AlphaGo beats Lee Sedol 3-0 [video]
youtube.com
youtube.com
As a Go player, I'm really excited to see what kind of play will come from that!
[1] http://www.theverge.com/2016/3/10/11192774/demis-hassabis-in...
The niggling thought in my mind was that AlphaGo's strength is built on human strength.
Neural networks are modeled after biological systems to begin with, I don't the that's a meaningful concept at all.
BTW: even humans don't just randomly pick up the game. They have teachers, who teach them the tricks of the trade and monitor their games.
You train many models. Then you "distill" their predictions into one model by using the multiple predictions (from many models) as targets (for the single model trained afterwards).
You're right to point out that humans don't do that.
I think it would be "cheating" if you train BetaGo on AlphaGo, for the purposes for doing that experiment. The goal would be to have some kind of "clean room" where people fumble around.
Of course, you can also run the other experiment to see how fast you can bootstrap BetaGo from AlphaGo. That's also interesting.
Isn't that the nature of human endeavor? Always looking for the next challenge?
Uncarved block
In the days when Sussman was a novice, Minsky once came to him as he sat hacking at the PDP-6. "What are you doing?", asked Minsky. "I am training a randomly wired neural net to play Tic-tac-toe", Sussman replied. "Why is the net wired randomly?", asked Minsky. "I do not want it to have any preconceptions of how to play", Sussman said. Minsky then shut his eyes. "Why do you close your eyes?" Sussman asked his teacher. "So that the room will be empty." At that moment, Sussman was enlightened.
(It seems based on a true story https://en.wikipedia.org/wiki/Hacker_koan )
https://en.wikipedia.org/wiki/Hacker_koan
Eric Raymond kinda butchered the Jargon File when he took over maintenance, so it wouldn't surprise me if some of the text there is invented. The original Jargon File does not contain any koans:
Making an AI that plays Go well is not (and has never been) the real goal. They're trying to learn how to build a AI that can solve any problem.
It can be a case that training and learning on just a learned policy is going to get you stuck in a local optimum that is of worse quality than the one with pretraining.
If they stored all of the AI played games their reference policy (the data) would be of extreme value. You could train a recurrent neural network, without any reinforcement learning, that you could probably run on a smartphone and beat all of the players. You wouldn't need a monte carlo search too.
There are algorithms [1] that have mathematical guarantees of achieving local optimality from reference policies that might not be optimal, and can even work better than the reference policy (experimentally) - assuming that the reference policy isn't optimal. The RNN trained with LOLS would make jointly local decisions over the whole game and each decision would guarantee that a minimization of future regret is being done. Local optimality mentioned here isn't finding a locally optimal model that approximates the strong reference policy, it means that it will find the locally optimal decisions (which piece to put where) without the need for search.
The problem is that for these algorithms you have to have a closely good reference policy, and given a small amount of human played Go games, reinforcement learning was the main algorithm instead, it allowed them to construct a huge number of meaningful games, from which their system learned, which allowed them to construct a huge number of more meaningful games, etc.
But, now when they have games that have a pretty good (AlphaGo is definitely playing on a superhuman level) reference policy, they can train the model based on that reference policy and they wouldn't need a search part of the algorithm at all.
The model would try to approximate the reference policy and would definitely be worse than AlphaGo real-search based policy, but it wouldn't be significantly worse (mathematical guarantee). The model is trained starting from a good player, and it tries to approximate the good player, on the other hand, reinforcement learning starts from an idiot player, and tries to become a good player, reinforcement learning is thus much much harder.
I definitely did not expect that.
Major credit to Lee Sedol for toughing that out and playing as long as he did. It was dramatic to watch as he played a bunch of his moves with only 1 or 2 seconds left on the clock.
That's on purpose to make full use of time and think about the next move
As a 3d amateur, I'm really curious about when he resigned. It really seemed like he was playing the position out to go for the win (or perhaps to see how AlphaGo would fare in ko). It didn't look like he was searching for a place to resign.
Edit: M5 was definitely played as a ladder breaker, so the above is correct.
A possible explanation is:
During self reinforcement learning, AlphaGo learned to minimize Ko potential by maximizing its probability of winning through diminished available Ko moves.
It would be interesting to see how AlphaGo would be able to capitalize on a game that emphasized Ko play, but that would take more time with AlphaGo to emphasize that kind of play.
edit: I'm not sure why, but I think Lee Sedol is partly holding back, or not playing at his maximum ability. It feels like these games are more along the lines of query games.
I look forward to the next two games because I'm 100% certain Lee Sedol is going to query the AI with some new queries.
So I'd be very surprised if that turns out to be the trick. Things that are hard for human players are not at all necessarily AlphaGo's weaknesses.
I think us humans made a critical error in that line of thinking.
It didn't avoid ko because of the risk of loss.
It avoided ko because of the lack of strategic win.
It's still a game that can be described in terms of clear state-machine rules. The real challenge for AI is making sense and acting in the real world, which can't be described in such way. I consider advances in self-driving cars much more interesting in that sense - even if, even there, there are at least some rule-based constraints that can be applied to simplify the representation of the "world state".
Anyway Marazan already pointed that out, but any computer system is a state machine, with 2^N states where N is the number of bits the machine can flip anywhere in its system (RAM, registries, disk, etc.).
Beating the average human Go player was probably accomplished decades ago, whereas it's not even clear if we're safer than the average human driver (under all conditions).
These tasks are just wildly different, and yes I think it's basically all due to the fact that Go's state is so easily represented by a computer, and the goal is so concrete.
The real problem is dealing with all the edge cases. Think of this edge case. You pull up to a red light, a guy with a gun starts running at your car in a manner you perceive to be threatening.
You as a human are most likely going to step on the gas and get the hell out of there saving yourself, at some risk of causing a traffic accident.
The car will just sit there till the light turns green while the windows get shot out and you get dragged out of the car.
Thing is, it wasn't. Go AIs were on the level of amateurs (and amateurs could win) only a two years ago.
edit. 'Decades ago', i.e. in 1990s, amateurs would crush the AIs. https://en.wikipedia.org/wiki/Computer_Go#Performance
(or something like that)
Politicians fool countries delivering empty promises about better health, education and security. A supraintelligent AI could promise making humans rich, healthy and powerful, to then break its promise and dominate the world.
