Need to let loose a primal scream without collecting footnotes first? Have a sneer percolating in your system but not enough time/energy to make a whole post about it? Go forth and be mid: Welcome to the Stubsack, your first port of call for learning fresh Awful youā€™ll near-instantly regret.

Any awful.systems sub may be subsneered in this subthread, techtakes or no.

If your sneer seems higher quality than you thought, feel free to cutā€™nā€™paste it into its own post ā€” thereā€™s no quota for posting and the bar really isnā€™t that high.

The post Xitter web has spawned soo many ā€œesotericā€ right wing freaks, but thereā€™s no appropriate sneer-space for them. Iā€™m talking redscare-ish, reality challenged ā€œculture criticsā€ who write about everything but understand nothing. Iā€™m talking about reply-guys who make the same 6 tweets about the same 3 subjects. Theyā€™re inescapable at this point, yet I donā€™t see them mocked (as much as they should be)

Like, there was one dude a while back who insisted that women couldnā€™t be surgeons because they didnā€™t believe in the moon or in stars? I think each and every one of these guys is uniquely fucked up and if I canā€™t escape them, I would love to sneer at them.

  • sc_griffith@awful.systems
    link
    fedilink
    English
    arrow-up
    11
    Ā·
    4 months ago

    this is simple. we just need to train a new model for every move. that way the adversarial bot wonā€™t know what weaknesses to exploit

    • BigMuffin69@awful.systems
      link
      fedilink
      English
      arrow-up
      10
      Ā·
      4 months ago

      In chess the table base for optimal moves with only 7 pieces takes like ~20 terrabytes to store. And in that DB there are bizzare checkmates that take 100 + moves even with perfect precision- ignoring the 50 move rule. I wonder if the reason these adversarial strats exists is because whatever the policy network/value network learns is way, way smaller than the minimum size of the ā€œtrueā€ position eval function for Go. Thus youā€™ll just invariably get these counter play attacks as compression artifacts.

      Sources cited: my ass cheeks

      • sc_griffith@awful.systems
        link
        fedilink
        English
        arrow-up
        9
        Ā·
        4 months ago

        i donā€™t think that can be quite right, as illustrated by an extreme example: consider a game where the first move has player 1 choose ā€œwinā€ or ā€œhypergo.ā€ if player 1 chooses win, they win. if player 1 chooses hypergo, begin a game of Go on a 1,000,000,000 x 1,000,000,000 board, and whoever wins that subgame wins. for player 1, the ā€˜trueā€™ position eval function must be in some sense incredibly complicated, because it includes hypergo nonsense. but player 1 strategy can be compressed to ā€œchoose winā€ without opening up any counterattacks

        • sc_griffith@awful.systems
          link
          fedilink
          English
          arrow-up
          8
          Ā·
          4 months ago

          more generally I suspect that as soon as you are trying to compare some notion of a ā€˜trueā€™ position eval function to eval functions you can actually generate youā€™re going to have a very difficult time making correct and clear predictions. the reason I say this is that treating such a ā€˜trueā€™ function is essentially the domain of combinatorial game theory (not the same as ā€œgame theoryā€), and there are few if any bridges people have managed to build between cgt and practical Go etc playing engines. so itā€™s probably pretty hard to do

          (I know thereā€™s a theory of ā€˜temperatureā€™ of combinatorial games that I think was developed for purposes of analyzing Go, but I donā€™t think it has any known relationship to reinforcement learning based Go engines)