Yet people will still tell that worrying about AI taking over is like worrying about overpopulation on Mars, and that this is a problem at least 50 years out.
Edit: I do understand that the techniques used to implement Alphago can be used to implement other single-function solvers. That doesn't make it a general purpose strong AI.
As far as I know there is nothing particularly novel about AlphaGo, in the sense that if we stuck an AI researcher from ten years ago in a time machine to today, the researcher would not be astonished by the brilliant new techniques and ideas behind AlphaGo; rather, the time-traveling researcher would probably categorize AlphaGo as the result of ten years' incremental refinement of already-known techniques, and of ten years' worth of hardware development coupled with a company able to devote the resources to building it.
So if what we had ten years ago wasn't generally considered "true AI", what about AlphaGo causes it to deserve that title, given that it really seems to be just "the same as we already had, refined a bit and running on better hardware"?
By that standard there's nothing particularly novel about anything. Everything we have today is just a slight improvement of what we already had yesterday.
World experts in go and ML as recently as last year thought it would be many more years before this day happened. Who are you to trivialize this historic moment?
A light bulb is just a metal wire encased in a non-flammable gas and you run electricity through it. It was long known that things get hot when you run electricity through them, and that hot things burst into fire, and that you can prevent fire by removing oxygen, and that glass is transparent. It's not a big deal to combine these components. A lot of people still celebrate it as a great invention, and in my opinion it is! Think about how inconvenient gas lighting is and how much better electrical light is.
Same thing with AlphaGo. Sure, if you break it down to its subcomponents it's just clever application of previously known techniques, like any other invention. But it's the result that makes it cool, not how they arrived at it!
All algorithms are incremental improvements of existing techniques. This isn't a card you can use to diminish all progress as "just a minor improvement what's the fuss".
People have used neural nets as function approximators for reinforcement learning with MCTS for game playing well before AlphaGo (!!).
Your lightbulb example actually supports my point. The lightbulb was the product of more than a half-century of work by hundreds of engineers/scientists. I have no problem with pointing to 70 years of work as a breakthrough invention.
A decade ago I was trying and failing to build multi-layer networks with back-propagation-- it doesn't work so well. More modern, refined, training techniques seem to work much better... and today tools for them are ubiquitous and are known to work (especially with extra CPU thrown at them :) ).
These things might seem like "small iterative refinements", but they add up to 100x improvement. Even when you don't consider hardware. And you should consider hardware too, it's also a factor in the advancement of AI.
Also reading through old research, there is a lot of silly ideas along with the good ones. It's only in retrospect that we know this specific set of techniques work, and the rest are garbage. At the time it was far from certain what the future of NNs would look like. To say it was predictable is hindsight bias.
Today lots of people-- ones with even less background and putting in less effort-- try and are successful.
This is not a small change, even if it is the product of small changes.
10 years ago no one believed it was possible to train deep nets[1].
It wasn't until the current "revolution" that people learned how important parameter initialization was. Sure, it's not a new algorithm, but it made the problem tractable.
So far as algorithmic innovations go, there's always ReLU (2011) and leaky ReLU (2014). The one-weird-trick paper was pretty important too.
[1] Training deep multi-layered neural networks is known to be hard. The standard learning strategy— consisting of randomly initializing the weights of the network and applying gradient descent using backpropagation—is known empirically to find poor solutions for networks with 3 or more hidden layers. As this is a negative result, it has not been much reported in the machine learning literature. For that reason, artificial neural networks have been limited to one or two hidden layers
http://deeplearning.cs.cmu.edu/pdfs/1111/jmlr10_larochelle.p...
If you asked people 10 years ago before the moon landing if it was possible, I too would agree it's impossible. But after that breakthrough it opened up the realm is possibilities.
I see AlphaGo more of an incremental improvement than a breakthrough.
I'm generally considered to be way over optimistic in my assessment of AI progress. But wow.. that's pretty optimistic!
It's still an incredibly achievement - but it's important to be accurate.
While the higher portions do share some similarities with the Atari system, at a basic level this is a machine that was designed and trained to play Go. AlphaGO is 'essentially the same' as the Atari system in the same way that all Neural Networks are 'essentially the same.'
Is this an extremely impressive accomplishment? Yes. However, doesn't seem to qualify as anything close to generalizable.
[1] http://googleresearch.blogspot.com/2016/01/alphago-mastering...
