back to article Felted! AI poker bot Libratus cleans out pros in grueling tournament, smugly trousers $1.8m

Machines have triumphed again. Libratus, a powerful computer program, has crushed its human opponents at a heads-up no-limit Texas hold’em poker tournament held at Rivers Casino in Pittsburgh, Pennsylvania, winning $1,776,250 over 120,000 hands. It’s a landmark achievement in AI game playing, said Tuomas Sandholm, co-creator …

  1. RIBrsiq

    Commence movement of goal posts in 3, 2, 1...!

  2. Aristotles slow and dimwitted horse Silver badge

    Nice results, but...

    Now put it in a chair at a World Series or Aplha 8 final table and see if you get the same result.

    1. emmanuel goldstein

      Re: Nice results, but...

      The article already made it clear that Libratus is only effective at heads-up poker.

  3. A Non e-mouse Silver badge


    All this AI quashing humans sounds very impressive, but can it brew a decent cuppa yet?

    1. Rich 11 Silver badge

      Re: But...

      Only after it becomes self-aware and decides whether we're worth brewing tea for.

      1. Michael H.F. Wilkinson Silver badge

        Re: But...

        Or it might present us with a cup filled with a liquid which is almost, but not quite, entirely unlike tea

        Sorry, couldn't resist. Mine's the one with the cassette tapes of the HHGTTG in the pocket

        1. Swiss Anton

          Re: But...

          Just wondering why anyone would downvote @Michael's post. I can only assume that there are some miffed AI's who pride themselves on their ability to make a really nice cuppa.

          Of course the two aren't mutually exclusive if the AIs have learnt their tea making skills from someone who isn't British. Yes I'm aware the Chinese have been brewing tea for millennia, but a really nice cuppa needs milk and at least 3 teaspoons of sugar. A Chinese cup of tea has neither.

  4. Aaiieeee

    "today's robo-players triumph in heads-up battles, not at a table"

    Until I got to this line near the end of the article I was under the impression it was 4v1. I am still impressed.. just slightly less so.

  5. jaduncan

    "Building a poker bot as good as Libratus is also a major task, as it requires a healthy sized supercomputer. Libratus probably won’t be playing anyone online anytime soon, as it costs too much to run."

    Paging Dr. Moore. Dr Moore to the white VOIP enabled computer.

    1. BarryUK

      Well, you can fire up an Amazon EC2 instance with 36 CPUs and 60Gb RAM for $1.60/hr. Not sure what sort of computing power is needed for this software, but it sounds potentially profitable to run this in the cloud.

      1. Brewster's Angle Grinder Silver badge

        "Not sure what sort of computing power is needed for this software"

        Do you know why the AIs will beat us? Because they actually pay attention to what they've just read:

        Over 20 days, four human poker players stared at multiple computer screens for ten hours a day....Libratus gobbled up approximately 19 million core hours of computing....

        1. asdf

          >19 million core hours of computing....

          Lets see 19 million hours * $1.6 (yes yes probably actually much more per hour but just for the sake of argument) to earn less than 2 million dollars sure looks like Cyanogen Inc (rot in hell Kirt McMaster) math to me.

          Edit: in retrospect not a great analogy because at least this computer had something tangible at the end instead of simply dog shit smeared on a defunct brand.

          1. JulieM Silver badge

            You don't make omelettes without breaking eggs

            The first attempt to do anything is always an enormous effort.

            A lot of the data gathered will turn out not to have any influence on the results, but you won't know this for certain until you have analysed all the data.

            On the next iteration, you need pay no <del>mind</del><ins>CPU cycles</ins> to anything that did not affect the results last time around. (If your success rate falls, you know that data was important after all ..... I suspect "Round Two" will last longer, as there are so many ways the decision tree can be pruned, and more of them will make things worse than better.)

            Once you have analysed many head-to-head games with different opponents, you can get a handle on what are person-to-person variables (different human players will invariably have their own preferred tree-pruning patterns, which will mean their behaviour will be predictable in given situations) and then you can build a machine capable of taking on several players at a table.

            I'd be properly impressed if a computer could work out for itself that it could start out playing defensively, losing a small amount, then turn the tables once its human opponents have started trying to take it for a sucker. Or decide why a bait-and-switch strategy might not be the best.

