Garry Kasparov: Open-Ended AI [video] | Hacker News Reader

Garry Kasparov: Open-Ended AI [video]

youtube.com

33 pointsAlanTuring21 comments

ITT: People who refuse to watch the video because kasparov can't possibly know anything else than chess. Being a great chess player doesn't make you an expert in everything, but it also doesn't mean you can't say anything useful about anything else than chess.

I think he has a valid point, then systems are specifically engineered to work in a domain, for instance chess. The rules and aspects of the game are programmed and then it can teach itself how to optimally play by those rules. It is a very limited scope and VERY fa from general AI

This limited scope is still progress though, toward the general AI he is saying will not come in the near future. The Alpha Zero project showed that it could be adapted for any 2-player perfect information game. This is still quite limited, but not nearly as much as a program like Stockfish is, which will basically only ever play chess.

Perhaps the next step is bringing the system further into the realm of the unknown in games, like chance and imperfect information. I don't know if Google is continuing this, but it would be a shame if it wasn't.

They showed that the system could be adapted to learn any two player perfect info game.

Once it has learned Go it won't play chess. Also, it has to be manually adapted.

By this logic minimax has also been shown to do the same:

https://en.m.wikipedia.org/wiki/Minimax

Minimax is a general rule on how to optimally play 2 player games. It's about as much of a learning algorithm, and as practical, as "buy low, sell high".

In pure form, it can't be applied to any game where you cannot enumerate every last possible board position (so Chess and Go are out). And with heuristics, those heuristics effectively become the algorithm and they're really complex.

The minimax principle is used in chess and go and any other competitive two player perfect knowledge game. What improves is the selection of moves to evaluate and the quality of the board evaluation. Chess and Go absolutely use minimax with sleight, but important, modifications. Deep learning extensions with Alpha Go were specifically for move selection and board evaluation.

Alpha beta pruning with heuristics is one of the most fundamental extensions of minimax. It is still minimax.

The alpha go system still clearly applies only to perfect information 2 player games, because it is fundamentally extended minimax.

Edit: I'm not saying the heuristic extensions of alpha go, or the reinforcement learning of zero, aren't brilliant and important. That doesn't take away from the basic fact that it is an algorithm for perfect info 2 player games and one of them at a time (even if modifications can adapt it to other games). Heuristics will be fundamental to general intelligence, but alpha go is not general intelligence.

A lot of comments dismiss Kasparov's thoughts because he is out of his element(being one of the greatest chess players of all time, but apparently his brain is too small to understand AI). Can anyone explain why he can't be expert on AI? Is it because he doesn't have a degree in CS?

He can be an expert on AI, of course, but this assumption requires proof. He's expert on chess and he proved it many times, but I'm not aware of anything related to AI. That said, he's smart man, of course, and it's interesting to listen to him anyway. I just have much more trust in people who actually work in this industry.

It's because what he says doesn't make sense. And when reality catches up he moves the target.

Take for example his conversation with Sam Harris[1] where Garry says that Human+Computer will always win over Computer only at chess. If you know about CS, if you've heard about AlphaZero, you can see that even if it might currently be that Human+Computer > Computer it surely won't be for long. It'll come a time that Human will be Computer's handicap, and we might be there already.

Now he moved on to speak about the narrow/general problem, but that doesn't make him any more of an expert. If you want to listen to an expert look for someone qualified, someone like Eliezer Yudkowsky.

[1] https://samharris.org/podcasts/the-putin-question/

Kasparov apparently continues to believe that his expertise in chess transfers to other disciplines. This isn't an uncommon phenomenon, even for people with more serious qualifications in a not totally unrelated field. I also shouldn't forget Hollywood actors - their views on scientific matters and foreign policy carry a lot of weight because they're, ummmm, famous.

For the same reason I'm not rushing over to Amazon to snag Garry's books on how to succeed in the boardroom, I won't be spending any time trying to digest his musings on AI.

Many many people believe their expertise in one field transfers to other disciplines. Chess expertise is not unique. Even though I generally agree about this particular case, I am aware that I am not an expert on Kasparov or AI. I am also aware that Kasparov is an expert on Kasparov and has practical professional experience interacting with AI's...I mean IBM built AI's specifically tuned for him and Kasparov has had decades to reflect on that experience. He didn't just fall off the turnip truck in Artificial Intelligence land.

Neither Deep Blue nor the superior chess playing engines running on PC hardware that came afterwards used machine learning. They relied on basic heuristics as humans do to numerically evaluate a position (material, activity, etc.) with the advantage that they could evaluate many more possible move sequences than a human. They also had opening books to avoid losing games from the outset and endgame tablebases to identify forced wins/draws with a small number of pieces on the board. Only recently did Google come out with a chess playing program that is actually ML-based, and it beat the top rated 'traditional' chess engine.

My point being, despite chess being considered a game that requires deep thinking, the use of actual AI in chess is very very new. As far as I'm aware, Kasparov had nothing to do with it let alone a deep understanding of it. He wasn't even involved in the earlier development of computer chess playing programs as they rose to the grandmaster level and eventually beyond his own level (super grandmaster). He along with many others had confidently predicted that machines would never beat humans in chess. So yes, I'm quite reluctant to believe that he has any kind of vision on this topic.

the use of actual AI in chess is very very new.

  Each generation thinks it invented sex.
  -- Heinlein

>Kasparov apparently continues to believe that his expertise in chess transfers to other disciplines.

There is a lot of irony in this statement, since he speaks about expertise of AI systems in Chess and other games transferring to more open domains.

I think there is some transfer learning from chess to politics at least. Kasparov is still alive after the 2010 "Putin Must Go" campaign.

> expertise in chess transfers to other disciplines

Some scientists think understanding quantum theories make them sudden experts in economics. I am always amazed at the arrogance displayed where humility should be the norm.

I see this phenomenon on HN and Reddit quite a bit. Some expert will publish a paper and there is inevitably a comment of the form, “The conclusion could be a result of...therefore the paper is trash.” People act as though their initial ruminations on a topic qualify them to properly critique an expert's research.

I try to keep in mind that if I could think of a possible objection or insight with a few minutes of thought then the expert is surely aware of this too. I think it’s common for someone who has acquired expertise on one area to think their insights apply to other unrelated areas. For instance Paul Graham has famously bashed philosophers but he certain doesn’t know what he is talking about.

These topics are discussed in greater depth in Kasparov's book, "Deep Thinking".

My view of the book: Very well written (and occasionally engrossing), well thought out, and strikes a more hopeful/reasonable tone that anything else I've seen on the subject.

All learning is reward-based. We can imitate human reward systems to start. The machines will act like humans do: basically random behaviors at the start, then following pleasure and rewards to shape their behavior.

what is the reward for learning for the joy of learning ?

hmm.. joy?

I think it's really funny that the video is Kasparov talking about AI working in closed systems having difficulty adapting any knowledge from a closed system to an open system.

While Kasparov is arguably the best player of all time and has a considerable intellect and vast amounts of chess knowledge, I think the the knowledge transfer from skill in chess to knowledge about other difficult, technical fields is quite difficult. I'm in no rush to listen to his thoughts on transfer learning in AI (although I did watch the video to confirm my suspicion).

Reply on news.ycombinator.com