Likewise NNs are uncaring what application you put them into. Give them a different input and a different goal, and they will learn to do that instead. Alphago gave it's NN's control over a monte carlo search tree, and that turned out to be enough to beat Go. They could plug the same AI into a car and it would learn to control that instead.
Note that even without the monte carlo search system, it was able to beat most amateurs, and predict the moves experts would make most of the time.
The best Go program before AlphaGo was CrazyStone, ranked at 5-dan ("high amateur" range).
The Atari-playing AI watched the pixels indeed, but it was also given a set of actions to choose from and more importantly, a reward representing the change in the game score.
That means it wasn't able to learn the significance of the score on its own, or to generalise from the significance of the changing score in one game, to another.
It also played Atari games, that _have_ scores, so it would have been completely useless in situations where there is no score, or a clear win/loss situation.
AlphaGo is also similarly specialised to play Go. As is machine learning in general: someone has to tell the algorithm what it needs to learn, either through data engineering, or reward functions etc. A general AI would learn what is important on its own, like humans do, so machine learning has not yet shown that it can develop into AGI.
Even humans have utility functions. For example, we get rewards for having sex, or eating food, or just making social relationships with other humans. Or we have negative reinforcement from pain, and getting hurt, or getting rejected socially.
You can come up with more complicated utility functions. Like instead of beating the game, it's goal could be to explore as much of the game as possible. To discover novel things in the game. Kind of like a sense of boredom or novelty that humans have. But in the end it's still just a utility function, it doesn't change how the algorithm itself works to achieve it. AGI is entirely agnostic to the utility function.
No, what I'm really saying is that you can't have an autonomous agent that needs to be told what to do all the time. In machine learning, we train algorithms by giving them examples of what we want them to learn, so basically we tell them what to learn. And if we want them to learn something new, we have to train them again, on new data.
Well, that's not conducive to autonomous or "general" intelligence. There may be any number of tasks that your "general" AI will need to perform competently at. What's it gonna do? Come back to you and cry every time it fails at something? So then you have a perpetual child AI that will never stand on its own two feet as an adult, because there's always something new for it to learn. Happy little AI, for sure, but not very useful and not very "general". Except for a general nuisance, maybe.
Edit: I'm saying that machine learning can't possibly lead to general AI, because it's crap at learning useful things on its own.
There is also unsupervised and semi-supervised learning, which can take advantage of unlabelled data. Even supervised learning can work really well on weakly labelled data. E.g. taking pictures from the internet and using the words that occur next to them as labels. As opposed to hiring a person to manually label all of them.
I don't know what situation you are imagining that would make the AI "come back and cry". You will need to give an example.
Of course they did. They trained it with examples of Go games and they also programmed it with a reward function that led it to select the winning games. Otherwise, it wouldn't have learned anything useful.
>> There is also unsupervised and semi-supervised learning, which can take advantage of unlabelled data.
Sure, but unsupervised learning is useless for learning specific behaviours. You use it for feature discovery and data exploration. As to semi-supervised learning, it's "semi" supervised: it learns its own features, then you train it with labels so that it learns a mapping from those features it discovered to the classes you want it to output.
>> I don't know what situation you are imagining that would make the AI "come back and cry"
That was an instance of humour [1].
Yes, but it doesn't need to be trained with examples of Go games. It helps a lot, but it isn't 100% necessary. It can learn to play entirely through self play. The atari games were entirely self play.
As for having a reward function for winning games, of course that is necessary. Without a reward function, any AI would cease to function. That's true even of humans. All agents need reward functions. See my original comment.
>That was an instance of humour
Yes I know what humour is lel. I asked you for a specific example where you think this would matter. Where your kind of AI would do better than a reinforcement learning AI.
That's reinforcement learning and it's even more "telling the computer what to do" than teaching it with examples.
Because you're actually telling it what to do to get a reward.
>> Without a reward function, any AI would cease to function.
I can't understand this comment, which you made before. Not all AI has a reward function. Specific algorithms do. "All" AI? Do you mean all game-playing AI? Even that's stretching it, I don't remember minimax being described in terms of rewards say, and I certainly haven't heard any of about a dozen classifiers I've studied and a bunch of other systems of all sorts (not just machine learning) being described in terms of rewards either.
Unless you mean "reward function" as the flip side of a cost function? I suppose you could argue that- but could you please clarify?
>> your kind of AI
Here, there's clearly some misunderstanding because even if I have a "my kind" of AI, I didn't say anything like that.
I'm sorry if I didn't make that clear. I'm not trying to push some specific kind of AI, though of course I have my preferences. I'm saying that machine learning can't lead to AGI, because of reasons I detailed above.
