Go Master Lee Se-Dol Quits Because AI 'Cannot Be Defeated' (bbc.com) 106
An anonymous reader quotes a report from the BBC: A master player of the Chinese strategy game Go has decided to retire, due to the rise of artificial intelligence that "cannot be defeated." Lee Se-dol is the only human to ever beat the AlphaGo software developed by Google's sister company Deepmind. In 2016, he took part in a five-match showdown against AlphaGo, losing four times but beating the computer once. The South Korean said he had decided to retire after realizing: "I'm not at the top even if I become the number one." "There is an entity that cannot be defeated," the 18-time world Go champion told South Korea's Yonhap news agency. Lee Se-dol is considered to be one of the greatest Go players of the modern era. "On behalf of the whole AlphaGo team at DeepMind, I'd like to congratulate Lee Se-dol for his legendary decade at the top of the game, and wish him the very best for the future," said Demis Hassabis, chief executive and co-founder of Deepmind. "During the AlphaGo matches, he demonstrated true warrior spirit and kept us on the edges of our seats to the very end."
Rage quit? (Score:5, Insightful)
I doubt I'm the 'best in the world' at anything, I still play.
Re: (Score:2)
I doubt I'm the 'best in the world' at anything, I still play.
For sure, fun or maybe to train the mind seems like good reasons.
From another perspective the scenario we have provides a rough measure of that area of cognitive capacity represented in a machine.
Re:Rage quit? (Score:5, Insightful)
I doubt I'm the 'best in the world' at anything, I still play.
Indeed.
People still do competitive weightlifting, although none can beat a forklift.
People run marathons, although none can beat a motorcycle or even a horse.
Re: (Score:3, Informative)
People run marathons, although none can beat a motorcycle or even a horse.
If you are talking longer distances than a marathon, an athlete can come closer to beating a horse than http://ultrarunninghistory.com... [slashdot.org]>you might first think.
Re: (Score:3)
No human athlete can beat a jet airplane in a race. Give up now, or who cares?
Re: (Score:2)
I can beat one in an underwater swimming contest.
Re: (Score:2)
Re: (Score:2)
No, you can't. A submarine is an underwater plane/blimp thingie.
Btw, for you Eve Online players, you're not in a space simulation but a submarine one.
Re: (Score:2)
Space
Keep accelerating forever (subject to relativity, a separate issue)
Shut off engines, cruise forever
Eve Online or a submarine
Max speed tops out for your engine power
Must keep engines on or you slow down
Re: (Score:2)
Ironic the title of this sub-thread is "Rage quit?" Eve is the queen of rage quit.
Re: (Score:3)
He should have picked a shorter distance for his example. The marathon seems to be about the point where human long distance endurance starts to overtake the horse.
For example Winning Brew ran 2 furlongs (402 m) in 20.57 s, whereas the fastest 400 m run by a human is 43.03 seconds by Wayde van Niekerk.
The fastest the Kentucky Derby was ever run was 1:59.40 (Secretariat), the human 2000 m record is 4:44.79.
Both are more than twice as fast.
Re: (Score:2)
He should have picked a shorter distance for his example. The marathon seems to be about the point where human long distance endurance starts to overtake the horse.
Judging by the results of man vs horse marathons (I googled it, and it's really a thing), it seems like horses still hold a slight advantage at marathon distances. They lose consistently by the time you get to ultrathon distances, though.
Re: (Score:2)
That would seem to be horse abuse unless you were trying to save Hailee Steinfeld's arm.
Re: (Score:2)
True. The best theory so far why humans have lost their body fur is run-walking down animals in Africa until they get heat stroke. This takes ~4 hours but is an actual thing still used and documented.
Re: (Score:2)
People still do competitive weightlifting, although none can beat a forklift.
People run marathons, although none can beat a motorcycle or even a horse.
Huw Lobb beat a horse [bbc.co.uk] over 22 miles.
Re: (Score:3)
I think the difference lies in recognizing physical competitors such as a forklift, horse, etc. as non-humans and clearly outside of any kind of direct competition. Many animals can outperform humans in some way and yet the ancient Greeks never saw fit to cancel the Olympics over it. But we can't r
Re: (Score:3)
I doubt I'm the 'best in the world' at anything, I still play.
