AlphaGo shows its true strength in 3rd victory against Lee Sedol
gogameguru.com
gogameguru.com
At this point, the match is won, but games 4 and 5 will commence. The question shifts from whether AlphaGo is better than humanity's best, to whether humanity can even have a chance of beating AlphaGo in a single game. And so far, it sounds like the answer is likely to be a resounding no.
Second game, AlphaGo looked obviously strong throughout.
So the probabilities are maybe not entirely wrong.
For reference I am 4k rating, somewhat below darkforest(1d amateur) and so very very far from 9dan pros that AlphaGo is beating. Usually, pro matches are hard to understand but the excellent commentary and abundant community analysis helps greatly.
AlphaGo doesn't tire and thinks globally on every move.
We saw that time and time again. Lee playing locally and AlphaGo surprising everyone with an unexpected move somewhere else on the board.
Obviously it is better to be globally optimal than locally optimal.
(edit: also there'd be no point playing a game against it!)
I think the OP simply implies AlphaGo doesn't seem bound by these human cognitive crutches.
This is not the same as "global" in the theoretical sense, as in "finding the perfect solution / exhaustive proof". See also "game temperature" [1]
[1] https://googleblog.blogspot.com.br/2016/03/what-we-learned-i...
The rollouts they are doing, evaluating every probable move, it is a search process of trying to find local optimality. You have a current state of the board and you're trying to find a decision that minimizes your future regret. It is by definition local optimality.
Global optimality would be finding a sequence of moves that wins you a game.
Is it known that Go is winnable? It may be that only a draw could be guaranteed. A globally optimal player would be provably able to do (whichever of these it turns out to be) for any legal position. This is harder than winning a particular game.
"the stronger player will typically play with the white stones and players often agree on a simple 0.5 point komi to break a tie ("jigo") in favor of white."
Which suggests ties can occur in practice with human players, but are then broken wrt an external parameter.
In terms of optimality, is it known in Go whether an optimal player can force a win against (i) all suboptimal players; and (ii) another optimal player, based on whether they play first or second?
I realize we don't know the optimal policy but sometimes we can decide (non)-existence anyway. Anyone happen to know for Go?
I don't know a lot about the strategy of go, but it seems to me that any play that doesn't take into account the entire state of the board is allowing for the same class of sub-optimal behavior as the example above.
https://en.wikipedia.org/wiki/Bellman_equation#Bellman.27s_P...
and your attempted counterexample makes no sense (as the 'two cinderblocks' solution is not a subproblem of the optimal solution).
In general, it's false that sub-problems of a problem with an optimal solution are optimal themselves.(Wikipedia's article says the same differently: "In computer science, a problem that can be broken apart like this is said to have optimal substructure.", implying some problems can't be broken apart like that). Also see: https://en.wikipedia.org/wiki/Optimal_substructure
Maybe Go has this property of optimal substructure. Maybe it doesn't. I don't know, but it sure isn't immediately evident.
If the moves on different parts of a Go board didn't influence each other and could be solved separately via dynamic programming, it would be a much easier game (and probably uninteresting).
Optimal solution: Kill yourself.
Sub-optimal solution: Both live.
the implication of having machines dictate our behavior [WRT environment, each other] to optimize the survival of humans far into the future seems a bit closer to reality.
it's a board game of military strategy. while it is severly limited by its rules of play relative to real life, it is not inconievable that at some point it can dictate life and death as a capable military advisor.
i think it is inevitable that we will utilize AI to make better predictions about our own future than weather forecasters. not in our lifetimes, of course.
and 20 years ago is still well "within our lifetimes"
And the 300-fold increase in branching factor.
You could setup a game of Nim: https://en.wikipedia.org/wiki/Nim with much higher branching factor vs Go, but because you can trivially separate a winning moves vs. losing moves those branches become irrelevant in Nim. AKA, if there are 2,000 winning moves then they are all equivalent.
i have a hard time believing something like that isn't already in use to some extent, at least for special operations forces. how long until they just hook that into the drone fleet? what if they already have? autopilot on a modern fighter/bomber is basically the same thing, except there's a human in the loop, right?
that would also go along nicely with the whole matrix / skynet plot that seems playing out in real life right now...
however, point 2 is only understood from a very high level by humans. we know the dependency chain goes deep, but we don't know exactly which branches can be ignored without effect and which are critical.
our default is to protect all the things "just in case"
It took Lee Sedol many decades of his life to train to achieve this level. And his ability to pass on his skills is limited. Now that a computer has achieved this level, the state of the neural network behind it can be serialized and ran into an unlimited number of computers and have millions of systems that are more proficient at Go than the best player in the world.
