The AI That Has Nothing to Learn From Humans (theatlantic.com)

An anonymous reader shares a report: Now that AlphaGo's arguably got nothing left to learn from humans -- now that its continued progress takes the form of endless training games against itself -- what do its tactics look like, in the eyes of experienced human players? We might have some early glimpses into an answer. AlphaGo Zero's latest games haven't been disclosed yet. But several months ago, the company publicly released 55 games that an older version of AlphaGo played against itself. (Note that this is the incarnation of AlphaGo that had already made quick work of the world's champions.) DeepMind called its offering a "special gift to fans of Go around the world." Since May, experts have been painstakingly analyzing the 55 machine-versus-machine games. And their descriptions of AlphaGo's moves often seem to keep circling back to the same several words: Amazing. Strange. Alien. "They're how I imagine games from far in the future," Shi Yue, a top Go player from China, has told the press. A Go enthusiast named Jonathan Hop who's been reviewing the games on YouTube calls the AlphaGo-versus-AlphaGo face-offs "Go from an alternate dimension." From all accounts, one gets the sense that an alien civilization has dropped a cryptic guidebook in our midst: a manual that's brilliant -- or at least, the parts of it we can understand. Will Lockhart, a physics grad student and avid Go player who codirected The Surrounding Game (a documentary about the pastime's history and devotees) tried to describe the difference between watching AlphaGo's games against top human players, on the one hand, and its self-paired games, on the other. According to Will, AlphaGo's moves against Ke Jie made it seem to be "inevitably marching toward victory," while Ke seemed to be "punching a brick wall." Any time the Chinese player had perhaps found a way forward, said Lockhart, "10 moves later AlphaGo had resolved it in such a simple way, and it was like, 'Poof, well that didn't lead anywhere!'" By contrast, AlphaGo's self-paired games might have seemed more frenetic. More complex. Lockhart compares them to "people sword-fighting on a tightrope."

Comments

For all IT knows...

[richrz]

stones are REAL PEOPLE. Be afraid.

Re:For all IT knows...

[Applehu Akbar]

No, that's not true. It has no concept of 'people', 'person', 'human' or 'alive'.

So it would make the ideal airline customer service bot, then.

AI will be alien

[Anonymous Coward]

I think this teaches us a great deal about what AI will actually be like when it inevitably arrives. It won't be r2d2 or c3p0 or data - it will be an alien mind that will be incomprehensible to the rest of us.

Re:AI will be alien

[fluffernutter]

If it can't do anything but see patterns of black and white dots, I think we'll be ok.

Re:AI will be alien

[mikael]

But it will see every human as a black or white dot.

Re:AI will be alien

[fluffernutter]

Now you're calling aliens racist.

Re:AI will be alien

[Jeremi]

I think this teaches us a great deal about what AI will actually be like when it inevitably arrives. It won't be r2d2 or c3p0 or data - it will be an alien mind that will be incomprehensible to the rest of us.

Except, of course, for the AIs that are trained to emulate human thought processes -- those will be comprehensible to us (or at least, we'll be able to pretend that they are ;))

Re:AI will be alien

[Maelwryth]

I was thinking along those lines the other day. There could be a true AI in the Internet right now but what would it see? What inputs and outputs would matter to it? Would we even be in its perceptions?

Remember when Go was unsolveable?

[jandrese]

It was only a few years ago people were saying that the best Go computers would never beat human players because the game was so much more complex. We're getting to the point where AI decisions, even when explained, end up being too complex for humans to follow. This is a scary path we are following.

Re:Remember when Go was unsolveable?

[MrDozR]

Ummm, Go still isn't solvable. Not in the mathematical sense; there are just way too many moves to determine who should win from any given position.

Re:Remember when Go was unsolveable?

[Dixie_Flatline]

I find it less scary than inevitable.

I think if Go playing programs were constrained to the world-view that humans have, somehow, maybe they wouldn't be able to outplay humans, but of course, they're not. They can look for solutions that we didn't even consider. Humans are really good at thinking we have all the answers, but we're obnoxiously bad at even understanding the SCOPE of the problem, let alone solving the problems themselves.

"The planet's other lifeforms reveal so many ways of being that we could never imagine them if they didn't already exist in reality. In this sense, other species don't only have the capacity to inspire our imaginations, they are a form of imagination. They are the genius of life arrayed against an always uncertain future, and to allow that brilliance to wane out of negligence is to passively embrace the death of our own minds." (JB Mackinnon, "The Once and Future World")

It's not new that we're bad at it, we've always been bad at it. But at the very least, we've been very narrowly successful in creating (and observing) things that can break past our own lack of vision.