Indeed.
People still do competitive weightlifting, although none can beat a forklift.
People run marathons, although none can beat a motorcycle or even a horse.
Actually people can beat a horse at running: https://en.wikipedia.org/wiki/... [wikipedia.org] A human runner would actually have beaten a mounted courier in the famous run from Marathon to Athens for the same reason persistence hunting is a viable hunting strategy.
Re: (Score:3)
People run marathons, although none can beat a motorcycle or even a horse.
can't they?
https://en.wikipedia.org/wiki/... [wikipedia.org]
Re: (Score:3)
It sounds like he does too, he's just not going to play professionally.
30 years ago... (Score:5, Interesting)
Re:30 years ago... (Score:5, Insightful)
Pure science always seems impractical, and that would include pure computer science. But there is plenty to learn from solving previously-impossible problems which can carry over to other, more practical, problems. It isn't always obvious at first what the practical payoff will be, but that doesn't mean it doesn't exist.
In my own opinion, there is a cultural benefit to silencing those who sneer at the thought that the special thing they are good at is something that computers will never be able to do. I think that a humble recognition of what we can actually achieve through computer science is culturally healthy, and will help give us an atmosphere of clarity around such issues as economic sustainability.
Re: (Score:2, Troll)
I think it was better before in a spiritual sense although I acknowledge the amazing achievements of the Deepmind team.
It's not an amazing achievement. Google literally just applied a supercomputer to solve go. It's not AI, and it's not even heuristic, it's just a pathing algorithm from any-possible-board-configuration to a-win.
Re: (Score:2)
Right, because AI is anything that computers can't do yet. [wikipedia.org]. The instant a computer can do it, it no longer qualifies as AI.
Anyway, I think the achievement is amazing just because it has been impossible to do throughout all of human history, and throughout nearly all of my own lifetime. I am a fan of any scientific achievement that makes the impossible possible.
Re: (Score:2)
AI is intelligence that is artificial. Full stop.
Consider that evolution has (evidently) managed to produce intelligence, so there is certainly no physical law precluding intelligence from existing... Can you cite any property of the physical universe which might precludes its artificial production?
Because barring that, saying that "if a computer can do it, then it no longer qualifies as AI" is just an appeal to history, and certainly not logically sustainable.
Re: (Score:2)
Oh, that statement I made, "if a computer can do it, then it no longer qualifies as AI" was intended as sarcasm. I apologize, I realize that sarcasm is sometimes hard to detect on the Internet. That quote was a link to a wiki page about the "AI Effect," which is a perpetual re-definition of the term "AI" such that it is never an achieved goal. I was trying to point out a cultural problem, not assert that as an actual definition.
The technical jargon use of the word "AI" varies hugely from the popular use.
Re: (Score:2)
Re: (Score:3)
AI is intelligence that is artificial. Full stop.
False.
AI is the stuff they teach at a University in the portion of the computer science department that is called "AI." It involves a basket of specific algorithm categories.
Re: (Score:2)
It involves a basket of specific algorithm categories
That doesn't mean it's not intelligence or artificial.
Re: (Score:2)
Re: (Score:1)
Re: (Score:2)
"Supercomputer" is pretty generous too. Alpha Go Zero runs on four TPUs and achieves master level performance in a few days.
Re: (Score:3)
Google literally just applied a supercomputer to solve go.
No. Google's Go programs do not run on supercomputers. The training is done on GPU compute nodes, and the resulting network can run on a high-end desktop.
it's just a pathing algorithm from any-possible-board-configuration to a-win.
No. That is not how it works. Alpha-Go does not provide a path to a win, nor does it provide any rationale or explanation for its moves. The advantages to the moves it makes are often not apparent until much later in the game.
Re: (Score:2)
. The training is done on GPU compute nodes, and the resulting network can run on a high-end desktop.
That is exactly what a super computer is in our days.
Re: (Score:2)
. The training is done on GPU compute nodes, and the resulting network can run on a high-end desktop. That is exactly what a super computer is in our days.