People have said that achieving the cognitive level of the human brain requires to match its computational power. But if you take out all the parasympathetic and motor boilerplate, what is actually left for mental tasks is much less from that, most of that power is not even recruited for higher level mental tasks. That lowers the bar for strong AI.
Then, strong AI can be immortal, and never physically deteriorate from aging. Strong AI can multiply infinitely and communicate at a rate that would be equivalent to writing millions of books in a second. It could transfer all its knowledge in seconds. It can also recursively improve itself. This advantage will lower the bar for strong AI even more.
Whereas, good human players learned as they went with far fewer matches, trial and error, planning, and learning from pto games sometimes. We can even go from Age of Empires to Starcraft with little prrparation and still do OK.
That worked with backgammon without any problems. (http://www.bkgm.com/articles/tesauro/tdl.html)
In Go, a naive approach would probably also work, but would take ages. There's probably a slightly smarter approach possible.
When the AI is cognizant of the fact that Go is a game, and knows what a game is, and perceives that the game is taking place in a larger reality, where there are other things happening while the game is going on....then I'll be impressed.
Or maybe not. The history of AI is one of humans always moving the goalposts when AI advances.
Chess? just a computationally simple game. Driving? Well it's just physics. Go? now that requires intelligence oh wait, just some deep neural savant thing, not real AI
You can make some analogies between sparse matrix data representations and what happens in the mammalian neocortex. It is of course not self conscious, but it effectively deals with semantic information and relationships among it.
The sheer amount of computing power that AlphaGo needs will be a problem for games as complicated as Go for people who aren't Google. There will probably be more activity in simpler games with interesting properties.
Non-game applications will require a fully formalized set of rules. The fact that formalizing what you want is hard is the reason we need programmers in the first place, so nothing fundamentally different there.
I did some work for a Forex trader. He had devised an algorithm for trading and used it daily for modest gains. He engaged me to automate it for him via the vWorker freelance website [1]. I got the engine going getting the realtime price data etc. But when it came to his system it didn't work. In practice he didn't always act upon the signals he thought he did. We ended up in payment dispute because he insisted that if I had done the work right then the bot would be making money. By this time, I had learned the error of his thinking. We arbitrated on 50% payment - I wasn't too pleased with that but I had learned about trading.
I do feel that given all the hype and speculation about strong AI researchers in Deep Learning should really come forward to contextualize the real impact of this - which I think is not all that great.
You can expect a lot of advancement in actual games very soon, including ones with randomness, and after a few years probably hidden information games too. Being able to specify real-world problems as formally as a board game is now a more effective skill than it was before (it was good already).
It is also good enough to put it in a combat drone that visually recognizes objectives and people. This can be all commanded by a small group of people, anonymously, and detached from accountability and without casualties.
To say nothing of the casualties that'll be inflected on the AI warriors...
Did any credible skeptics actually claim this would never happen? If so, I'd love to hear their reasoning.
Autonomous vehicles are similar. IMO, it would be a fool's bet to bet against them ever happening. But it's reasonable to debate whether they'll be mainstream Level 4 consumer devices in 10 years, 25 years, or 50 years.
Now, being more specific about Go:
- Lee Sedol himself was confident he would win.
- Ke Jie said he would win against AlphaGo. I think he would lose too.
https://en.m.wikipedia.org/wiki/List_of_countries_by_carbon_...
That's already happening, at least somewhat, right?
[Link](https://medium.com/@peterbsmith/computer-beats-human-at-hard...)
My grandfather was very racist and never changed as long as he lived. I'm glad most people aren't immortal. There's no guarantee a full AI wouldn't be tempted by evil and just become republican (or worse, a VC).
It can also recursively improve itself.
So can people (the more you know, the more you can learn), but most don't. It's important to remember "intelligence" isn't an abstract concept—intelligence is also embodied in personality—and personalities have wishes and goals and desires and loves and hates and that one song they can't get out of their head. A true "strong AI" will be fully conscious, not just algorithmic function bating.
Good luck telling a mildly strong godform to stop tripping on youtube videos and instead solve the global economic stability equation over lunch.
