this statement seems to imply there's some "general level" that can't reached by the incremental approaches that yielded the results.
At this point in time, the belief in "incremental" general AI looks like a kind of pseudo-religion emerging in IT. People go way beyond postulating it as a possibility. They fervently defend it as some kind of obvious fact and flaunt this attitude as progressive.
What really disturbs me is the way these bridge the gaping void between existing AI and biological brains. On one hand I see absolutely insane amounts of hype around artificial neural networks, exploding way beyond the optimism warranted by the actual research. On the other hand I see the insistence that biological brains are "nothing special". I wonder how deep that goes. Are these people ultimate sociopath who truly believe that everyone around them is a mere pattern-matching device?
All this bullshit is actually detrimental to the usage of "AI" algorithms as programming techniques aimed at solving real-life problems. For one, many managers look at the hype and mysticism and conclude that AI is something that is too complex for mere mortals to handle. I've seen this on many occasions.
One is whether simple progress on neural nets is enough to close-in on real intelligence and I think those who really understand neural nets generally think not.
The other issue is whether incremental progress in general can achieve AI - there I don't think anyone can be sure, especially the relative vagueness of "incremental" and so I don't one should dismiss incremental progress or uncritically assume it.
Go ranks are in "stones" that one player can give another as a handicap - except at the professional level where there's more relative equality of play and the stone-ranks are honorary and ELO ratings are more accurate.
Monte Carlo tree search had advanced computer Go from 2-3K to 6-7 dan, a gain of 8 stones. Alpha Go apparently has advanced to 9-10 stones, professional level, by using neural networks to enhance the final position and search policy used in Monte Carlo tree search.
In many ways, it seems to me that Monte Carlo Tree Search was the primary answer to how computers could deal with Go [1].
Chess essentially conquered through a combination of red-black and other smart pruning algorithms, incremental advances in hand-tuned final position policy and improved hardware [2].
So it seems like the "conquest" of Go has involved a more generalized, self-learned version of the original approach to chess (tree-search strategies plus final position heuristics). That might enough for just about any deterministic game one can find.
It should be noted that Arimaa, a game designed specifically to be hard for computers without having the large board of Go, was "conquered" last year but without any neural net techniques (apparently).
Is MCTS algorithmically interesting, or is it only powerful because it is embarrassingly parallel and so can leverage all the computing power as you have availalbe?
I really think Gary agrees with you, and you're just picking an unnecessary fight. Having advocated for this kind of approach decades ago, he fully appreciates the novelty of the approach. He points out the simplicity of Jeopardy and Chess as reasons why they haven't translated to the real world well, and maintains optimism about DeepMind. I don't read "the real question is whether the technology developed there can be taken out of the game world and into the real world" as a criticism, but praise; the question is worth considering, whereas so often it is not.
I don't intend to pick a fight. I am pointing out one deficiency I perceive in this widely shared article.
On the philosophical questions: I think even with the greatest technological advances, the question remains: will we adopt AI? I have seen too many situations where machine learning is not adopted even though algorithms can enable great functionality: usually the reason is lack of financial incentives or simple entrenched human interests (low-grade Ludditism). This resistance will lead to more (to use terminology from the AI debate) paperclip maximizers and fewer general AIs. This is already happening, the future of 2001 didn't happen to be HAL 9000, but an ingenious model and linear algebra algorithm to deliver better search results.
I see quotes to use it as an analogy. I'm not claiming that their algorithm over fits. It is just another way of saying the system is too specific and optimized to chess to be generalizable.
In my opinion while Watson-Jeopardy and Deep Blue were 'over-fitting' for Jeopardy and chess respectively, AlphaGo algorithms are more general and over-fits for the larger category of 'games'.
[1] http://brenocon.com/watson_special_issue/05%20automatic%20kn...