Depends on the number of GPUs involved. I don't know if Google has revealed the size of the system used for training. Actually, I'm sure they used TPUs, not GPUs, but the difference isn't enormous.
Re: (Score:2)
You have no idea what you're talking about. Go was widely regarded as a task that would require actual AI because, unlike checkers and (conceivably) chess, a brute force method such as you describe was out of the question.
Re: (Score:3)
It is absurd to presume that a pattern-matching algorithm is limited by the problem space just because you use a certain term to describe it.
You don't need "intelligence" for an algorithm to work across a wider range of inputs than you have memory. You just need a good algorithm.
When chess computers are described as "brute force," it doesn't have the same simplistic meaning it did 25 years ago when it was actually true. A "brute force" chess computer doesn't check everything, it has a whole bunch of domain-
Re: (Score:2)
Mmm hmm.
Re: (Score:2)
Re: (Score:2)
You know, there are a fair number of papers written about how AlphaGo actually works, if you'd rather learn than make shit up on Slashdot. There are also some open implementations that describe the process in more detail, and let you experiment with it yourself.
Re: (Score:2)
yeah, and i've read them. "covering all the patterns" is a fair description (albeit very, very simplified) of what alpha-go actually does. the magic of ML was to couple pattern-recognition applied to state-space, to reinforcement learning on the action-space. the latter is old hat, the former was the "special sauce". it's not unfair to call that compression, because that's what it is. the ML extracts salient features from state-space and feeds it into the decision-maker.
Re: (Score:2)
no, you're quoting Noam Chomsky.
Re: (Score:3)
You clearly don't play Go.
And probably never even looked at the problem.
No, there no way in the universe that a Chess like program ever will play Go better than a human.
Go is usually played on a 19x19 field, each field has 3 states: white, black, empty. That corresponds to a ternary number with 361 "tri-bits". Obviously many positions are invalid or make no sense to play (as in suicide), nevertheless the possible combinations are more than atoms in the universe. Aka impossible to hold in memory or prune in
Re: (Score:2)
"or prune in a search tree" is the stupidest thing you could say.
You're saying the problem space is too big... to prune? That's just daft.
The same wrong arguments were made about chess, which also has a search space too large to search.
Re: (Score:2)
It's not an amazing achievement. Google literally just applied a supercomputer to solve go
Supercomputers existed long before Deepmind solved Go, so they did a bit more than that.
Re: (Score:2)
Learning humility is never a bad thing. It does hurt the ego though.
If knowing that a non-human can play a game really well kills the joy of it for you, then did you only play for hubris?
Re: (Score:2)
Maybe they should have just left it alone and worked on something useful to humans like driving or medical analysis rather than killing the joy of an ancient game.
They need to learn with increasing complexity. Chess has few branches, Go has many. A game like DotA a continuum and yet every rule is exhaustively defined. I think their next challenge is to open up the rule book and make an AI refuse to do things outside its limitations. Like say you're teaching it to drive a car, when should it stop driving in a blizzard? Today I think it's an entirely external process that cuts it off, the AI would just keep trying even in whiteout conditions no matter how poorly it wou
Re: (Score:2)
I think their next challenge is to open up the rule book and make an AI refuse to do things outside its limitations.
You could try limiting it to more "human" capabilities. An AI for Chess searches through millions of positions a second. A human searches through two.
The computers aren't winning on intelligence, they are winning based on raw computing power.
Re: (Score:2)
If you want, right now, you can play against Leela Chess Zero, set at 2 nodes search, and you'll probably lose.
Or you can try to play against this version here: https://lichess.org/@/MiniHuma... [lichess.org]
Re: (Score:2)
If you want, right now, you can play against Leela Chess Zero, set at 2 nodes search, and you'll probably lose.
What rating does that play at?
Or you can try to play against this version here: https://lichess.org/@/MiniHuma [lichess.org]...
Looks like it's running on a Raspberry pi. A Raspberry pi can look through several orders of magnitude more positions per second than a human can.
Re: 30 years ago... (Score:2)
Re: (Score:2)
Your'e not dumb, just ignorant.