            Oh, yeah, and there are two "l"s in "gruelling".

      2. Anonymous Coward
        Anonymous Coward

        I'm working on a way to do this in live competition. If it works out I'll change my name and move to New Zealand - I might buy the North Island.

  6. Anonymous Coward
    Anonymous Coward

    Imperfect text

    "Poker is a difficult game for machines to master as it’s an imperfect information game. Players do not have equal knowledge about the game state due to hidden cards"

    "Poker is an imperfect information game, as players do not have equal knowledge of the game state due to hidden cards"

    Did you know that poker is an imperfect information game, and that players do not have equal knowledge? Due to hidden cards apparently.

    1. Andrew Moore

      Re: Imperfect text

      I heard that hidden cards make poker an imperfect information game.

  7. Anonymous Coward
    Anonymous Coward

    Can it decide it doesn't like poker and instead raise sheep in Wales?

    On the AI playing games, I'm looking forward to the star craft games. As that game as long as the AI is limited by the physical limitations of the human body (it needs to use a mouse pointer, hot keys, limited to how hands work, the hands of a progamer of course (you're looking at upto 600 apm in bursts)

  8. Magani


    Do AI poker players have a particular poker face screen?

    1. earl grey Silver badge

      Re: Visage

      Guess you'll just have to ask Lady Gaga....

    2. Ashley_Pomeroy

      Re: Visage

      On a tangent I remember seeing a video on Youtube a while back. It was of a poker player - a really good one - who was having an off day. I've forgotten the details but he was convinced that his hand was about to win, because only one combination of cards could beat it, and he was boasting about it.

      And inevitably that one combination of cards came up, and he had the most incredibly crestfallen facial expression. It was awesome. (googles)

      It's this one:

      He is apparently a top player but in this one match he let his poker face drop, and fate stomped on him.

  9. ChrisElvidge

    Card counting

    Doesn't a computer by definition count cards. And isn't this forbidden in Casinos?

    1. Anonymous Coward
      Anonymous Coward

      Re: Card counting

      Card counting restrictions apply to games where they aren't shuffling each time.

      In poker hands are dealt from a single, shuffled deck each time and you're playing against other players. Card counting is part of the game.

      1. BoldMan

        Re: Card counting

        They still have a better chance of calcualting the odds vs a human player.

        The usual mantra when playing poker is "Play the Player, not the hand" - Poker is a game of psychology so how can the humans stand a chance when their most important skill has been nullified by a computer screen?

        Its why online poker is a mugs game.

        1. Anonymous Coward
          Anonymous Coward

          Re: Card counting

          Exactly. All this demonstration shows is that computers are better at calculating odds. Real poker isn't about odds, that's just the playing field. It's about people scrutinizing other peoples' play and trying to out-guess them. When poker is reduced to pure statistics all the fun has been sucked out of it.

          Let the machines win their way. No machine will ever truly participate in what goes on around a table of humans holding cards, any more than they will ever 'get' comedy or tragedy.

          1. Version 1.0 Silver badge

            Re: "real poker"

            I seem to remember hearing all these arguments being trotted out to explain why humans would always beat computers at Chess, Go etc...

            1. Anonymous Coward
              Anonymous Coward

              Re: "real poker"

              I don't remember that. Chess and Go are pure strategy games with no element of chance, so they fall in a different class anyway. Poker between humans involves 'tells' that no program has (unless programmed to), so a purely programmatic game of Poker is a different game than when humans play it. Not that much point in comparing them.

        2. Destroy All Monsters Silver badge

          Re: Card counting

          "Poker is a game of psychology"

          Well, apparently not.

          "Giving a fuck about flesher psychology because I don't have to"

    2. a_yank_lurker Silver badge

      Re: Card counting

      Blackjack has odds slightly against the house if one counts cards correctly. Normally poker is not played against the house but other players.

      1. Anonymous Coward
        Anonymous Coward

        Re: Card counting

        That's only true in Blackjack when a deck is played all the way thru without shuffling. Thus many casinos require shuffling before all deals, while others advertise that they don't, so the wanna-be card-counters can give it a try. Probably the stakes are kept low in those games and it's treated as a loss-leader.