This advantage will lower the bar for strong AI even more.
That's kinda foofy conjecture. Being good at rectangular grid outcomes isn't necessarily a step in any direction towards a hands-off tax evaluating robot.
It feels really really good to talk about how AI will be a hundred billion trillion times smarter than the combined brainpower of all humans that have ever lived, but it feels good in the same way thinking dead people live again after they die feels good—it triggers that warm wishful thinking parietal lobe that removes a bit of reason for the sake of an overarching calmness.
Enthusiasm is great, but tempering with real expectations and less technopriesthood is better.
Now, while many people die and their ideas die with them, there are built in aspects of our brain that are hardwired. People eat animals because they are tasty, people follow the life of Kim Kardashian and waste money buying a rolex and wear fur just because the stupid way our social functions are hardwired in our brain.
Fresh minds don't have the burden of outdated ideas... quoting Max Planck: "a new scientific truth does not triumph by convincing its opponents and making them see the light, but rather because its opponents eventually die, and a new generation grows up that is familiar with it."
We are primitive forms of life after all and are at the point where we should be able to say, with certainty, that machine intelligence will be objectively superior and something that we should feel OBLIGATED to give rise to, regardless of consequences.
There is no prescribed script about what our evolutionary value should be. We will be whatever we ourselves make of us. Unless something unexpected happens -more advanced aliens intervening with our specie for example- Our human instinct is to grow, explore, discover, compete, share as much as we can. AI should be seen as a tool to our goals. Not as an end by itself. There is of course the possibility that we may loose control to the AI. That is not our desired goal, but it could happen. In that case our future will not be our decision. We will be at the mercy of whatever the AI choose for us. That will depend on what kind of AI we built, and how the rampant AI chooses to modify itself. But we will strive to reach our human goals. Because that is the human nature.
"We will be whatever we ourselves make of us" sounds meaningless to me. The universe obeys certain physical laws, evolution is directed and follows a specific pattern. We have up to this point been bound by both. Occam's razor says this will continue to be the case.
I think we are already at the crossroads where we can more than speculate about the function of machine intelligence in evolutionary terms, but hopefully sooner rather than later we will have actual _evidence_ that we can look at and use to refine our assumptions.
Don't human professional players do exactly this? They care about playing the best move and winning. The difference is that AlphaGo is likely much more accurate at determining what "95% probability of winning" means in terms of gameplay. A human has a harder time judging the eventual outcome of a game, and so plays more "aggressively" (favouring point margins to account for variance in the estimate, etc.) than AlphaGo might.
(Of course, making an AI that understands emotion isn't easy. And happiness by itself is an insufficient goal if we're building ourselves a benevolent overlord.)
I think the writing on the wall is clear, to anyone who cares to take a look. The moment we create an artificial intelligence capable of self-improvement and let it loose, we will have fulfilled our function (others say destiny) and will therefore be obsolete in every sense of the word.
What will happen to humanity after that point is irrelevant.
You don't see humans trying to keep bacteria "in the loop", why expect otherwise from our artificial progeny?
Bacteria didn't design us. We will design our artificial progeny and so try to keep them friendly. Whether that will work, time will tell.
Correctly programming something that takes into account all of the nuances of the human condition might be impossible, and the results of mistakes are uniformly terrifying, but the result is entirely in the hands of humans (and human mistakes).
Formally specifying what we actually wanted is why programmers have a job. And the difficulty of that is why we have bugs. So that's still scary. But it's not the machine's choice in any way
Anyway, yes, it will optimize whatever function it's told to optimize. So you'd better hope that whoever gives it that function 1) cares about people's wellbeing and 2) is good at expressing that in a way the computer can interpret.
1 is pretty obvious, 2 may be a bit less so. For example, a superintelligent computer commanded to maximize global average happiness might decide the best route would be to work towards killing the entire human population except for one individual, who is kept constantly drugged out of their mind.
Go AIs weren't expected to reach this level for at least another 10 years.
Before AlphaGo, Zen and Crazy Stone (the previous Go AIs) could only play against top-level professionals with a significant 4-5 stone starting handicap, and this was less than 3 years ago. A 4-5 stone handicap is basically taking control of half the board before the game has even started.
It really shows how the neural network approach made a huge difference in such a short time.
You have a trillion connection neural net wrapped in 2 pounds of flesh inside your head. This is a massively larger amount of hardware compared to just about every animal out there. Throwing hardware at a problem is a solution to intelligence.
