• kromem@lemmy.world
    link
    fedilink
    English
    arrow-up
    4
    arrow-down
    1
    ·
    22 hours ago

    You really think that an Othello board could represent the world?

    No, you seem to be misunderstanding what a “world model” is. That research was on a toy model a fraction of a fraction of the size of modern models and exclusively trained on board notation moves. And yet it modeled the geometry of a board and positions of pieces and ‘my’ vs ‘opponent’ status of those pieces (there’s been multiple follow-ups and replications looking deeper at what encoded).

    The question it answered is that transformers don’t simply learn surface statistics from the data but build world models from the data they are trained on.

    So when you take a model orders of magnitude larger and train them on collective human culture orders (of orders) of magnitude larger than the board moves in the Othello-GPT paper, what do you think happens? No world modeling at all? Well there’s been a number of interpretability papers since 2023 finding that modem LLMs have functional models of all sorts of things, even including tracking emotions for themselves vs others.

    Also, you are assuming that a trained world model is a lossless identical version, which it isn’t and nobody claimed. In fact, arguably the best way to identify AI generated videos these days is by over-expressed emotions of characters. Why? Because the training data is over indexed on influencer exaggerated faces which game clicks so video and image world models of human emotions are miscalibrated.

    Transformers have a world model, but often it’s of a kind of parallel imagined/extrapolated world that’s not the real one. Another good example of this is how most models are certain that real world news after their cutoff isn’t true because extrapolating from their internal world model (from the training data before their cutoff) those events seem improbable. The real world doesn’t exactly match the internal model and that drift grows the more they need to extrapolate.

    If you want to continue to discuss (I’m happy to help explain and educate you), I’d recommend reading up a bit on what “world model” means in relation to the tech.

    • Wildmimic@anarchist.nexus
      link
      fedilink
      English
      arrow-up
      1
      ·
      1 hour ago

      I know very well what a “world model’” is, please stop trying to undermine my credibility.

      If you feed these things thousands of datasets that only consist of a specific set of coordinates, your output will also only consist of these coordinates, which will create the illusion that they actually have knowledge about how an game board looks (as i said, these things are remixers after all). And feeding it a championship data set, where not one invalid move will appear, should ensure at least a close to 100% of valid moves if they were actually able to have an actual internal model of the system they are trying to manipulate. (and in reality, it should choose pretty good moves too, but this was not evaluated at all…).

      Their test cases (and the error rates!) were not even “play a game of Othello”, but feeding them gamestates and asking them for a legal move. IMO, this whole setup does not require any form or shape of a world model, and if it actually existed, the error rates would be a ton lower - the championship data will not contain a single invalid move, and the concept of choosing an empty space on the board, touching the other players piece with one of my own ones on the other side of the mass of pieces, is really not a hard one.

      I agree that current models will probably have a better performance, but this is NOT because of a better “world model”, but because the statistical sieve has become a lot finer by blowing up the amount of parameters.