Re: (Score:2)
To be honest, it is not a "computer algorithm" that beats humans at go, but mostly an neural network.
And the fact that we once would build neural networks in hardware with thousands was 10, 15 years ago not foreseeable.
Well, you could say: after GPUs it was the next logical step, but if you compare Go to Chess it was pretty "clear" no one will ever build a Go computer. And now, as we have one, you can ask: how much did it cost? Will we ever have potent Go programs on tablets (since the NN is trained, it sho
Re: (Score:2)
In other News: All runners quit (Score:2)
Um... no shit machines can be humans (Score:5, Interesting)
Re: (Score:2)
Because board games are considered smart work, comparable to the processes that make humans special and put them at the top of the food chain. The idea that a computer could eventually brute force every task that humans are good at gives the feeling of redundancy.
Re: (Score:2)
Still waiting for the day George Jetson gets home from work and complains how these 2 hour workdays are killing him.
Re: (Score:1)
Re: (Score:2)
It will only matter at first in some aspects, because unlike computers, humans have egos that are easily bruised. I'd guess it was a point of pride for many Go players to know that computers still couldn't compete with the best of them. Kids that grow up in a world where it's natural for computers to be superior at games of any sort probably won't be bothered by this at all.
In other aspects, it will matter quite a bit in the long run. Note how the chess world is trying to deal with electronic cheating.
Re: (Score:2)
I remember back in the dark ages, around the time Windows first came out, fighting the Microsoft Go AI against some 3rd party Mac one. On hardest, the Mac one destroyed the Windows one.
The next game I turned the Mac one down to drooling until the Windows one pwned the board and was seconds from a win, then cranked the Mac one up to Commander Data, could it perform a miracle? It could. I watched, unbelievably, as the god strode through there as against a chimpanzee.
Re: (Score:2)
I had to sit there plugging each one's moves into the other's board.
Yes Shit (Score:1)
Um... no shit machines can be humans
You can't no-shit this one, people were claiming Go was of a complexity insoluble by AI not that long ago...
Re: (Score:3)
this is why we don't let cars compete in the 1000 yard dash. I don't understand why it matters that computers can beat people at anything beyond being a scientific curiosity and an opportunity to automate boring, repetitive work.
Back in the 1950s when computers really first started to come into some use, it was thought that the only way a computer could ever beat a human at chess (a simpler game than go) was if AI got really good and basically the computer would out think the human. Nobody realized for a long time that the problem would be solved by what is, effectively, cheating. AI doesn't really out think the humans. It just finds maybe novel ways of doing searches or estimating future probabilities, but it's not a fair con
Re: (Score:2)
Nobody realized for a long time that the problem would be solved by what is, effectively, cheating. AI doesn't really out think the humans. It just finds maybe novel ways of doing searches or estimating future probabilities, but it's not a fair contest
This used to be true with brute force engines like Stockfish, capable of searching through billions of positions.
Modern neural net engines, such as Leela Chess will beat you easily while only looking at a handful positions.
Re: (Score:2)
"we don't let cars compete in the 1000 yard dash"
However, we do let them compete in the 440 yard dash.
Re: (Score:2)
I am reminded by this great scene in Use of Weapons by Iain M. Banks where a character comes across some dude cleaning a table, a task that (in the Culture universe) could easily be done by technology. The response from the table wiper when it is suggested that wiping a table does not seem interesting or impactful:
âoeI could try composing wonderful musical works, or day-long entertainment epics, but what would that do? Give people pleasure? My wiping this table gives me pleasure. And people come to a c
Komidashi (Score:2)
best of eight...
recalculate komi...
best of five.
Se-dol wins.
profit!
To Be Fair (Score:1)
Re:To Be Fair (Score:5, Informative)
Nonsense. Go has 2 * 10^170 legal positions. That's a completely ridiculous number that's not solvable by brute force now, nor ever. Storing just one byte per position to indicate what move to play next (which is very generous given that a 19*19 board has 361 positions) would require 2 * 10^170 bytes of storage. A petabyte is 10^15 bytes for comparison.