Basically, why would an AI want to dominate the world? Humans would have to both very stupidly give the AI values that encourage it to dominate the world and very luckily (or unluckily) give it values that actually converge to a horrible outcome against human intentions by random chance (since the AI designers certainly won't be tuning the value set for that outcome).
Humans are going to program their AI's to try to make as much money as possible. Many corporations are already mindless and reckless amoral machines that relentlessly try to optimize profits despite any externalities. Try to imagine Exxon, Wal-Mart, and Amazon run by an intelligence beyond human understanding or accountability.
Ha, if that were true we wouldn't be constantly extending the law and putting people in jail because they keep breaking the law for profit.
>ivilisation can't work because humans will want to make as much money as possible
And yet we keep running into issues with long term pollution and environmental degradation because of the growth of civilization.
>whether they feel they already have sufficient money for their own needs,
Does greed have bounds?
The problem is when you have multiple AIs. Then same evolutionary principles apply. Paranoid and self-sustaining AIs survive, and the circle goes on...
World domination is a similar situation. For any goal you give an AGI, one of the big risks that may prevent that goal from being accomplished will be the risk that humans intervene. Humans are a big source of uncertainty that will need to be managed and/or eliminated.
Of course, a cat is not conscious. But compared to an AI, we might also be considered pretty low consciousness beings, or at least beings in front of which you don't justify yourself.
How do you know?
If you take the view that consciousness somehow arises from the brain and neural connections (which is intuitively plausible, but I personally am skeptical), it stands to reason that other species with complex brains are conscious as well. Perhaps "less conscious" (if that means anything) in proportion to how much less complex their brains are.
Now experts can discuss details or semantics, but do you truly suggest cats might be conscious?
In particular, we all know that we're conscious, but can't really explain what that means.
Arguing they are not conscious in the sense of a more obscure definition is a bit pointless unless you specify your definition.
Why do you believe this? I don't like cats, but I wouldn't argue that they're not conscious.
Instead of debating the suitcase word "conscious", let me ask: 1) Do you believe that toddlers are conscious? 2) Is there a more precise way to state your belief that doesn't use the word "conscious"?
If some unforeseen event occurred and I had to abandon my cat, thereby breaking my promise that I would take care of her, I would definitely feel guilty about it.
Of course, a cat is not conscious.
Either this is a nonstandard definition of "conscious", or you haven't met many cats.
Or it could devote its entire power to making human lives the best and most comfortable they can be because humanity is some super-precious resource in the universe and it feels it's unimportant because it's just a bunch of silicon and electrons.
Supraintelligent AI being evil is FUD imho because we can't reason about supraintelligent AI.
There is a difference between being evil and incomprehensible intelligence. You are not being evil when you accidentally step on an ant or dig up an ant-hill to build your shed. The ants won't be able to understand what you're doing, or why.
Maybe there's a second AI on the same level as the first and thinks the first AI is evil. We're still dumb as rocks compared to them, but something certainly has that opinion.
That's why e.g. the Future of Life Institute's open AI letter is so important: http://futureoflife.org/ai-open-letter/ We need to be thinking in advance about how to solve the "value loading" problem for future AIs, and how to architect them so they can be deployed to solve big problems without being undermined by subtle but catastrophic bugs.
It feels like computers have taken one aspect of humanness: logic. Computers could do arithmetic, do algebra, play chess, and now they can play go.
It hurts because logic is usually thought to be one of the highest of human characteristics. Yes computers might never be able to replicate emotion, but even dogs have that.
There's still some aspects we have left to call our own. Computers perform poorly at language-based tasks. They can't write books, write math papers, compose symphonies. I hope it stays that way.
Edit: iopq, pretending to be a dumb human (or one with a language barrier) is cheating for a headline. A real Turing test would require a computer imitate a human for longer than 5 minutes (although currently that is plenty of time) and without any caveats or limitations on the computer's skill.
I'm sure there's always someone that can write books or maths papers or symphonies better than you. I don't think this robs you of purpose, unless your purpose is to be the absolute best at something.
Anyway, I find it curious that you would say logic is a quintessentially human trait, because humans are naturally quite bad at logic.
So a more apt analogy would be if there was someone inside every cellphone who could write books, papers, or symphonies better than you. That day is coming.
People have been doing his for decades, but as far as I'm aware, no-one has tried it with thousands of distributed servers and millions of songs.
AlphaGo needed a training set of perhaps a billion games to be as good as it is. The dataset of master Go games is perhaps a million games. So AlphaGo played at tons of games against a half-trained version of itself to reach the billion game mark.
This doesn't work for songs, because there's no one to tell AlphaBach whether any of the billion symphonies it makes are any good. AlphaGo can just look at the rules and see if its move lead to a win, but there's no automatic evaluation function for music.
Perhaps the Matrix wasn't using the humans for power, but rather the computers wanted to get good at writing music, so they gave each human in it slightly different music and watched their emotional responses.
This is possibly my favorite comment of the whole thread.
It's a super interesting idea and could make for some fascinating science fiction. Poorly programmed AI might not wipe out humanity, because it still needs humans to evaluate its fitness function.
don't you think that a team trying to build this could provide a free offering where users get free algo-generated music in return for 1-10 voting on a song-by-song basis. given enough time and votes, i suspect that the algo could get remarkably good at delivering satisfaction.
On emotions, that's a characteristic of life. With the consciousness we possess, without emotions we would quickly realize that life isn't worth living. I doubt that a "true AI", one with consciousness, will want to live without emotions. And about dogs, we haven't built anything as sophisticated yet ;-)
On AlphaGo, personally I'm not impressed. It's still raw search over the space of all possible moves, combined with neural networks and these techniques do not have the potential to yield human-level intelligence.
On logic, we have enough as to be able to build AlphaGo (also aided by computers and software that we've built, in a man-machine combination, get it?). Can a computer do anything resembling that yet? Of course not, because for now computers are just glorified automatons.
It doesn't 'always'. Advanced chess is already dead, and judging from the pro commentaries, they currently are worse than useless in an 'Advanced go' setting. That may change, but given how much faster computer Go is reaching superhuman levels than computer chess, the 'Advanced go' window may have already closed.
There is only selection.