No one tells the computer what to do. They just let it do it's thing, and give it a reward when it succeeds.
>Not all AI has a reward function. Specific algorithms do. "All" AI?
Fine, all general AI. Like game playing etc. Minimax isn't general, and it does require a precise "value function" to tell it how valuable each state is. Classification also isn't general, but it also requires precise loss function.
Sure they do. Say you have a machine learning algorithm, that can learn a task from examples, and let's notate it like so:
y = f(x)
Where y is the trained system, f the learning function and x the training examples.
The "x", the training examples, is what tells the computer what to learn, therefore, what to do once it's trained. If you change the x, the learner can do a different y. Therefore, you're telling the computer what to do.
In fact, once you train a computer for a different y, it may or may not be really good at it, but it certainly can't do the old y anymore. Which is what I mean by "machine learning can't lead to AGI". Because machine learning algorithms are really bad at generalising from one domain to another, and the ability to do so is necessary for general intelligence.
Edit: note that the above has nothing to do with supervised vs unsupervised etc. The point is that you train the algorithm on examples, and that necessarily removes any possibility of autonomy.
>> Fine, all general AI. Like game playing etc.
I'm still not clear what you're saying; game-playing AI is not an instance of general AI. Do you mean "general game-playing AI"? That too doesn't always necessarily have a reward function. If I remember correctly for instance, Deep Blue did not use reinforcement learning and Watson certainly does not (I got access to the Watson papers, so I could double-check if you doubt this).
Btw, every game-playing AI requires a precise evaluation function. The difference with machine-learned game-playing AI is that this evaluation function is sometimes learned by the learner, rather than hard-coded by the programmer.
>The "x", the training examples, is what tells the computer what to learn, therefore, what to do once it's trained. If you change the x, the learner can do a different y. Therefore, you're telling the computer what to do.
But with RL, a computer can discover it's own training examples from experience. They don't need to be given to it.
>I'm still not clear what you're saying; game-playing AI is not an instance of general AI.
But it is! The distinction between the real world and a game is arbitrary. If an algorithm can learn to play a random video game, you can just as easily plug it into a robot and let it play "real life". The world is more complicated, of course, but not qualitatively different.
People said for years that Go would never be beaten in our lifetime. They said this because Go has a massive search space. It can't be beaten by brute force search. It requires intelligence, the ability to learn and recognize patterns.
And it requires doing that at the level of a human. A brute force algorithm can beat humans by doing a stupid thing far faster than a human can. But a pattern recognition based system has to beat us by playing the same way we do. If humans can learn to recognize a specific board pattern, it also has to be able to learn that pattern. If humans can learn a certain strategy, it also has to be able to learn that strategy. All on it's own, through pattern recognition.
And this leads to a far more general algorithm. The same basic algorithm that can play Go, can also do machine vision, it can compose music, it can translate languages, or it could drive cars. Unlike the brute force method that only works one one specific task, the general method is, well, general. We are building artificial brains that are already learning to do complex tasks faster and better than humans. If that's not progress towards AGI, I don't know what is.
Those arguments absolutely are wrong. For one thing it's classic hindsight bias. When you make a wrong prediction, you should update your model, not come up with justifications why your model doesn't need to change.
But second, it's another bias, where nothing looks like AI, or AI progress. People assume that intelligence should be complicated, that simple algorithms can't have intelligent behavior. That human intelligence has some kind of mystical attribute that can't be replicated in a computer.
Whenever AI beats a milestone, there are a bunch of over-optimists that come out and make predictions about AGI. They have been wrong over and over again over the course of half a century. It's classic hindsight bias.
And the optimists are being proven right. AGI is almost here.
(the answer is Rocky Road by the way)
Many of the chatbot Turing test competitions have heavily rigged rules restricting the kinds of questions you're allowed to ask in order to give the bots a chance.
I know this is true, because there are already a lot of people that think the Turing test isn't valid. They believe it could be beaten by a stupid chatbot, or deception on the part of AI. Just search for past discussions on HN of the Turing test, it comes up a lot.
There is no universally accepted benchmark of AI. Let alone a benchmark for AI progress, which is what Go is.
No one claimed that Go would require a human level AI to beat. But I am claiming that beating it represents progress towards that goal. Whereas passing the Turing test won't happen until the very end. Beating games like Go are little milestones along the journey.