Re:To Be Fair (Score:4, Interesting)
Nonsense. Go has 2 * 10^170 legal positions. That's a completely ridiculous number that's not solvable by brute force now, nor ever. Storing just one byte per position to indicate what move to play next (which is very generous given that a 19*19 board has 361 positions) would require 2 * 10^170 bytes of storage. A petabyte is 10^15 bytes for comparison.
For comparison, the number of atoms in the observable universe [universetoday.com] is on the order of 10^80 (give or take 2 orders of magnitude), so exhaustively searching Go is doable if you have a spare couple of universes to dedicate to the task. But then, it's Google we're talking about.
It's not AI. Go is solvable (Score:2)
No, it's not AI. AI is an optimization, that's it.
Go is a simple game, and despite it's widely branching move tree that makes it almost impossible to compute a good next move, it can be solved. It might take a ton of memory and a ton of CPU power, but eventually Go will be solved so the most advantageous move given any board state can be rapidly chosen.
(For a simpler example of solving a game, try tic-tac-toe, a game which is trivially solved in a few minutes and one where only the starting player can win o
Re: (Score:2)
As you say, perfect play in tic-tac-toe by both players will result in a draw, but if player one plays poorly, then the second player can easily create a forced win where player one cannot stop player two from winning on their next move.
I was going to paste an example of such a
Re:It's not AI. Go is solvable (Score:5, Insightful)
It really is amazing how many posters on this tech site don't have even a beginning of an idea of what they're talking about, and zero inclination to go find out.
Go specifically has far too many possibilities to solve the way you suggest, not now, and very possibly not ever in the future. Tic-tac-toe is a game that can be brute forced by a talented highschooler on a calculator. Checkers takes more resources but has been solved. Chess has not, not even close. Go is many, many orders of magnitude more difficult to brute force than chess.
Re: (Score:2)
First you say it's a simple game, because you can just search the tree, and then you contradict yourself by saying "AI just makes it quicker to traverse the tree by pruning branches that lead to suboptimal results."
What you're missing is that pruning branches is not done with a tree search, but by looking at the board, and deciding it's not a good move, simply based on the pattern of the pieces, and nothing else.
That is the hard problem.
Re: (Score:2)
No, it's not AI. AI is an optimization, that's it.
Nothing is AI because as soon as it works it's not AI any more.
That aside, why isn't it AI? Go was a problem required that required intelligence to solve. Now we have an artifice to solve it. Hence, artificial intelligence.
It's like calling a steam train an iron horse because it could do some of the jobs horses did much better. No one expected it to be equivalent to a horse in all respects.
Stupid, and a bad example to others (Score:2)
Re: (Score:2)
I have to believe it's different when you have been at the top of a profession and acknowledged worldwide as such. He believes that very notion has been rendered meaningless by the AI just running riot over all human opponents now, and frankly, I can understand where he's coming from. He was big fish in a moderate size pond, and someone has gone and filled the pond with sharks. Realistically, he was probably never going to be king of the hill again even without the threat of AI, just because he's past his p
Re: (Score:2)
Re: (Score:2)
Tell that to Vince Lombardi, who said "Winning is not everything, it is the only thing."
Baseball next? Nascar? (Score:2)
Machines have been able to beat humans at most all sport for a very long time. By this guys logic, no one should play any sport, because it would be pretty trivial to build a robot team that would totally dominate pretty much any physical sport that exists (who is going to be able to hit a 250mph fastball???).
The only reason it is not done is because rules do not allow it.
Re: (Score:2)
Garry Kasparov did not quit (Score:2)
Kasparov retired after he lost his World Champion title to another human, not after he was played by IBM and lost to Deep Blue.
The fact that in the mid-80's he could play a simul against 32 computers and win every game is not a proof of superiority of the human brain. It is a proof of the pathetic state of computing in the mid-80's.
After all, if computers do not outperform humans, what would be the point of building them?
And the whole point of man vs machine matches was only a sales pitch by IBM, capitalizi
Re: (Score:2)
You don't see... Usain Bold racing against a car.
Are you absolutely, 100% sure that they didn't already make this commercial?
