Let's say AlphaGo beats Sedol - maybe now, maybe in 1-2 years, but Go will be "done" soon. What is the next challenge for AI? Is there any zero-sum-game left? Where is AI still really bad? I just tried a bot called "cleverbot" who's supposed to talk to you but it's a sorry try. Does someone have an overview here?

From my point of view the law business is still the most distance AI has to go...agreed?

I heard that the DeepMind team is going to tackle StarCraft next.

_________________
be immersed

If you are a man you would say the biggest challenge is to understand women. If you are a women you would say the biggest challenge is to understand men. When that is solved then there is nothing left.

_________________
Still officially AGA 5d but I play so irregularly these days that I am probably only 3d or 4d over the board (but hopefully still 5d in terms of knowledge, theory and the ability to contribute).

I think the financial industry would be a good candidate.
Big data, data mining and cloud are already very hot in the financial industry.
Combination Google/funds/AlphaGo/finance is already almost unbeatable if they pursue in this area, I think.
Sorry to say Pippen, but IMHO that's more interesting (and promising!) than law business.

I think zero-sum games are "done". I think games - relying heavily on luck like coin toss - will never be "done" for AI for natural reasons. But speaking and talking is the real deal, because when you look at a decision tree from Go and from when we talk it's another ball game, it even worse than Chess vs. Go. And in fact up to now AI can only speak naturally like a 5 year old. Or am I mistaken?

Maybe not an overview, but there is something to discuss about what you just sadi.

There are currently a number of AI programs that come close to passing the "Turing test" (can you tell if you are talking to a human or a computer). That was true even a decade ago if the realm of discourse was restricted (limited to some subject) but I understand that there are programs today that are pretty good in general conversation BUT we are talking about "when running on VERY powerful machines".

This "cleverbot", what was it running on? How "normal" did the "person" have to be. That latter is important. Even decades ago the program "Parry" might have been able to fool you even though running on a modest machine, but the "person" Parry rather sick so you expect some weirdness in his responses. Just as you expect a drunk not to always be following you.

BTW -- over time the "Turing test" has become more refined in terms of goals. Being able to fool you for a short time or usually fool you no longer the goal. Rather, to be able to fool you in situations where a program could be expected to give itself away by "non sequitor". Some examples of these harder tasks:
1) If told a joke, respond with a joke that is related in some sense (recognize a joke, respond not just with a random joke but with one that has some relationship to the first).
2) Same with a story.
3) Make sense of elliptical speech, a conversation where much is being left out because assumed by general knowledge of the situation and relevance. For example .....
"Honey, I'm home"
"The stove just died."
"There's a special at Dinardos"
"I'll get changed."
<< As a human, you can recognize that is not a "non-sequitor" conversation -- but can the computer program? Understand the connections between those statements? >>

Emanuel Lasker, the famous chess world champion, once said that his success was in part due to his ability to "play the man", not the board. Go and chess are games of perfect information (not zero-sum games). Humans playing these games can notice personality traits of their opponents and can take advantage. One assumes that DeepGo plays the board and, being a machine, it does not have personality traits to be exploited. Lee Sedol is a fearsome opponent for a human player partly because of his reputation. Computer programs have not "mastered" poker because winning at poker is heavily influenced by ability to "read" the other players. It might be possible for a bneural net based method to play poker. Instead of feeding it a lot of patterns of game play as was done with go, the computer will have to watch people playing poker and somehow correlate game play with what is observed in the behavior of the players. Seems difficult.

They're working on poker: https://www.theguardian.com/science/2015/jan/08/poker-program-cepheus-unbeatable

Right now it's a very limited subset of poker (limit heads up Texas hold'em), but the research is promising. One interesting thing from related articles is that even though the bot will always win in the long run, it doesn't yet exploit human psychological weaknesses to gain an advantage against weaker players. That might be necessary for any long term success in more open forms of poker, and that's a much harder challenge.

And work out that this conversation probably took place in the USA and not in England and respond appropriately.

From the Verge interview with Demis Hassabis

I think for perfect information games, Go is the pinnacle. Certainly there are still other top Go players to play. There are other games — no-limit poker is very difficult, multiplayer has its challenges because it’s an imperfect information game. And then there are obviously all sorts of video games that humans play way better than computers, like StarCraft is another big game in Korea as well. Strategy games require a high level of strategic capability in an imperfect information world — "partially observed," it’s called. The thing about Go is obviously you can see everything on the board, so that makes it slightly easier for computers.


So IMHO no game is safe from AI right now, and soon, the real world too.

Really anything requiring interacting with the real world is hard. There has been a lot of progress in machine vision and locomotion, but still, I don't think a robot could do my laundry within the constraints that my home imposes, and I honestly wish one could. I think I'll either die or switch to spray-on clothing before that becomes a reality.

I don't consider Poker a big deal for AI and in fact I read that Poker bots did already beat pro's. Poker is all about the expectation value of winning the pot. It depends on the odds of your hands and how much money is involved. Pretty easy to calculate for a machine. Then sometimes you can bluff, but I do not consider that vital if you play many many hands.

Poker bots are (AFAIK) nowhere near beating pros. There has been some success in limit hold'em heads-up. Poker is way more than just odds and EV, and it seems that you don't really know the game.

_________________
Offending ad removed

Can u try to explain what Poker is all about? I don't really know the game. For me it seems to consist only of

1. calculating the right odds & ev's
2. being cautious that 1. is not too obvious to your opponent
3. plain luck.

p.s. I just checked state space complexity: No limit Holdem: 10^140, Go 10^170, so Go should be still more difficult for AI than Holdem. Is there any more complex game left than Go?

"State space" in No Limit Hold'em seems to not be well defined just by saying "No Limit Hold'em". No limit means you can bet all you have. So in theory, you have infinitely many possible bet sizes at each step, if you are allowed to make real number sized bet. In practice, you are playing with chips and are not able to bet fractions of chips, so only finitely many bet sizes are possible. Now it ovioulsy depends for the number of those bet sizes, how big the stacks of players are compared to the smallest unit in chips(which we call 1 for now for siplicity, though it cound well be called 1000 in a tournament, to make it sound more fancy, or because it incresed during tournament). This also shows that just counting the "state space" doesnt really reflect difficulty well, as it is not changing actually much. Consider the following: If the big blind is 10, someting like a little less than 30 is a reasonable opening raise(with 20 being the minimum). Now you can also make it 25 or 24, if you want. If the big blind is 10000 instead (again compared to the smallest unit, in pratice the bb would never be so much bigger then the smallest unit, it makes no real sense) , you are able to bet weird things like 24564 instead of 25000 or 24000. There are a lot of added possibilties, while the are "almost the same" as the ones you had before, thus not making the game really much more complex.

What makes poker much more difficult, is the fact that it is a game of incomplete information, and therefore fundamentally different from chess and go. As the learning/NN approach doesnt seem to be tied to that sort of game, it seems reasonable to try wheter the problem are gone if you just try to apply them there.

OTOH computers have a bit of an advantage, as they are on home ground with mixed strategies, probalities, and so on(this is basicaly a calculation, humans will always round off a bit there), and don't have to worry so much about giving away tells(if they play reasonable mixed strategies, a strategy like "bet exactly this amount in exactly this situation" would of couse give away tells)

It's coming