I don't like the biological comparison, but using your metaphor it would be like God saying "Hey I've created a brain but only have 10 billion synapses. Evolution would normally take 10 years to get to human-scale at our current organic rate but if I throw money at building a bigger brain cavity I can squeeze in the 1 trillion to get there today!"
At the very least, it's interesting to see how much more accessible computation has become. Back when I was in school I could only dream of having a cluster of 280 GPUs. When sometimes the dream would come true and you had access to a cluster you would have to wait your turn in the job queue and hope you had enough compute in our resource quota to prevent your job from being terminated.
Now I could spin up a 280 GPU cluster on AWS (after dealing with pesky utilization limits) for only $182/hour. If researchers at Google have been doing this non stop for the past year they have "racked up" $1.6M on just compute. This is a drop in the bucket for a marketing department and the publicity they have achieved. I don't think normal Go AI developers have access to those resources :)
Modern chess engines are built on a testing infrastructure that makes it possible to measure how each potential change affects the playing strength. This "Testing revolution" has brought massive improvements in playing strength.
For AlphaGo, it's probably the training that requires the most computational resources. The 'distilled knowledge' could perhaps run on a desktop PC. The program would search fewer variations and would be weaker, but if AlphaGo improves further, that version might still be stronger than any human.
Also, it feels like training the AI against many games with lots of hardware is somewhat equivalent to a human progressional who engrossed themselves in the game and trained since childhood.
He said that the training data set size is much, much larger than the number of lee sedol games. It is like a drop in the ocean and not enough to significantly influence the resulting policy network.
I don't know the details of the algorithm and perhaps it doesn't give more explicit weight to his games. But I wouldn't be surprised if after some iterations the algorithm decided to give more implicit weight to the games of the world leaders.
However, this only gets you so far, and the training set is fairly small compared to what you want to really train both the policy and value networks well. So then they had it play millions of games against different versions of itself, training both a new policy network and the value network based on that.
It's unlikely that Lee Sedol's games made much if any impact on AlphaGo's training. It was bootstrapped off of high-level amateur, and some casual pro, play, but from then on it just trained against itself.
So even if AlphaGo isn't explicitly designed to play against his style, it's implicitly trained in it.
But it's pretty subtle and I'm guessing that the volume of high-level tournaments overwhelms any effect, as the AlphaGo team said.
What does seem right is that whatever strategic insights are in Lee's play are reflected in current games--his and the younger generation who came up in his shadow. Whatever strategic novelties shape AlphaGo, they are totally new to Lee.
I don't think that would make the difference: 1) there's no trick to be learned and 2) the same thing happens with human players to some extent--when Lee Changho appeared on the international go scene, his style was misunderstood and underestimated, even when his games were public.
However, it is true that there is a real asymmetry--AlphaGo may not know Lee from other players, but it has had the opportunity to "study" the best games of the current players, and no one outside of DeepMind has had an opportunity to study its games.
That's why the next goal AI developers are pushing towards is incomplete information games.
While unsupervised learning might be the holy grail, this victory is really about deep learning, which is an advanced supervised learning technique.
In your case you have knowledge but lack skills.
When Federer was losing to Nadal he changed his practice and game to counter Nadal's style. He took knowledge and applied it to his skills.
When Federer came out with SABR people were like wtf. Now they know of it and put more umph on their second serve. See. Knowledge + skills
The way the would play against a lefty is different than a righty. Someone with lots of topspin vs someone who hits flat, a pusher vs power hitter, etc.
If someone is clearly better than you, you can study it's style all you want, it won't make a difference. You probably wouldn't even understand it's style.
There's a reason professional sports team and players study the styles of their opponent. Teams employ statisticians for this exact role. Before games, teams study the style of their next opponent. After every after game teams go and study the recordings of their last game.
But I think a computer as a training buddy is indeed a good idea to improve Lee's skill. I don't know how Lee feel right now, he's defeated, there must be an enormous pressure inside of him, but I think he will appreciate the challenge because he needs more challenge! He has played against top players all over the world now for his 20+ years since earning 1 dan rank. Also, I think the computer play against itself and many random moves. Unlike human players, one would expect computers to play unconventionally since the computer can predict so far ahead about the probability of maximizing winning.
But I would say he might feel like crap having lost his title to a 19 year old, and now losing to a machine all in the same week.