I'm waiting for something else (Score:2)
I'm waiting for the world's best doer of household chores, whoever he or she is, to retire from competing, because robots can do it better.
Isn't it interesting that we can automate the work of some of the most intelligent and trained people on Earth, while we cannot have a robot that does household chores at least as well as a pre-teen.
Re: (Score:2)
I've got a robot that's a hell of a lot better at vacuuming than I am.
*Cannot* be defeated? (Score:2)
What happens when you pit AlphaGo against itself?
Is the game almost always a draw? Or does one side usually win?
If the latter, then it is a fallacy that it cannot be defeated, only that nobody has reliably figured out how.
Re:*Cannot* be defeated? (Score:4, Insightful)
AlphaGo Zero is trained by playing itself.
It's quite difficult for Go to end in a draw, unlike chess.
Re: (Score:2)
That's kinda what I thought... and since AlphaGo can defeat itself, and does so routinely, then it follows that it *IS* somehow possible to reliably defeat.
My point is that just because nobody has figured out how to do something is not, by itself, a reason to say it cannot be done.
And when you have an example of a computer regularly figuring out how to do it, there's even less of a reason to conclude such a thing.
Re: (Score:2)
Alpha Go can certainly be defeated. Go has far too many states to create an actually undefeatable player. Se-Dol apparently doesn't believe that a human can defeat it.
He's probably mostly right. Alpha Go achieves grandmaster play levels in four days of training. It's not programmed, it learns. If you find a weakness in its play, it can be trained to counter. Just like a person, except much, much faster.
That doesn't mean Go isn't worth playing anymore. I'm sure Se-Dol will keep playing. But he's been world
Re: (Score:2)
That's kinda what I thought... and since AlphaGo can defeat itself, and does so routinely, then it follows that it *IS* somehow possible to reliably defeat.
Automobiles regularly beat one another in races, too.
Re: (Score:2)
Tournament rules give a half-point to white to prevent draws. Currently it's around 6.5 points for white in exchange for going second.
Draws could still happen with basic rules, but tournaments completely eliminated them.
Re: (Score:2)
The players nevertheless know that they played a draw and that only the "artificial rule" made one the winner.
congratulations to the number 2 human go player. (Score:2)
Cat (Score:2)
"On behalf of the whole AlphaGo team at DeepMind, I'd like to congratulate Lee Se-dol for his legendary decade at the top of the game, and wish him the very best for the future," said Demis Hassabis, chief executive and co-founder of Deepmind. "During the AlphaGo matches, he demonstrated true warrior spirit and kept us on the edges of our seats to the very end."
Anthropologists point out the victors ascribe ferociousness to defeated states. This is the source of naming so many sports teams after Indian tribes, something that applies in a lot more places than North America.
That's silly (Score:2)
Lee has a point (Score:2)
Why bother if a computer is going to beat you every time?
Game designers long ago realized that and have kept game AI pretty limited. If they didnâ(TM)t, who would buy the game?
That being said, what was the point of making a computer good at Go? Has it advanced CS theory and practice? Does it have other actually helpful potential applications?
Re: (Score:2)
The AI in the past felt more like it was weak through being incomplete instead of by design, as it often fell for the same trick again and again and being unable to adapt. In case of RTS games, it was both limited and cheating at the same time, almost as if they needed to give a hidden boost.
Since then, I've seen at least two games give fine control over AI player's strength, allowing
Re: (Score:2)
The machines also have us beat on quantity. They can play so many more games than human players that you can expect practically all theoretical novelties to arise out of lines indicated by computer exploration from now on, in any game that has been "beaten". Certainly I find Leela chess more interesting than human chess, although it requires wading through twenty draws to find that one brilliant win (I let other people do that wading). I watch analysis of TCEC matches on a fairly regular basis. I watch abou
Usain Bolt Must Feel Stupid (Score:2)
All that running and he's still slower than a motorbike.
AlphaGo isn't even a good player - it's just a huge lookup table. Who cares about that? Give it a game of Advanced Third Reich and it won't know what to do with it.
the only way to win (Score:2)
is to stop playing.