Meat is just a phase.
Don't worry about the machine. Even in Star Trek TNG, Data can outperform everyone in every task, but was never truly happy!
Use the computers to engineer ourselves to "superhuman" capabilities.
You think we have one now?
"Naches" from our Machines https://www.edge.org/response-detail/26117
>Naches is a Yiddish term that means joy and pride, and it's often used in the context of vicarious pride, taken from others' accomplishments. You have naches, or as is said in Yiddish, you shep naches, when your children graduate college or get married, or any other instance of vicarious pride. These aren't your own accomplishments, but you can still have a great deal of pride and joy in them.
>And the same thing is true with our machines. We might not understand their thoughts or discoveries or technological advances. But they are our machines and we can have naches from them.
AI is more likely to evolve into a tool to be used by the few to control the many.
But, then also consider that many of us are working (i.e. as in 'working for the man') many more hours than our parents did in fields that require significantly more focus, concentration, and mental energy. Even the article notes that people have been facing increasing stress and feelings of being rushed since 1900 and 1965.
Maybe you're in a cushy field. But, most people that I know only have time for 'zoning out' and recovery, rather than in pursuit of true leisure.
At some point it figured that the Ko fight at the bottom was already won. Hence that white move at the top which nobody saw coming.
Another interesting moment was when Michael Redmond said "A human would typically not spend too much time thinking on this obvious move". This was the move on the right-hand side somewhere. What this tells me is that human players rush through some moves because they seem obvious but since Alpha-Go is a machine, it does not care about obvious and non-obvious. It's calculating the entire board through to the end and is not interested in "local fights".
To some relatively small depth, right? I hear the estimate that all possible moves in a Go game probably can't be physically represented in the universe (unless we learn much more about the structure of games' evolution).
This objective function is why Go playing AI jumped hugely in the last 10 years.
Also, that forced move it's very obvious to us that it's forced, but AlphaGo might not have this concept.
Also I'm not entirely sure how AlphaGo's time management works, but it's doing the same thing for every move—populating the game tree as deeply and intelligently as it can. It may just look for "30 seconds" on every move and then take the best bet meaning it's a more thorough and exhaustive reader than any human.
I wonder if even the idea from the AGA stream today, to get all the best pros in the world together and challenge AlphaGo as a team, is enough.
Perhaps releasing the core AlphaGo as open source (to the extent it's not dependent on internal Google machinery), or at least publishing its trained model, may be the next step. Let people "challenge themselves" however they want.
EDIT: Also, Lee Sedol had his time in the sun, but commiserations to Ke Jie. He's just 19, already number #1 in the world, his whole career in front of him... and this happens.
Has this been tried? That is, have Players 2-9 (or some subset) ever competed as a group against a dominant Player 1? Unless it's been tested, I wouldn't take it for granted that a group would beat an individual.
> He later said, "It is the greatest game in the history of chess. The sheer number of ideas, the complexity, and the contribution it has made to chess make it the most important game ever played."
There Go Seigen, the genius who brought about the Go revolution that Redmond talked about at post-game conf today, was beaten by Honinbou Shusai + his students.
Also, "lesser" pros routinely call out mistakes in master games (and the masters agree). Games are often decided by easily avoidable mistakes, even at the highest level.
Also I mentioned in a previous post...the human style of playing go needs to adapt to AlphaGo. That's why the commentators say "oh that was odd" since a human would not make that move as its unorthodox, but turns out to be right.
If the top 1-10 had a chance to play AlphaGo privately for months they may have a better chance.
I can think of a worthy goal for AlphaGo: make a program which can play better than top pros which runs on a macbook pro.
I'm an avid tennis fan. If we build a humanoid that can run faster, hit harder, hit more accurately and never gets tiredI would say...good job...Now get out and put the humans on the court.
We enjoy relating to the players, see how far they push their boundaries, see them make mistakes, recover from mistakes...
And on that note, hopefully there will not be cheating scandals like in chess where players have an ear piece and someone in the back communicates what move to place based on computer output.
Chess competitions are still going strong.
Also, human runners do not compete against cars.
Lee Sedol said he could beat AlphaGo, based on the Fan Hui games. Ke Jie said he could beat AlphaGo, based on the Lee Sedol games.
Ke Jie belongs to a similar category than Lee Sedol, and we could see how Lee Sedol was completely dominated by AlphaGo, 3-0 so far. It is not unreasonable to say AlphaGo will most likely beat Ke Jie, and even if that doesn't happen the first time, AlphaGo can be improved by adding more infrastructure and training time.
Because the process might tell us something interesting and useful about AI and ourselves.
There has been an annual humans vs computers challenge match every year since 2004, and in 2015 computers won with David Wu's software "Sharp". Despite the very high branching factor, standard chess AI techniques turned out to be applicable when combined with high quality hand written heuristics for positional evaluation and candidate move generation. The software is described in detail here:
In 1978 chess IM David Levy won a 6 match series 4.5-1.5 - he was better than the machine, but the machine gave him a good game (the game he lost was when he tried to take it on in a tactical game, where the machine proved stronger). It took until 1996/7 for computers to match and surpass the human world champion.
I'd say the difference was that for chess, the algorithm was known (minimax + alpha-beta search) and it was computing power that was lacking - we had to wait for Moore's law to do its work. For go, the algorithm (MCTS + good neural nets + reinforcement learning) was lacking, but the computing power was already available.
Humans have a tendancy to want to win the battle, or to get too focus in a local area. I think that's a way AlphaGo is coming up with an extra move here or there which is making a difference in the fight later.
It's certainly a feature of the best Magic: the Gathering pros, for example - their play is marked by the cards they play around, even when seemingly far ahead.
[1] komi is a number of points that is added to white's score to compensate for the disadvantage of moving second. When a non-integer komi is used, such as in this match, you cannot have a tie because scoring on the board is always integral.
https://en.wikipedia.org/wiki/Peak_Performance_(Star_Trek:_T...
This is our Deep Blue moment folks. a history is made.
This time the computer did not win out of pure bruteforce. Deep Blue relied on an opening book and massive computational power to explore the game tree. After the opening it was pretty much on its own, bruteforcing moves.