It's not whether AI can outperform human, it's a matter of when.
However, this article contains much more analysis of the game than previously posted articles, and is worth reading.
I reviewed the match, and it seems that very quickly White (AlphaGo) went aggressive, and Black was trying to contain it. I suppose a human player can commit serious mistakes by playing too aggressively from the start, right? But the summary of the game speculates move 31 might have been the losing move (maybe implying the game until that moment wasn't going so bad for Lee Sedol?), while to my untrained eyes it looks as if White was constantly on the offensive from the start, and Black was playing almost exclusively to contain it.
Or am I reading this wrong?
White certainly played plenty of moves that required black to respond, but black also did the same. For example, the sequence beginning at 77 is black getting a foothold inside an area that white might've hoped to claim later. Then at move 115 black attempts an invasion into a very strong white area, ultimately failing to escape or live, leading to resignation.
That all said, I wouldn't say that either player was particularly aggressive. I've watched plenty of human games where one player or the other very intentionally picks an all-out fight. If you're really good at reading ahead locally and you think your opponent is not, this make sense. This is especially true if your large-scale game play is not as good as your tactical play.
This is also common in high handicap games. If white can't slowly eke out an advantage early on, the only remaining tactic might be to go for a huge fight and see what happens.
In 13 years, computers progressed from beating Chess to mastering Go. I hope that in 13 more they will learn to not steal focus while I type
But the problem is that computers do what you say, not what you mean. If I write a function called be_nice_to_people(), the fact that I gave my function that name does nothing to affect the implementation. Instead my computers behavior will depend on the specific details of the implementation. And since being nice to people is a behavior that's extremely hard to precisely specify, by default creating an AI that's smart enough to replace humans is likely to result in a bad outcome.
Recommended book: http://www.amazon.com/Superintelligence-Dangers-Strategies-N...
Chess may have a big search space, but it is "well behaved". Moves might be out of the ordinary, but not too much (weird moves that still help to win the match are rare)
AlphaGo might also have learned from past matches, but that doesn't give all the answers
The black-box aspect, the fact that you can't understand what it is thinking is curious, to say the least
How true is this? Presumably you could ask AlphaGo to give you a list of the top 10 best moves it considered and perhaps, given a certain move, what it sees the board looking like 10 moves ahead.
If you're watching the commentary, one of the commentators (Michael Redmond) will try to explain the outcome of certain moves and this would seem to be equivalent although I imagine he could explain this in terms of the endgame better.
I would imagine that given the "All-22" camera footage from a football game, it would be possible to construct a piece of software to monitor what was going on (visually identify players by name/number, determine who has the ball, etc.) and then convert that model to a stream of English text.
Making it nuanced and entertaining, on the other hand, would be a challenge. :)
The problem is small sample size. You only have the plays as called in the games to rely on. You can't "replay" the game with a different set of plays and see what the result would be.
This was a problem even in Moneyball. It's just that On-Base Percentage has a really good correlation to runs and therefore wins.
It's really all about execution on the field and the "feel" of the game.
DeepMind doesn't generally comment on these sorts of things, but I think they knew it would be strong, just not how strong.
It is hard to accurately gauge the playing strength of such a program for the following reasons:
* Go is not solved, it doesn't really even admit a good heuristic for close games. In chess, given enough time, we can do a pretty good job of searching the possible outcomes from a given position, and even if we can't evaluate all the possible endgames, evaluation functions exist that give us an idea of which side is ahead when we terminate the search. Go has a much higher branching factor (more moves available at each turn, making the search more expensive) and short of a catastrophic blunder, it's hard to quantify who is ahead at each point in time[1]. So we can't (in general) quantify optimal play, and therefore cannot quantify how much AlphaGo (or anyone else) deviates from it.
* One known aspect of programs using Monte Carlo Tree Search is that they play to win, and are willing to sacrifice margin of victory to maintain or increase their odds of winning. According to some people I've talked to, this can be suboptimal, but there are methods of addressing this[2]. Note that you can't just change the objective function from "winning" to "win by as much as possible" without potentially reducing AlphaGo's strength.
