Larian will no longer use GenAI for Divinity concept art, and any genAI used for other games will be "trained on data we own"

Comments

• null@piefed.nullspace.lol ⁨1⁩ ⁨day⁩ ago

  Is it even possible to create a model on just their content?

  source

  • Grimy@lemmy.world ⁨1⁩ ⁨day⁩ ago

    Maybe if they mean fine tuning but from scratch, no.

    It’s the main reason I think the whole anti-AI movement is going in the wrong direction. If we all don’t get open access to it, that means the access is dictated by sites like shutter stock and deviant art. It doesn’t go away.

    source

    • null@piefed.nullspace.lol ⁨1⁩ ⁨day⁩ ago

      That’s my thinking, if the concern is that they want to be positive that no copyrighted materials were used in the toolchain, that really just doesn’t sound feasible for a single game dev to pull off.

      source
    • KoloradoKoolAid75@lemmygrad.ml ⁨1⁩ ⁨day⁩ ago

      deviantart

      Does that mean all of the ai pictures will become a variety of gay furry inflation porn?

      source
  • NuXCOM_90Percent@lemmy.zip ⁨1⁩ ⁨day⁩ ago

    Yes and no.

    Yes in the sense that they could write a model completely from first principles as it were. The algorithms to train the models are pretty trivial. Providing source material to train the model on to specialize it is also trivial… if you have it (which Larian presumably would).

    The (vastly simplifying so anyone who wants to “well ackshually” can go suck Yurgir’s fat one (negative)) initial weights are the problem. Think of it like what is required for the model to even understand what “give me a weathered stone exterior texture” means. THOSE are fundamentally built on stolen IP (and the uncredited work of grad students around the world…).

    How much you care about that is up to you. But that is what facebook et al had seedboxes running 24/7 to steal. They might not train “their model” on your favorite author’s work. But they used your favorite author’s work and previous generations of their model to create the initial weights they optimized on.

    And a “from scratch” model will not have those. Many are trivially easy to find but those are also very poisoned.

    source

    • null@piefed.nullspace.lol ⁨1⁩ ⁨day⁩ ago

      Exactly what I’m thinking. If it’s an ethical line then the answer is to just not use AI full stop. If they can’t do that, then this is really just about optics.

      And if they can’t stop using AI, it makes me wonder why I keep seeing people say it’s so useless…

      source

      • -> View More Comments
  • wholookshere@piefed.blahaj.zone ⁨1⁩ ⁨day⁩ ago

    yes?

    I’m sure they have enough content in the data store.

    source

    • null@piefed.nullspace.lol ⁨1⁩ ⁨day⁩ ago

      I’m sure they can train an existing model with that, but my understanding is that basic functionality requires a huge amount of broad data to train the base model

      source
  • warm@kbin.earth ⁨1⁩ ⁨day⁩ ago

    Why even bother? New concept ideas are better coming straight from the mind rather than using a model to regurgitate existing things.

    source
• Philharmonic3@lemmy.world ⁨1⁩ ⁨day⁩ ago

  Or… Just don’t use it? What the fuck?

  source
• eleijeep@piefed.social ⁨1⁩ ⁨day⁩ ago

  CEO needs to step down.

  source
• belated_frog_pants@beehaw.org ⁨1⁩ ⁨day⁩ ago

  We want art and ideas from humans, not slop.

  source