This technology used a neural network trained with hundreds of thousands of games which provided the pattern matching aspect, combined with the bruteforce move sequence reading, the montecarlo tree search... and 1200 CPUs + 600 GPUs.
Note though that AlphaGo almost certainly uses single precision arithmetic -- for neural networks even single precision is overkill.
Also the Top500 list is based on Linpack, which measures performance for computations that are pretty strongly interconnected across the different processors of the system. AlphaGo's Monte Carlo tree search problem is more embarrasingly parallel, with evaluation of different positions really being independent computations.
It is much easier to make systems that can handle embarrasingly parallel loads than the highly interconnected loads handled by the top500 supercomputers. So even though the flops are comparable, the systems are not.
While DNN+RL+Tree search is cool, the hardware requirements for AlphaGo to play at this level are staggering and only supported by large marketing budgets :)
IBM lost in 1996, 2-4, and then won in 1997, 3.5-2.5. If they had played a third match with Kasparov, especially a longer match, it is not at all clear that they would have won.
Kasparov asked for a third match of 10 games, to be played over 20 days, but IBM would not give it to him.
Of course, you can change rules as much as you want and generate new marketing events. Personally, I would like a match where the AI is only allowed the energy a human uses. I guess, AlphaGo would lose with only 2000 kcal (2.3 kWh). Not sure about chess.
Pocket Fritz (HIARCS) was winning Grandmaster tournaments and got rated at 2800-2900 Elo - on a smartphone in 2009 using 1W worth of power without offloading to a GPU. The GPU in a more modern phone like the Samsung Galaxy S5 can crunch 140 GFLOPS - an order of magnitude more than Deep Blue (11 GFLOPS). Allowing 200W of power draw raises your computation power into the TFLOPS area (using a PS4, say). No contest - even without two decades of chess engine improvement over Deep Blue.
http://www.nature.com/news/the-go-files-ai-computer-wins-fir...
For me, the key moment came when I saw Hassabis passing his iPhone to other Google executives in our VIP room, some three hours into the game. From their smiles, you knew straight away that they were pretty sure they were winning – although the experts providing the live public commentary on the match that was broadcast to our room weren’t clear on the matter, and remained confused up to the end of the game just before Lee resigned. (I'm told that other high-level commentators did see the writing on the wall, however).
But maybe this is all just human prejudice... i.e. what this really goes to show is that in the final analysis all board games we humans have inveted and played are "trival", i.e. they are all just like tic-tac-toe just with a varying degree of complexity.
Same could probably be accomplished with Go.
What is most interesting for me is that the nature of solving the problem "how do I win at Go?" is one that has not been, historically, one that computers could solve. Compute the ballistic trajectory of an artillery shell? Easy. Compute a winning strategy on the fly? Impossible. But by creating tools that allow computers to work on those problems we open up the things that can be improved and automated and that has historically improved the experience.
https://www.facebook.com/yudkowsky/posts/10154018209759228
Sample:
At this point it seems likely that Sedol is actually far outclassed by a superhuman player. The suspicion is that since AlphaGo plays purely for probability of long-term victory rather than playing for points, the fight against Sedol generates boards that can falsely appear to a human to be balanced even as Sedol's probability of victory diminishes. The 8p and 9p pros who analyzed games 1 and 2 and thought the flow of a seemingly Sedol-favoring game 'eventually' shifted to AlphaGo later, may simply have failed to read the board's true state. The reality may be a slow, steady diminishment of Sedol's win probability as the game goes on and Sedol makes subtly imperfect moves that humans think result in even-looking boards...
The case of AlphaGo is a helpful concrete illustration of these concepts [from AI alignment theory]...
Edge instantiation. Extremely optimized strategies often look to us like 'weird' edges of the possibility space, and may throw away what we think of as 'typical' features of a solution. In many different kinds of optimization problem, the maximizing solution will lie at a vertex of the possibility space (a corner, an edge-case). In the case of AlphaGo, an extremely optimized strategy seems to have thrown away the 'typical' production of a visible point lead that characterizes human play...
One could make the same statement of two humans of different ranks playing each other. In some cases, the more highly ranked player might be the only one who knows who's winning.
This doesn't mean that the lesser ranked player can't improve over time. It just means that at that moment, that player is inexperienced or less skilled, which we already knew.
I respect your work a lot. I've studied and used ML myself, including gensim, in industry. I've given nowhere near your level of contribution to society / the field. My opinion is true AI is quite a ways off. I haven't read anything from a ML researcher that says it isn't. Perhaps you weren't saying true AI is nearer.
So let me ask you this. What would you consider to be "true AI"? At what point are you willing to say, "Okay, that's it, computers are just plain smarter than we are?" Because, frankly, it seems to me that that day is getting closer and closer.
Saying that AIs can't be smarter than humans because they don't think and act like humans is like saying that airplanes don't "truly" fly because they don't flap their wings.
"The question of whether a computer can think is no more interesting than the question of whether a submarine can swim."
– Edsger W. Dijkstra
[0] http://www.antipope.org/charlie/blog-static/fiction/accelera... (readable online at http://www.antipope.org/charlie/blog-static/fiction/accelera...)
Other things that humans can do that machines can't yet:
* change other human's minds
* contribute to the state of the art of human knowledge
* determine the difference between a human and a machine.
Computer says no
> contribute to the state of the art of human knowledge
Genetic algorithms have designed circuitry that we failed to even understand at first but that did work.
> determine the difference between a human and a machine.
For now.
> Genetic algorithms have designed circuitry that we failed to even understand at first but that did work.
This is an excellent example of what I think of as not machine intelligence. If humans can't understand it then it's something entirely different that we need a different word for - an "artefact", perhaps. Meaningfully contributing to the state of the art of human knowledge requires being built upon. If these genetic algorithms can explain how they can be incorporated into the design process by humans, that's intelligence. If they are similar to being able to evolve a mantis shrimp by fiddling with DNA, that is marvellous but not what I would regard as intelligence.
We apply the same standard to human intelligence: someone who can multiply numbers very fast but not explain how they can do it is a savant; someone who can discover and teach other people a faster way of multiplication is intelligent.
Savant literally means 'one who knows', and they're not required to explain to you how they know, it's up to you to verify that they do. Just like a chess grand master doesn't have to prove to you he or she is intelligent, it's enough that they beat you. They are under no obligation to prove their intelligence to you by teaching you the same (assuming you could follow in the first place).