  10. Simon Rockman

    There has been some interesting AI research done by Snowie which produces AI based poker training

    They do "poker for business" and other interesting things, but I suspect Snowie already has bots playing online.


  11. breakfast

    But can Libratus read Lady Gaga's poker face?

    1. User McUser

      How do you wake a sleeping Lady Gaga?

      P-P-P-Poke Her Face.

  12. Ashley_Pomeroy

    Colossus: The Forbin Project

    Imagine if Libratus was allowed to keep its winnings. It would plough the money back into improving its hardware - after a few tournaments it would be ordering the security guards to execute the other players, and then it will offer us a choice between "the peace of plenty and content or the peace of unburied death"

    We'll try and fight it, but it will just toy with us.

    1. John H Woods Silver badge

      Re: Colossus: The Forbin Project

      "Imagine if Libratus was allowed to keep its winnings" --- Ashley_Pomeroy

      It would pay its debts: as noted above 19M core hours cost more than the winnings.

    2. allthecoolshortnamesweretaken

      Re: Colossus: The Forbin Project

      Alternative title: IoT With Nukes

  13. montyburns56

    What happens in Vegas...

    Yeah, but unfortunately a week later Libratus had spent all of it's winnings on hookers and cocaine and ended up married to an Atari ST with a damaged MIDI port.

  14. Sooty

    not sure about this

    Poker is a game of 2 halves, one half is playing the odds, which the AI can do very well, a lot better than any human.

    Also there is playing the other players, intimidation, reactions etc. it's why online poker is so different to in person poker. I can't imagine the AI isn't going to take into account how the opponent places chips into the pot etc. It might analyse length of time, but it's not going to get clues about body language etc.

    Yes you can consistently win money by just playing the odds, but it's called grinding for a reason.

    Not only that, heads up is very different as well...

    1. JulieM Silver badge

      Re: not sure about this

      A computer is going to be able to make an accurate estimation of probabilities in every situation, untarnished by emotion. A human is going to be forever thinking about the consequences of winning or losing, and occasionally lose sight of the relative probabilities. A computer might not be able literally to see the pound signs in your eyes, but it can certainly detect an ever-so-slightly-over-eager mouse click -- while deliberately timing its own actions to indicate nothing of its own hand. It will not show hesitation when bluffing, nor urgency when betting on a good hand. It isn't influenced by thoughts of lying naked on a beach sipping cocktails, nor searching in a garbage skip behind a supermarket for something edible.

      Imperfect shuffling of the cards could even create a pattern which an AI might be able to pick up on, giving it a slight advantage. (If it's even allowed to remember past cards; making use of this information might be considered unsporting behaviour. Which is not unreasonable, seeing as the computer is most assuredly not in it for the sport, but to win as much as possible.)

      I think only an exceptionally talented actor -- or the right combination of hyper-aggressive play and sheer blind luck -- could hope to beat that.

  15. Simon Ward

    This isn't AI ...

    When Libratus can engage me in a conversation on, say, the price of cheese in Paraguay or deduce the existence of income tax and a recipe for rice pudding before it is switched on then it isn't an AI.

    An expert system which happens to be good at poker? Definitely.

    An AI? Definitely not.

  16. JimM

    Cart before horse?

    Surely one of the features of GAI is that it can learn how to deal with uncertainty and doesn't need a human to teach it.

  17. Destroy All Monsters Silver badge

    Rise of the Cardshark

    According to Top 500: Bridges Supercomputer Used to Build AI Model for Beating Humans at Poker , Libratus is running on the Bridges Supercomputer, which is a fat piece of kit:

    Bridges is a new kind of supercomputer being built at the Pittsburgh Supercomputing Center (PSC) to empower new research communities, bring desktop convenience to supercomputing, expand campus access, and help researchers facing challenges in Big Data to work more intuitively. Funded by a $9.65M NSF award, Bridges consists of tiered, large-shared-memory resources with nodes having 12TB, 3TB, and 128GB each, dedicated nodes for database, web, and data transfer, high-performance shared and distributed data storage, the Spark/Hadoop ecosystem, and powerful new CPUs and GPUs.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Biting the hand that feeds IT © 1998–2020