3 ms·

I agree that that would be impressive, but beating a human being at Go is also very, very impressive, and had been considered a high benchmark for many years. I

• dmreedy 10y ago
  The Go victory is definitely impressive, for what it was, but I think this is a pretty unfair comparison; it's an entirely different class of problem, and a strictly less difficult one at that (at least given current knowledge of the mind). It seems likely to me that strategy games are accepted as an intelligence gauge partly because they're so hard for humans to do; thinking so many plys ahead by exploring and pruning across the game tree seems really hard for humans. For some reason, Natural Language and its semantic binding (I think the notion that these two things are often treated separately is part of a lot of the bigger deltas between expectation and reality when it comes to 'AI'. Honestly, they're probably the same thing anyway) doesn't seem really hard for us. That doesn't strictly imply that 'strategy games' are strictly harder than 'natural language understanding'; in fact, as far as research goes, strategy games are easy to tackle, because we know all the rules; we designed them! And often, perfect information is available to all agents involved. It's about as black and white as problems get. Sure you can't brute-force Go before the heat death of the universe, but it still theoretically -can- be brute forced. The rules of the game (known and accepted by all parties) encode the algorithm, and it's just a matter of reverse engineering it step-by-step for a given situation. What a brain is doing when it speaks or listens or thinks is currently a much bigger, much more primitive mystery. As far as we know right now, you can't just brute force sentence permutations until you hit a correct answer to a question like the one posed by the parent. So I don't call this goal-post moving (though acknowledge that the Vanishing Problem is systemic in AI Research and Pop Science); instead, I think this is more akin to the work in say, separating problems into classes of computational complexity. Finding the upper limits of things can be very helpful when we're dealing with such an old, deep, controversial problem.
  • bbctol 10y ago
    For sure: as much as Go really was widely-talked about as an AI goal, real language processing has been the goal since, well, Turing. Even more importantly, it remains to be seen how far the current neural network approach extends, and it could still be only successful at certain tasks, with a bigger paradigm shift needed to crack language and so on. But I don't think it's worthwhile to make a hard, fundamental distinction between game-playing and language, such that one requires "real" intelligence and the other is something else. After all, we possess a single machine that can both play Go and use language, and do far more things besides, and it really does look like the brain is re-using a lot of its structures for these different problems. You can theoretically solve Go, and you can't solve language, but the brain doesn't try to solve either (and neither does Deepmind); there's at least some hope that one paradigm can be applied to both these problems, and the reason we originally turned to neural networks was because of their (admittedly highly superficial) resemblance to our one positive case of intelligence. So I do take issue with the idea that language understanding is "strictly" harder than board games; that's an artifact of the paradigm used. The idea that the number of rules and exceptions to rules is the 'objective' mark of a task's difficulty only applies to a certain kind of processing agent, and our one example of intelligence clearly uses a different one. For most of the past ten millennia, the only information processing agents on the planet had far more ease communicating with natural language than crushing at board games; ask some now to either write an essay or beat Lee Sedol 4-1, and they'll tell you which is easier. The fact that board game mastery is easier for our AI attempts than basic speech isn't a sign that they've secretly been easier this whole time; for a certain kind of processing agent, communicating in a poorly-defined space has been much easier than mastering a clearly-defined one, and if our AIs don't match that pattern, it's only another reason to wonder if we're on a different path.
    • dmreedy 10y ago
      The 'kind of processing agent' emphasis is a very good one, and I think worth bringing up in every one of these conversations. That said, I want to niggle a little bit on some of these points There certainly does seem to be component re-use in certain parts of the brain between seemingly disparate tasks. The left brain/right brain divide has been largely debunked, and I'm not entirely sure what the current state of the art is, but I'm not sure I'd go so far as to say that that re-use in this particular case (if it exists, I don't know) is anything other than a 'best approximation' (and maybe not even best) with the available hardware. Like using a fork to eat soup. Sure you can kinda do it, but you're always gonna be thinner than the fellow with the spoon. That suggests to me then that using games as a metric of intelligence is actually a pretty poor way of getting at the deeper issue; there are too many other variables beyond "how much (general/amalgamated) smarts have you got?". It might be semi-convenient when you have cause to believe that the two models you're testing have largely similar structure (i.e., two humans), but it gets pretty noisy when you've got a less than solid idea how similar the two models are. Because there is an obvious, perfect solution to the problem, you end up with a kind of algorithmic continuum, with the system you're trying to model on one end and perfect brute force on the other. And you don't know how far along the continuum your model is; it might be pretty close to the system (the brain), in which case, hey! Progress. But it might be closer to the perfect solution, and thus closer to dedicated hardware for the problem at hand, instead of adapted hardware doing its best. It might have a spork, which makes it rather less worthy of admiration when it comes to soup-eating. Especially when you see it struggle with the steak. So, in short; I think it actually -might- be a sign that strategy has secretly been easier this whole time. I'm not even sure it's that much of a secret. At least when it comes to the particular processing agents in question, the ones capable of general intelligence as we understand it. If you want to get into philosophical conversations about Hyperintelligent Shades of Blue, well, that's another story.