> Meaningfully contributing to the state of the art of human knowledge requires being built upon.
No, it requires us to understand. But we will not always be able to (in the case of those circuits we eventually figured it out, but not at first). And in Chess we did too, computer chess made some (the best) chess players better at chess. But there is no reason to assume this will always be the case and that's a limit of our intelligence.
Though in all honesty, I think a lot of people just want to see a machine with emotional "instincts" and an understanding of tribal status-hierarchy dynamics such that you can empathize with it. A lot of people would consider a machine that accurately simulated a rather dumb chimpanzee to be "smart enough" to qualify as AI, even if it couldn't do any useful human intellectual labor.
Alan Turing would give the system the "Turing test". If a computer can fool a human into thinking it's a human, then it is true AI, according to Turing.
I think that's a pretty good test. Some would argue that this is already possible with some advanced natural language processing systems. But these are not extensive tests, from what I've seen. People have to decide if the system is machine or human after just a few minutes of interaction. Turing probably meant for the test to be rigorous and to be performed by the smartest human. Deciding a conversational partner is a human after 5 minutes of interaction is not enough. 10 years might not be enough. I honestly couldn't say when enough is enough, which is part of what makes Turing's definition so complicated, even though it seems simple on the surface.
I would add that currently, systems cannot set their own goals. There is always a human telling them what to be good at. Every machine-learning-based system is application-specific and not general. There are some algorithms that are good at generalization. You might be able to write one algorithm that's good at multiple tasks without modifying it at all. But from what I've seen, we are nowhere near being able to write one program that can be applied universally to any problem, and we are even further from one that can identify its own problems and set its own goals.
As humans, do we even know our own goals? Stay alive, right? Make the best use of our time. How does the quality of "intelligence" translate to computers which are, as far as they know, unconstrained by time and life or death? What force would compel a self-driven computer to act? Should we threaten them with death if they do not continue learning and continue self-improvement? If I hold a bat over my laptop and swing at it, does it run my program faster? If I speak to it sweetly, does it respond by doing more work for me? Further, are animals intelligent or are they not?
It gets pretty philosophical. What are your thoughts?
> Saying that AIs can't be smarter than humans because they don't think and act like humans is like saying that airplanes don't "truly" fly because they don't flap their wings.
That's just semantics. I think any conversation about this must define intelligence really carefully. We all perceive things differently, so it's impossible to be sure we're talking about the same thing. Maybe that's one other quality of intelligence that separates us from computers. Every computer perceives a given input the same exact way. Can we say that about humans? If there were another dimension with the same atomic makeup as our own, would I think the same things as I do in this dimension? Are my thoughts independent or dependent upon my environment? Is anything truly random?
Anyway, for me, independent goal setting is a key element of true AI. And philosophically speaking, I believe we can't guarantee that we set our own goals independently. Most of us have a strong feeling that we act of our own volition and fate does not exist. And I think that's right. But what if there is no randomness and we are entirely products of our environment? Then under this definition, we don't have independent goal setting and we are not true AI.
Thanks for asking my thoughts.
Look, there's no doubt that computers can outperform humans in specific tasks. There's no doubt that AlphaGo is intelligent when it comes to Go, but on the other hand it would be completely incapable of tackling a different congitive task- say, language, or vision, or discriminating between say two species of animal [Edit, since there seems to be confusion on this: you'd need different training data and another training session to perform well at a different task].
That's a limitation of our current systems. They generalise badly, or they don't model their domain very well. You have to train them again for each different task that you want them to undertake and their high performance in one task does not necessarily translate in high performance in another task.
Humans on the other hand are good at generalising, which does seem to be necessary for general intelligence. If we learn to play a board game, we can take the lessons from it and apply them in, I don't know, business. If we learn maths, we can then use the knowledge [Edit: of what constitutes a well-formed theory] to tackle physics and chemistry. And so on.
So, let's say that "true AI" is something that can show an ability to generalise from one domain to another, like humans do, and can be trained in multiple cognitive tasks at the same time. If we can do that, then computers will already be super-human, because they can already outperform us in terms of speed and precision.
There's a lot that suggests that humans and machine learning algorithms learn in very different ways. For instance, by the time a human can master a game like Go they can also perform image processing, speech recognition, handwriten digit recognition, word-sense disambiguation and other similar cognitive tasks. Machine learning algorithms can only do one of those things at a time. A system trained to do image processing might do it well, but it won't be able to go from recognising images to recognising the senses of words in a text without new training, and not without the new training clobbering the previous training.
To make it perfectly clear: I'm talking about separate instances of possibly the same algorithm, trained on a different task every time. I'm not saying that CNNs can't do speech recognition because they're good at image processing. I'm saying that an instance of a CNN that's learned to tag images must be trained on different data in a different time if you also want it to do word-sense disambiguation.
And that that is a limitation, that stands in the way of machine learning algorithms achieving general intelligence.
A human brain (or any other animal brain for that matter) is almost infinitely more advanced and computationally efficient than state of the art machine intelligence, even without taking things like thoughts, emotions and dreams - which we currently do not understand at all - into account.
It's a huge accomplishment for machines to be able to win over humans in games like chess and go and <insert game here>, but these are games originally designed for humans - by humans - to be played recreationally and I think we shouldn't read too much into it.
A simple game that this architecture would fail at is Simon [0], where you are presented a sequence and then are tasked to replay the sequence.
I'd also like to point out that you didn't define "smart" or "intelligent" either. The fact is, it's a very very difficult concept to define.
Reposting this comment from another thread:
The "moving goalposts" argument is one that really needs to die. It's a classic empty statement. Just because other people made <argument> in the past does not mean it's wrong. It proves nothing. People also predicted "true AI" many times over-optimistically; probably just as often as people have moved goalposts.
When a computer itself makes a persuasive argument that it is intelligent (without coaching.)
I have always believed AI is possible, and I am undecided whether current techniques alone will get us there, or whether other breakthroughs are needed, but I have no time for premature claims of success.
while I agree that this does happen, I don't think it did in this case - that is, I don't remember anyone saying that they would take a computer beating the top human in Go as evidence of "true AI"
So perhaps the problem isn't our brain's hardware as such but the operating system that runs on top of it.