See this interview between Kurzweil and Minsky: https://www.youtube.com/watch?v=RZ3ahBm3dCk#action=share
We don't know that, actually. Maybe GAI isn't one shining simple algo, but cobbling together a bunch of algorithms like this one.
You're making the mistake of assuming anything about how the human brain learns.
Despite my optimism, the writing is on the wall. AlphaGo and algorithms like it will only improve as you throw more CPU time at them. I actually want Lee Sedol to win, not because it would uphold some kind of human supremacy but because I want to see the AI guys put some more effort (and CPU time) into it. It would be a real shame if they'd win on their first attempt.
This 10 years to beat go prediction shows that our time estimates are wildly ignorant.
One way of looking at the significance of this is that it might tell us that relatively simple machine learning algorithms can capture key aspects of the versatile human cortical capacity to learn things like Go using sheer pattern recognition. (It's amazing that human visual cortex can do that.) If the human brain were more mystical in its power, then human-level ability to recognize Go patterns wouldn't have been penetrable at all to a comparatively simple neural algorithm like deep learning.
From another standpoint, this could show that we're reaching the point where, when an AI algorithm starts to reach interesting levels of performance, a much-encouraged Google dumps 100,000 tons of GPU time and 5 months of a few dozen researchers' time to improve the algorithm right past the human performance level. In N years from now when it's a more interesting AI system doing more interesting cognition, we could see a more interesting result materialize when a big organization, encouraged by some key milestone, invests 5 months of time and massively more computing power to further improve an AI system.
Really?
Why not?
EDIT: clarified my language to address below reply.
http://arxiv.org/abs/1509.02971
The fact that the (reinforcement) learning problem is hard or not is not directly related to whether the observation and action spaces are discrete or continuous.
- discrete spaces such as atari games and go, - continuous spaces such as driving a car, controlling a robot or bid on a ad exchange.
When you're talking about some other kind of evaluation function, such as Monte Carlo rollouts, you usually prefix that to the tree search (in the case of Monte Carlo rollouts, "Monte Carlo tree search" or MCTS) to indicate that besides the basic fundamental task common to almost all AIs (finding the optimal branches in a decision tree) it functions completely differently from the expert systems.
So is the case with this program, which (in a first pass) approximates subtrees by a trained neural net, rather than Monte Carlo rollouts or an expert system. So using terminology that suggests classical expert system tree search is bound to cause confusion (as you noticed).
Monte Carlo Tree Search was necessary and itself a massive improvement over minimax but not sufficient for creating a Go program to challenge professional players. The true innovation here is the neural networks. Without those networks to guide it AlphaGo plays far worse than existing programs.
The fact that those networks are sufficient is pretty incredible. We already knew that by inventing them we had created a very pure form of pattern recognition, but it's surprising that the pattern recognition coupled with some tree search seems to be all you need to play Go as well as humans.
It's not impressive to you that we've now reproduced a piece of human intelligence, "intuition", which was previously considered out of reach?
What if you train an AI to play an RTS where matter, energy, and time are the resources and the goal is to take over the world?
Heh, nice one.
Is it really all we need? Or it is more that they threw a lot of hardware to it? What if if a part of its efficiency is because they threw a lot of GPUs with a huge network, rather than having a NN efficient by itself?
We see that: "AlphaGos Elo when it beat Fan Hui was 3140 using 1202 CPUs and 176 GPUs. Lee Sedol has an equivalent Elo to 3515 on the same scale (Elos on different scales aren't directly comparable). For each doubling of computer resources AlphaGo gains about 60 points of Elo."
It's a lot of hardware.
It is the same question for the data they used. Facebook, Google and others seem to agree that, at the end, the quantity and quality of data are more important than the algorithm itself. So how much is it at play here? Knowing that will be able to show us why it is performing well and how much we can appreciate their work.
Basically, the ("lots of hardware") distributed implementation gets ~3100 points in the Elo rating against ~2900. ~2900 is still sufficient to win against Fan Hui. So I would say, that yes, this algorithm has most of the merit here.
As hard as writing AI for a problem space like "how to win at Go" is, it's several orders of magnitude easier than creating a general AI with the self-awareness required to see us as a threat.
That's the difference between a neural net -- which until about an hour ago was a phrase that reliably set off my snake-oil detector -- and a simple tree search. DeepMind just beat a Go master... and nobody can say exactly how it did it.
That's a big deal.
Remember that "solve computer vision" was considered a summer project.
And they're pretty much there. Have you seen some of the latest results in that field?
We certainly got nearly there, but it was nearly 50 years later, not 3 months. Similarly, something that might look somewhat simple to us right now, might also be a lot more difficult.