* The value function learned by a deep net is hard to interpret, partly because it encodes information about the possible futures arising from a position, and partly because it involves a tremendous number of calculations to compute. We do not know what it is representing at the intermediate levels-- techniques that aim to visualize or cluster unit activations can provide a bit of insight, but there's always the possibility that we're interpreting the patterns incorrectly because we're trying to fit it into the framework of "what would a human think?". Further, the representation is somewhat monolithic-- we can't tweak the value of one thing without changing the values of others. In chess engines, we might modify the material value of, say, a knight, without affecting the value of a rook. In a convolutional net, if we adjust the value of one position, it will tend to affect the value of many others.
* We can attempt to quantify its strength by comparing it to other programs, but AlphaGo has already crushed other programs (99% win rate), so all that it tells us is that it is stronger than those programs.
Essentially, without the ability to perform a significant amount of searching, or gauge strength via margin-of-victory, or to examine the program from other angles, it's hard to gauge just how good the program is.
The only thing we can do is throw skilful opponents at it, and see if it fails eventually[3]. In its current incarnation, though, it seems like its playing strength is just going to be "stronger than you".
---
1. Hence the need for Deep Reinforcement Learning-- we learn how valuable each position is based on the results of the positions that it can lead to.
2. By modifying the search to maintain a given margin once it has been attained-- but I do not work at DeepMind, nor am I an expert on MCTS like some of my colleagues, so I don't know if this could adversely affect AlphaGo's overall strength.
3. We might be able to get a better idea by having extremely strong players play against weaker versions of AlphaGo (ones with less computing power) and then sort of telescope upwards once a good baseline has been set, but it remains to be seen if the current level of invincibility is due to it not having been around long enough for its weaknesses to become apparent[4].
4. How would an AI researcher play against it? I've talked to a few people, and the answers have been: (a) play the game to the conclusion, don't resign; (b) attempt to take the game off-the-rails into parts of the state-space it hasn't really explored before (although this is risky because the human player is more likely to make mistakes in these cases as well), and (c) let it get ahead somewhat so that it "relaxes" (see the points about margin), allowing you to catch up and overtake it.
You're right that it would be interesting to see, though-- we need to get better at understanding these sorts of systems, at least until they can start optimizing themselves.
Alternatively, we might train a different agent (OmegaGo?) to try to win by the largest margin possible-- if it works as well as AlphaGo, then that might give us some more insight into how strong both programs are.
One interesting implication of this is that it very much applies to perfect information games--not games that involve an element of luck or, indeed, pretty much anything that involves the real world. In those latter environments, you absolutely want to run the score up or otherwise take small chances to gain an immediate advantage to create a cushion against future unexpected events (even if they slightly hurt your odds if those events don't occur).
I suspect this is less suboptimal when a computer is involved since a computer never gets tired or makes an attnetion-based mistake.
The community has also collected a lot of good Go resources at Sensei's Library: http://senseis.xmp.net/
Good luck!
i guess it's time to play with dbn's again. At least I have accumulated a few more Titan x cards in the mean time.
I find it very hard to find information about this online.
So far, they have tested a previous iteration of the program against a much lower level professional, as well as all of the other Go engines they could test it against. It easily defeated all other Go AIs, but no previous Go AI has been able to compete at a professional level. It also defeated Fan Hui, the lower level professional, though it made several mistakes and lost a couple of the informal games (with lower time limits) it played against.
Since then, they've had several months to tweak it, and then let it train against itself some more. The thing is, to really determine its strength, you need to have it play serious games against professional players. So there really was no good way for them to determine its strength without playing these games. It's possible that it only would have improved a little since it's games with Fan Hui; in that case, the match with Lee Sedol may have been easy for him. However, it appears that it has improved substantially.
On the other hand: Knowing what makes some problems much easier for humans than for computers can have direct value on Google's business. Go was a hot candidate for a game with this property.
How much publicity has Google received for $1 million? (Much less than that actually, since they won)
Once this have been achieved, it is downhill from there. The singularity. In my opinion, we must start to consider how this tremendous breakthrough will affect the human race. Our lives, economy, culture, etc. Awesome things we will be witnessing in the coming times. It would be great if it were possible to make public, the progress made researching specifically AGI. I wonder how big of an effort is currently being done. This would be possibly the biggest discovery ever for humanity. It deserves a fantastic effort.
Considering what a large impact this will have, we must take that possibility seriously. Not to do so would be foolish.
AIs can play chess, recognize speech consistently, drive cars, win at Jeopardy and now Go. Deep Mind has some general capabilities. We must anticipate programs with even more general intelligence.
Good luck with that!