Yes, but all computers do those things millions of times faster than even savants. And the computers are getting faster at it every year, savants today aren't any more clever than savants 100 years ago.
If the utility function is about winning, going though easily-evalutable board positions might be the straightforward route.
If the utility function is to win while minimizing your estimate of of you losing, you might see different results.
Plus if you know that you are playing against a stronger opponent, your prior might bias your perception of the board situation.
It's common for the opening to pass by without feeling behind--Go doesn't have set openings to the same extent as chess, but if you play joseki, you might have an opening where the weaker player feels ok. Even that is not guaranteed, but once you start fighting, you will quickly feel the strength difference.
Of course you can play a move that looks good when you play it, but against a much stronger player, it doesn't take long for it to look bad.
I also wonder if when Alpha Go plays itself, does it always win when it is black or white, does it always tie, or is it a mix?
So much to learn from this.
It would be interesting if AlphaGo's maneuvers remain opaque to humans for an extended period of time. Can anyone at this point say with confidence that its strategies will indefinitely remain unknown?
Given the computational power (=number of cpu/gpu) is fixed and the time for each move to be done is fixed.
Perhaps a future "dr. evil / skynet" A.I. will first try to conquer the microchip production plants to increase its computational power and memory. goodbye taiwan, goodbye south korea, goodbye usa...
I still don't agree that this is the case, and I don't care what a thousand Google-hyped press releases say, beating the best human player in anything is not "superhuman" and "superhuman" performance has not been achieved by anything yet. [1]
Why do I think so? Two reasons.
One, because you can be entirely human and still beat all other humans, without fail for a very long time. Long "winning streaks" in professional sports are very well documented. For instance Rocky Marciano went entirely undefeated in his whole heavyweight boxing career. In chess, Mikhail Tal went undefeated for 95 games. And so on, so forth.
Of course most humans' winning streaks end eventually. That's because our performance degrades over time. When a computer wins against the best player, it keeps on winning, and in fact it actually gets even better over time.
Still, and this is number two: in most instances where a computer does better than a human, we don't claim superhuman performance. Automatic calculators, going back to mechanical calculators, have been better than humans at arithmetic for a very, very long time. I would wager that nobody discusses pocket calculators as exhibiting "superhuman" performance. You only hear this sort of claim when it comes to Deep Blue, AlphaGo or Watson.
So maybe we need a better definition of what it means to be "superhuman" that covers both pocket calculators and AlphaGo. Without one, I don't accept that the performance of AlphaGo can be said to be superhuman, unless pocket calculators' performance is also celebrated as superhuman.
_______________________
[1] I'm perfectly willing to go even further than that and say that we can't make machines that have superhuman intelligence and that even if we did, we wouldn't be able to recognise them (this last bit is similar to what the GP says).
AlphaGo is rather different in that it actually has a representation of the game of Go and it knows how to play. I don't doubt at all that it's intelligent, in the restricted domain it operates in. But I do doubt that it's possible for an intelligence built by humans to be "superhuman" and I don't see how your one-liner addresses that.
Why would an intelligence built by humans not be able to be superhuman? The generally accepted definition seems to be "having better than human performance" in which case it seems we've done it many times (like with calculators).
Hm. Strictly speaking I believe my pocket calculator has an FPGA, a general-purpose architecture that in my calculator happens to be programmed for arithmetic, specifically. So I think it's accurate for me to say that, although the calculator has a program and that program certainly is a representation of arithmetic, I have to provide the interpretation of the program and reify the representation as arithmetic.
In other words, the program is a representation of arithmetic to me, not to the calculator. The calculator might as well be programmed to randomly beep, and it wouldn't have any way to know the difference.
(But that'd be a cruel thing to do to the poor calculator).
I don't think there's a generally accepted definition and I don't agree that performance on its own is a good measure. Humans are certainly not as good at mechanical tasks as machines are -duh. But how can you call "superhuman" something that doesn't even know what it's doing, even as it's doing it faster and more accurately than us?
Take arithmetic again. We know that cat's can't do arithmetic, because they don't understand numbers, so it's safe to say humans have super-feline arithmetic ability. But then, how is a pocket calculator super-human, if it doesn't know what numbers are for, any more than a cat does? There's something missing from the definition and therefore the measurement of the task.
I don't claim to have this missing something, mind you.
>> Why would an intelligence built by humans not be able to be superhuman?
Ah. Apologies, I got carried away a bit there. I meant to discuss how I doubt we can create superhuman intelligence using machine learing specifically. My thinking goes like this: we train machine learning algorithms using examples; to train an algorithm to exhibit superhuman intelligence we'd need examples of superhuman intelligence; we can't produce such examples because our intelligence is merely human; therefore we can't train a superhuman intelligence.
I also doubt that we can create a superhuman intelligence in any other way, at least intentionally, or that we would be able to recognise one if we created it by chance, but I'm not prepared to argue this. Again, sorry about that.
Ok flippancy aside, there are two problems that make techniques like this single-domain: network design and network training.
The design, uses multiple networks for different goals: board eval (what boards look good) and policy (which moves to focus on). Those two goals, eval and policy, are very specific to go. Just like category layers are specific to vision and LSTM is specific to sequence learning.
Network training is obviously hugely resource intensive -- and each significantly complex problem would need such intensity.
It is amazing the variety of problems DNNs have been able to do well in. However, the problem of network design and efficient training are significant barriers to generalization.
When network design can be addressed algorithmically I think we may have an AGI. However, that is a significant problem where you automatically add another layer of computational complexity so it is not on the immediate horizon and may be 50+ years down the road.
If top humans get beat 5-0 with significant handicaps then it is likely AlphaGo is superhuman. However, it is expensive to run AlphaGo so it is unlikely that we will know the true strength of AlphaGo for a while (more challenges) or until hardware catches up.
Update: typos and clarifications
[1] https://en.wikipedia.org/wiki/Harry_HaftTo be honest- I don't have one. My intuition is that we can't have a good definition of "superhuman intelligence" because having one would require us to demonstrate superhuman intelligence ourselves. Which is obviously a contradiction.
So I think the word "superhuman" must imply a fair competition, in the sense that the participants are competing using comparable approaches. For some definition of comparable.