Good chess programs definitely use self-play to improve themselves, eg to tune parameters and heuristics.
We should be taking steps to outlaw black box algorithms right now but I'm sure we won't.
I can't understand the call to regulate a technology when it is decades (much more that 10-20 years) away from possibly existing in a state that we can't even imagine. Add to that legislators with zero technical literacy. That's essentially advocating for shutting down all research into a technology that can improve millions of lives.
Basically any algorithm where you are training it with tons of samples and then it creates a logic tree that is not written by a human is a black box algorithm.
By the time AI technology gets to the point that we realize we should regulate it we will be too late.
"We've beaten some hard problems more quickly than expected therefore we'll likely beat other hard problems more quickly than expected" is logical induction. It's equivalent to "I just flipped a coin and got heads. I'll probably get heads again on the next flip." Don't do that. :)
Do you really believe there's less than a 50% chance strong AI won't be invented in your lifetime?
I don't know. I don't really see the value in speculating.
There is all kinds of value in speculating. We do it all the time in things like war games. 'If neighbor $x attacked us, what would happen?, would they win?, what can we do to prevent this?'.
Of course this may in your mind hold very little value now, but I promise if and when it occurs you will change your mind quickly on that topic.
Which is corrrect, if you don't know the true probability of flipping heads. You might find, for example, that it's a trick coin with two heads and no tails.
You absolutely can predict future progress from past progress. E.g. Moore's law held true for decades after the observation was made. If you see a technology advancing rapidly, then there is no reason to say it will stop in the near future!
>I don't really see the value in speculating.
Because everything depends on this prediction. The invention of AI will be the most significant event in the history of humanity. It will totally change the world. Or likely, destroy it. Being prepared for it is absolutely necessary.
By your logic, you can predict the failure of AGI predictions by past failures of (every) AGI prediction.
Why would you have a strong prior belief about the invention of AGI? Now you are claiming to have far more certainty than I am.
>By your logic, you can predict the failure of AGI predictions by past failures of (every) AGI prediction.
This logic is extremely flawed. First not every prediction was wrong. Many people predicted it would happen in 2045, a few in 2030.
Second there's no reason past predictions represent the accuracy of future predictions about the same thing. Predictions should get more accurate over time, and early predictions are expected to be wildly wrong.
And third there's anthropic bias. If they were right, then we wouldn't be here to speculate about it. We can only ever observe negative outcomes, therefore observing a negative outcome shouldn't update your priors at all.
The vast majority of predictions were wrong.
Yes, the logic is flawed, that's why I said it was your logic.
I for one did not say to not worry, and I bet $1500 vs. $2250 on Sedol winning the match before getting cold feet and arbing my outstanding bets down to $400 vs. $700.
What actually happened was the reverse of that; AI moved faster than I publicly bet a large sum of money on it moving, and I was already worried before then.
I'm not aware of the reverse-reverse having happened.
If you're claiming something as a successful advance prediction to bolster belief in a general model, it's fair play to ask for a record of that advance prediction.
[EDITED to add:] Oh, I see, your point is that it looks like AlphaGo is doing better than an "AI moves fast" advocate expected, rather than (per Teodolfo) no better than some "AI moves slowly" advocates expected. Fair enough.
Beyond that, it seems reasonable to call for discussion about how AI is used [as he also does in that article].
Neither does the Friendly AI crowd.
Infinity is a real problem. When you try to learn from examples, you first need to see "enough" examples of whatever you're trying to learn. If there are infinitely many such examples, no matter how clever you are in tackling your search space there will always be infinitely many examples of infinitely many situations you've never come across, and that you won't be able to learn.
The typical example of this is language. You could give a learner all phrases of a given language every produced and it would still be missing an infinite amount of necessary examples. Somehow (and it's freaky when you stop to think about it) humans get around this and we can produce and understand parts of infinity, without sweating it.
Machine learning is simply incapable of generalising like that and anyone who thinks AGI is just around the corner has just failed to consider what "general" really, really means.
Though to be fair, now that I had my little rant I have to admit that you don't need to go "general intelligence" before you can be really, really dangerous. Even if AI doesn't "take over" it can do a lot of damage, frex if we start using autonomous weapon systems or hand over critical infrastructure maintenance to limited and inflexible mechanical intelligence.
Look- take Monte Carlo methods. You can sample a very big number of events and hope to get some useful information from that. If you sample infinity, though, what do you get? Infinity.