So don't look at this as a scary time, this is just us finally building the tools so that we can hope to comprehend the universe around us. This is no scarier than the advent of the telescope or microscope.

Re:Remember when Go was unsolveable?

[JoshuaZ]

Yes, and this is happening faster and faster with more tasks. Unfortunately, once a task is well done by computers, we cease to think of it as impressive. Thus, it was a big deal when computers beat the best humans at chess, and now you can literally get an app that beats grandmasters at chess on your phone, and it just fades into the background. Even scarier, Go is (depending on your version of the ko rule you use) either EXPTIME complete or EXPSPACE complete https://en.wikipedia.org/wiki/EXPSPACE, https://en.wikipedia.org/wiki/EXPTIME, so from a computational complexity standpoint, this is one of the hardest types of problems humans ever face. If this isn't strong evidence that we should be worried about the capability of AI, I'm not sure what would be.

Re:Remember when Go was unsolveable?

[philmarcracken]

A closed problem has definable limits. For example, I struggle to see how an AI system would accurately render assistance as a psychologist; a field in which I feel we barely understand ourselves.

Then there is the simple bodies we can give them. A nurses work is varied but still well defined and considered closed. But do we have a body for the AI system to do that job and it's many tasks, as quickly and cheaply as a human employee?

Re:Remember when Go was unsolveable?

[fluffernutter]

If you can't write all the rules in a 20 page booklet, it is an open problem.

Re:Remember when Go was unsolveable?

[Wycliffe]

So, how do you decide what is a "closed" problem and what is an open problem? More to the point, I suspect that for whatever definition you are using of "open ended" by the time an AI can beat humans at a bunch of them, it may be too late.

Not only is Go a closed problem, it is a well defined problem with a well defined solution and well defined scoring. Even something like starcraft has well defined incremental scoring and a well defined goal at the end. The biggest limitation to AI right now is open ended problems. Even if solving Go is very complex, Go is a very simple game where the board can easily be represented by a tiny two dimensional array. How do you make a digital representation of an unfolded pile of laundry or the stuff in a kitchen needed to bake a cake? One of the main advantages of a digital representation is that an AI can use trial/error and solve it a million times. Even if you gave a computer the dexterity needed to bake a cake and came up with some form of scoring to determine if the cake was "good", a computer can't do one million trials in the real world.

Not all that surprising

[foxalopex]

AI hasn't really changed much in years, computers have gotten more powerful so some pattern recognition tricks that use to take more effort don't anymore. The big breakthrough was figuring out how to write a decent AI to play Go optimally against modern equipment. After that all they did was create random legal scenarios to play against the system creating patterns that are unusual to us and using the results of those outcomes to fine tune the AI algorithm. There really isn't that much magic if you've taken some AI courses.

Think about fractals, some of the resulting art or images generated from some relatively simple equations create things which we could have never imagined.

Re:Not all that surprising

[Gorobei]

You sound like an RNN trained quickly on a small dataset of popular science articles.

Re:Not all that surprising

[raftpeople]

There have been significant advancements in ANN methods, one specific area is deep belief networks and the training algorithm (Hinton). Prior to that, it was known that multi-layer networks could out perform simpler networks for things like image recognition, but there wasn't a good way to train them.

The newer models/methods are outperforming the previous simpler models/methods.

An interesting thought

[alvinrod]

Here's an interesting thought. Would these people still say the same thing about the games if they were told there were AI games, but in reality were actually games by two human players?

It reminds me of the recent story where some kids put a pineapple in an art exhibition as a joke and people thought it was art. Most people will believe and/or spew pure bullshit if they think it's what's expected of them.

Re: An interesting thought

[CustomBuild]

It reminds me of the recent story where some kids put a pineapple in an art exhibition as a joke and people thought it was art. Most people will believe and/or spew pure bullshit if they think it's what's expected of them. Are you trying to be ironic?

AlphaGo Zero self-play games

[muninn]

They have actually released a trove of game records from the newest version, with two actually commented by a professional:
http://www.alphago-games.com/

Not quite

[nospam007]

"Lockhart compares them to "people sword-fighting on a tightrope."

If you see that you lose, you can always cut the rope.

Re:Fascinating. Sounds like the AI Circuit Design.

[mikael]

https://www.damninteresting.co...

They wanted a circuit that detected between two frequencies. The system was supposed to be digital, but the artificial evolution had made use of analogue computing using magnetic fields and harmonics.