I wonder how many handicap stones do Lee Sedol, or Ke Jie for that matter, need to have a shot at winning even one game.
Minor quibbles: 1) taking territory or being up on points are the same thing, and 2) in go there are many other things to optimize for besides territory/points - influence, shape, efficiency, and sente being examples of this.
Here's where I think you're right. It seems from the commentary during the matches that the pro player thought the game to be close...on every match.
Then the tides turned...or so we thought...and the game always swayed in AlphaGo's favor.
I like how they were saying that the computer changed up the game style as well.
All of this is very interesting, and I see that they have games 4 and 5 listed however I wonder if they are going to go thru with them and play them as scheduled.
Maybe to see if AlphaGo can sweep the series? Or show it wasn't a fluke of any kind?
I hope they play the last two...just to see how AlphaGo plays them. :D
What do you mean by a one-stone handicap? Just no komi?
> In the case of AlphaGo, an extremely optimized strategy seems to have thrown away the 'typical' production of a visible point lead that characterizes human play. Maximizing win-probability in Go, at this level of play against a human 9p, is not strongly correlated with what a human can see as visible extra territory - so that gets thrown out even though it was previously associated with 'trying to win' in human play.
Pros have been aware for centuries that "visible extra territory" is a poor indicator of win probability. They use the term "thickness" (in English) to denote the positive potential resulting from a strong, safe, influential position, independent of current territory. Quite often a pro game will be a battle of territory vs thickness.
What is this?
http://lesswrong.com/lw/n4l/safety_engineering_target_select...
Hang on. Where are we going to find this magic agent that can do "even more" in "a rich and complicated domain whose rules aren't fully known" (the real world, as opposed to a Go board) than in a game of Go? What are we going to train such a learner with, if we ourselves don't fully know the rules of the domain, as you point out?
Even if a learner somehow magically found a superhuman path to perfect reasoning which is unavailable to us, entirely by accident and purely on its own, why would we select it from other trained learners to keep and foster further, if we think it's actually pretty dumb, rather than magically smart?
You're saying at some point that a fantastic paperclip maximiser might achieve superhuman intelligence and then lie in waiting, poised to turn us all into paperclips only when it knew it was safe to make its move, basically. But, how is it going to become that smart in the first place? It has to be smart enough to know that it must bide its time, but dumb enough that its time hasn't come yet. Sounds like a bit of an impossible double-bind there.
Just because sufficiently advanced AI may look like magic at first, it doesn't mean we should start all believing in magic because it may just be sufficiently advanced AI in disguise.
This does not make any sense if you assume general intelligence. A physicist who makes a discovery also was not trained to know about this new law or rule beforehand (how could he?).
Woa there. The Deepmind Atari-playing AI was too specialised for each particular game. It had a reward function that translated the score for it. It couldn't learn the importance of the game score on its own, just from "looking at the pixels" and it couldn't learn the significance of the score display and how it changed as a result of its actions on its own. All this had to be hand-coded. And if I've seen a hint that this means it couldn't really-really generalise to other games, like Deepmind claimed, then that's this bit that you report yourself:
>> Deepmind ... did reuse the widely known core insight of Monte Carlo Tree Search.
That is no mere detail. That is the crux of the matter, right there. Deepmind used their architecture to improve MCTS far enough that it could beat Lee Se-Dol. They didn't just add MCTS to their already general-game playing system.
Because the original Atari-playing AI was completely useless for playing Go, a game that doesn't have a score and looks nothing like an Atari game. So it wasn't very general at all, despite Google's and DeepMind's claims to the contrary.
"We also tested against the strongest open-source Go program, Pachi, a sophisticated Monte Carlo search program, ranked at 2 amateur dan on KGS, that executes 100,000 simulations per move. Using no search at all, the RL policy network won 85% of games against Pachi."
AlphaGo does use MCTS, but it seems that most of its improvements are actually coming from the deep reinforcement learning approach.
But in any case, I'm not necessarily disputing that. I'm particularly refuting the claim that the AlphaGo architecture is identical to the one that learned to play Atari games and that Deepmind have advertised as a general game-playing agent.
My comment here is specifically in reply to the GP who repeated this claim, but I'll dig up the relevant link if you're interested.
Hassabis has said that in the next few months they want to try and get up to current AlphaGo performance without using any MCTS at all.
Keep in mind that as well as AlphaGo plays, it's still extremely far from playing optimal moves.
This suggests that Go space was already very well understood and there aren't radically different play styles we've somehow overlooked. AI is not a magic wand that radically changes how the game works.
However there may be some bias due to AlphaGo having trained on human games and playing against a human. The real proof will happen when they redo the retraining from scratch.
We put a man on the moon, which encouraged decades of optimism about the near-future colonization of space... unfortunately it turns out we are nowhere near solving the hard problems associated with being able to live in space.
We spelled out the letters IBM in xenon atoms, five atoms tall, which encouraged optimism about near-future molecular manufacturing... unfortunately it turns out we are nowhere near solving the hard problems associated with making useful things with atomic precision.
We achieve superhuman performance in simple games, which encourages what should be optimism that becomes reflected by the distorting mirror of the zeitgeist of the times into pessimism, about near-future AI... unfortunately it turns out we are nowhere near solving the hard problems associated with making our tools something better than depressingly dumb.
In each case, the flurry of optimism turns as the decades pass into confusion, anger, soul-searching about why the expected follow-on progress has failed to materialize, fading slowly into despair, then into nothing as the optimists die of old age, one by one, and the rate of technological progress becomes slower with every passing decade. For this conversation is necessarily happening at a tech level that allows it, but that provides no guarantee whatsoever that the species that reached tech level X will proceed to tech level X+1. Moore's law has officially expired, and that was the last major area where we were making rapid sustained progress.
It's hard to say we shouldn't do 'flag and footprints' demonstrations; am I really willing to bite the bullet and say we shouldn't have put a man on the moon? I don't know the answer to that. I am, however, convinced that we should remember not to take them too seriously.
We put a man on the moon in 1969, and we are just now getting privately-funded spaceflight and satellites. We made the first computers in 1945; we got them on every desk in the 1980s. The first steamboat was built in 1783; the transatlantic shipping industry didn't transition to steam until the 1840s & 1850s.
