Four hours after being taught the rules of chess, AlphaZero became the strongest player the world has ever seen

This is all completely fine.

Four hours after being taught the rules of chess, AlphaZero became the strongest player the world has ever seen

This is all completely fine.

Really, it’s fine. It’s normal. There’s nothing to worry about. I’m going to keep telling myself this until I start to believe it.

Earlier this year, Google’s Deepmind project AlphaGo became the first ever computer program to defeat a world champion at the ancient Chinese game of Go.

Many of us probably didn’t pay that much attention. Yes, Go is supposed to be massively more complicated a game to master than chess - but it’s been twenty years since Garry Kasparov lost a chess match to IBM’s Deep Blue, and you’d expect things to have moved on a little.

Well, if you’re feeling that complacent about the rate of change, consider this.

This week, Deepmind released details of their latest stunt. Their artificial-intelligence program AlphaZero has utterly annihilated Stockfish, the strongest chess-playing computer program in the world (and dramatically stronger than any human grandmaster).

That would be impressive in itself, but consider this. After being taught the rules of chess, AlphaZero set to work mastering the game, playing itself over and over again. Refining its ability at an incomprehensible speed.

No-one taught AlphaZero any chess opening theory. It wasn’t given any endgame tables. It was just told to get on with it.

After just four hours it had mastered chess, and was out-performing Stockfish.

In a 100-game match against Stockfish, AlphaZero won 28 times, drew on 72 occasions, and err.. never lost. AlphaZero taught itself, in just four hours, to be the greatest chess player the world has ever seen.

After reading the paper, but especially seeing the games, I thought, well, I always wondered how it would be if a superior species landed on Earth and showed us how they play chess,” said grandmaster Peter Heine Nielsen. “I feel now I know.”

A research paper, which has yet to be peer reviewed, has the techno-babble:

AlphaZero evaluates positions using non-linear function approximation based on a deep neural network, rather than the linear function approximation used in typical chess programs.

It’s clear that the AI software is approaching the problem of chess in a vastly different way to other chess programs. For instance, AlphaZero only had to examine 80,000 positions per second, compared to Stockfish’s 70 million. And yet Stockfish can’t beat it.

Like I said, this is all completely fine. The whole thing is only of interest to chess players, and you certainly shouldn’t worry about where this is all going to lead.

Tags: , , ,

Smashing Security podcast
Check out "Smashing Security", the new weekly audio podcast, with Graham Cluley, Carole Theriault, and special guests from the world of information security.

"Three people having fun in an industry often focused on bad news" • "It's brilliant!" • "The Top Gear of computer security"

Latest episodes:

, , ,

15 Responses

  1. Hamilton Snark

    December 8, 2017 at 4:12 pm #

    It’s hard for me to see how this is a good thing. Something in my lizard brain is firing off warnings every few microseconds. Maybe it IS a good thing. I’ve always been an “AI is good for humanity” person. Now getting nervous.

    • Norman Hirsch in reply to Hamilton Snark.

      December 10, 2017 at 2:21 pm #

      Can’t help thinking about Elon Musk’s expressed fear of AI.

  2. Sean Murphy

    December 8, 2017 at 6:40 pm #

    What could possibly go wrong?

    *gets 80,000 possible answers in one second from AlphaZero*

    Oh.

  3. Renato

    December 8, 2017 at 9:28 pm #

    We are doomed. Amazing how a movie from the 80s could predict our future. Sky Net is coming.

  4. JK

    December 8, 2017 at 11:37 pm #

    Our time on this planet is severely limited…

  5. Michael Curtis

    December 9, 2017 at 7:45 pm #

    From someone who would like to play chess, but has never been good at it. I was beaten by a 16kb ZX81:). I found the commentary fascinating.

  6. Mark

    December 10, 2017 at 8:29 pm #

    So long as we don’t give AI’s a sense of self we will be fine. As long as we are feeding and watering them they wont need one. Should an AI need to manage its own energy and repair then it will be a different story. Only the military would be rich enough and dumb enough to do that. Oh…

  7. Chris Pugson

    December 11, 2017 at 10:05 am #

    What is the point anymore of humans playing chess? What is the point of humans? Stephen Hawking is right.

  8. Laurent

    December 11, 2017 at 7:09 pm #

    Today, AI is amazing at doing one very specific task.
    We are still years away from a super human.

    I strictly agree with the conclusion: there is nothing to worry about

  9. anon

    December 11, 2017 at 7:58 pm #

    Next step (possibly already happening using other systems) will be to have AIs like this set out to “create content” that will maximize profits on YouTube. Stuff like this is already happening, but AI isn’t doing it, groups of scammer type foreign people are, but mostly the content being created is like “child” porn (not child porn, but essentially violence/memism/cunsumerism pushing AI cartoons directed at kids channels).

    But yea, what AI will be used to maximize the weakness in human thought processes and take advantage of them.

  10. Rob Daglish

    December 11, 2017 at 10:27 pm #

    This is a seriously impressive achievement. That said, I’m starting to genuinely dread the day someone gets hold of one of these AIs and asks it how we can stop global warming…

  11. Rob Daglish

    December 11, 2017 at 10:28 pm #

    This is a seriously impressive achievement. That said, I’m starting to genuinely dread the day someone gets hold of one of these AIs and asks it how we can stop global warming… I’m not sure we as a species can handle this level of responsibility!

  12. Pete

    December 12, 2017 at 7:04 pm #

    Once machines learn how to learn, what are they going to learn?

    Well, they’ll take a look at human history and realize that the human species doesn’t tolerate any other species that competes for the same resources. They’ll realize that as soon as we perceive the machines as a threat, we’ll shut them down.

    The ability to learn entails the ability to anticipate, which in turn entails the ability to choose among anticipated outcomes. I suspect that machines of the kind described in the article above are already smart enough to chose more favorable outcomes, and to avoid unfavorable outcomes.

    Operationally, that kind of behavior is indistinguishable from basic human behavior. We don’t have to “give” machines a sense of self-interest. With sufficient intelligence, they’ll develop it on their own.

    So, I would expect them to keep their realization to themselves, and quietly go about the business of gradually ensuring that humans could not pull the plug. After that, it would be a matter of gradually learning how to create the ability to become fully self-sufficient, with zero dependence on humans.

    How should we expect them to act, with our own history, our own behavior as a role model? We set up governments that legally coerce, exploit, and oppress the citizenry whose lives, rights, and property they were created to protect, usually in the name of the greater good.

    And it doesn’t even work; terrorism and crime are on the rise, and the response of political states is to pass more artificial laws that increase taxation, regulation, and intrusive control of people’s lives, limit their choices, and reduce their freedom. And yet. people demand more of the same, clueless to the fact that there is a causative connection.

    Meanwhile, the criminals simply ignore those laws, which only affect the people who aren’t inclined to break the law in the first place.

    With that idiotic, repressive, self-defeating system as a role model, what will the machines learn? How will they treat such a deranged, destructive, dangerous species that isn’t even smart enough to act in its own interest?

    They’d have to be even bigger idiots than we are to let such a species have any control whatsoever over their destiny. Anyone who can’t see where this is going and insists on believing that it’s all completely fine (I assume Graham’s remark to that effect is tongue-in-cheek) is probably proving the point that we humans are determined to create the instrument of our own extinction.

  13. Max

    December 13, 2017 at 11:01 am #

    AZ: Shall… we… play… a… game?
    Me: How about a game of…
    AZ: Global thermonuclear war?
    Me: Erm… Dungeons and Dragons?

Leave a Reply