Probably some kid who's watching AlphaGo's game today is the person who, late in life, invents the first strong AI.
For example, long stays in large space stations, landing craft on Mars, successful isolation sustainability experiments, 14nm chips and understanding of quantum concerns, multi-layer chips, beginnings of optical computing, biological/genetic chemical synthesis, neuromorphic chips, etc.
Deep reinforcement learning is undeniably a major step forward for general intelligence.
You and others with your belief system will still be in denial up to and past the point where your species becomes irrelevant as the superintelligent AIs arrive within a few decades.
The one solace was that Lee Sedol got his ko =) however, AlphaGo was up to the task and handled it well.
Would StarCraft be the last game AI has to beat?
After that... then we have games that involve physical movement. Although I'm not sure what would be fair for strength & speed of robots vs humans. I'm imagining a team of robot soccer players, for example.
The from playing Call of Duty to controling a robot with a gun in the real world it's a small step :)
With an AI, if we base it on pixels, what they will do is spin around 360 degrees about 60 times a second while moving forward so that it can maximize all of the pixels inputted at a time. It would compare it against the stored level design and shoot at anything that is off (or against character models). I can just see that it's not going to behave like anything we'd expect.
If you look for an RTS game, StarCraft is an obvious candidate. It has a relatively active community of AI writers and AI competitions, it's a classic of competitive play, ...
Real-time games are hard to compare, because the interface really makes a difference. Does the AI get to talk directly to the game engine? Or does it have to use screen capture, understand the image and then use a robot arm to press keys, or something in between? If it gets to see precise positions and to input commands directly, even relatively "simple" programs can do crazy micro tricks and get massive advantages that way, even if their general macro-strategy is bad. If you do the other way you need image analysis and robotics, Google should have the tech for both ;)
My feeling is that AlphaGo could be a couple of stones stronger than the best pros. Considering that dans are capped at 9 but there are 9 dans that always win against some other 9 dans, the best pros could be 10 dan on their scale. That would also be 10 dan on the amateur scale (European, Americans are usually one stone gentler and Japanese much more than that). AlphaGo could be 12d amateur or, don't know, 14-15 dan pro? (pro levels are on a finer scale)
I wonder if Google will make it available to pros outside these official matches. It's not that Deep Blue played so much after winning with Kasparov in 1997, right? Kasparov asked a rematch and IBM refused. Deep Blue never played again.
Our computers are stronger than it now, mainly thanks to advances in the algorithms. Deep Blue used to crunch many more positions per second than the current chess engines do on standard PC. Still they're stronger. AlphaGo is running on 1920 CPUs and 280 GPUs, so either research trims down its algorithms or we wait for CPUs with 1024 cores. But at the current size they won't fit below the keyboard of a laptop :-)
Game 1: https://www.youtube.com/watch?v=6ZugVil2v4w
I recall Andrew mentioning that Myungwan will be back to comment games 4 and 5?
Anyway, I think AlphaGo is a great training companion. I think Lee felt he's learning.
Finally, I also feel that while experience is crucial, the older generation would flush out by the younger generation every decade. I wonder if age really play a role in championship - not that AlphaGo isn't considered a 1000 years old "human" given it has played thousands of games already.
Our model of representation of Go fails at expressing the game/strategies of AlphaGo is showing, we are communicating in the board in different languages, no wonder everyone looking the at games is stomped by the machine "moves".
Our brains lack the capacity of implementing such algorithms (understanding such languages), but we can still create them. We might see in the future engine A played against engine B and enjoy the matches.
No one is surprised by a machine doing a better job with integer programming/operational research/numerical solutions etc.
http://gizmodo.com/how-ashley-madison-hid-its-fembot-con-fro...
Considering the large training data set, handicap games would certainly be included. It's a great way to test aggressiveness and the effectiveness to reduce the opponents' territory whilst building your own.
As an amateur, handicap games where you try to overcome 3 or 4 stones are some of the most educational experiences. It requires a lot more creative thinking and perseverance to come back from such a disadvantage.
Given AlphaGo has learned all played games from a database, its more like AlphaGo is the essence or the sum of all experience gained from those games played. AlphaGo is better than every player (perhaps), but it has not pushed the game further by itself.
It would be fascinating to see how early AlphaGo assigned very high probability of it winning. It would also be interesting to see if there were particular moves which changed this assignment a lot. For instance, are there moves that Lee Sedol made for which the win probability is very different for the AlphaGo moves before and after?
http://www.shanghaidaily.com/national/AlphaGo-cant-beat-me-s...
"Demis Hassabis, Google DeepMind's CEO, has expressed the willingness to pick Ke as AlphaGo's next target."
https://www.youtube.com/playlist?list=PLqYmG7hTraZA7v9Hpbps0...
(It also includes the videos of the first 2 matches)
In 2 years? In 1 year? In 3 months?
The current answer is, still a long way off. With massive forecasts in GPU power over the next few years it may be reasonable for people to have desktops in their home with this kind of power possibly within a decade. I won't even give a timeline for phones because of the power requirements.
Perhaps longer given the restrictions (and huge problems) in further down scaling of the known silicon technology. GPUs are now at 28nm and the newest Intel CPUs on 14nm.
[1] http://techreport.com/news/27978/nvidia-pascal-to-feature-mi...
Unless they can optimize it quite a bit, we won't be seeing this on a mobile phone anytime soon. Perhaps you'll be able to pay a good hourly rate for the cloud version.
If you provide terrain(elevation etc.) information, AlphaGo can be used to corner opponents into an area surrounded by mountains where AlphaGo is sitting on the mountains. We all know what happens after that.
Don't want to kill the party but I am completely surprised with the lack of chatter in this direction.
Source: http://www.theguardian.com/technology/2016/feb/16/demis-hass...
I'm no where near a strong player but it seems like AlphaGo is far ahead of Lee Sedol.
- It doesn't make mistakes due to pressure or fatigue; it's always playing at its sharpest
- It can read deeper than humans can
- It seems to have better judgement than most current professional players
But maybe it could have 40 of those moves to play an incredibly confusing game for humans. I think giving the machine a big handicap against pros might reveal it's true colours ...
If the player forces it to a fight, it will fight back, but that's only because losing the fight will definitely lose the game.