• 0 Posts
  • 38 Comments
Joined 1 year ago
cake
Cake day: June 14th, 2023

help-circle






  • While it’s true that linear algebra and vectors are used in learning models, they’re not using the term correctly in a way that says they know something about the subject (at least, the modern subject). Concepts aren’t embedded as vectors. In older models (before the craze), concepts were manually embedded as numbers or a collection of numbers, which could be a vector (but could be something else as well), and the machine would learn by modifying weights. However, in current models (and by current, I mean at least more than a couple years), concepts are learnt by the machine (weights are still modified by the machine as well) and the machine makes its own connections between features presented to it.

    For example, you give it a dataset of 10x10 pixel images (with text descriptions) and it reads that as 100 pixels split into 3 numbers (RGB) and then looks for connections between those numbers and in which pixels. It’s not identifying what a boob is, but knows that when an image has ‘boob’ in the text description then there’s a very high likelihood that there will be a circular collection of pixels with lots of red somewhere in the image that are also connected to other pixels that are often also lots of red. That’s me breaking down what a human would think given the same task/information, but the reality is the machine will come up with its own connections/concepts which are both often far better than humans (when the model works, at least) and far more ineffable to humans.








  • People dismiss AI art because they (correctly) see that it requires zero skill to make compared to actual art, and it has all the novelty of a block of Velveeta.

    I look at art because I find it pretty, not because someone toiled over it for hours on end. Sure, I respect the artist who made it and think their effort commendable, certainly worth a sum of money, but if something is made such that the art of the craft requires less skill and time surely that is a good thing?

    Novelty of the tool doesn’t matter. What’s new changes daily, and the point of a tool is not to be new but to be useful.
    If you mean the art itself that is generated being samey or problematic in that sense of non-uniqueness, I disagree wholeheartedly. You can do a lot with learning models, and the sameness people perceive is from inexperienced novices dipping their hands in and flooding the ecosystem with beginner works, in much the same way DeviantArt was once flooded with drawings on the level of stick figures and box people.

    If AI is no more a tool than Photoshop, go and make something in GIMP, or photoshop, or any of the dozens of drawing/art programs, from scratch. I’ll wait.

    A hammer is a tool, and so is an electric jackhammer. You don’t tell a construction worker to go use a hammer when an electric jackhammer gets the job done far better and far more efficiently, and not everyone is suited to using a hammer just as not everyone is suited to using an electric jackhammer. They also have different purposes, but certainly the electric jackhammer did replace some of the uses the hammer once had, but it doesn’t make the hammer obsolete. I view learning models that generate art in the same manner as an electric jackhammer. Useful and powerful, but ultimately lacking in refinement and the work will certainly need other tools to finish the job.

    This phrase of yours just doesn’t mean much. I don’t see how making something in GIMP proves anything for anyone?


  • Humans certainly don’t make new things out of nothing. They also take from different sources and combine them together to make something new, whether that’s direct inspiration or on a more abstract level through the brain.

    Learning models aren’t generating art any more than GIMP or Photoshop is. It’s the person behind the tool that makes the art, not the tool. There’s certainly an art to prompt smithing.

    I feel like a lot of people dismiss generated art simply because it’s new (and because as a byproduct is spits out dozens of junk pieces before getting anywhere good). I don’t see how it’s that different from someone using photo-editing software built with dozens of algorithms instead of a ‘pure’ drawing pad, or someone using a drawing pad instead of a pencil, or someone using a pencil instead of chalk. It’s a tool, and a great one at that in comparison to many digital tools for artists.


  • It’s to keep design space open and to minimize developer work.

    Let’s say we decide to keep an overperforming gun. It does all the things. It has all the ammo, all the damage, all fire rate, all the reload speed. Now, all future weapons have to be made with that as a consideration. Why would players choose this new weapon, when there’s the old overperformer? The design space is being controlled and minimized by the overperformer. Players will complain if new weapons aren’t on the level of the overperformer.

    Now, let’s say we have ten weapons with one clear overperformer. Now, we can either nerf a single weapon to bring it in line with the others, or buff nine weapons to attempt to bring them up to the level of the overperformer. Assuming the balance adjustments of each weapon are the same amount of work, that’s 9x the effort. However, if we assume we do this extra work to satisfy players, now we have ten overperforming guns and players find the game too easy, so now we also have to buff enemies to match. However, the game isn’t designed to handle these increase in difficulty. Players complain if we just add more health to enemies, so we have to do other things like increase enemy count, but adding more enemies increases performance issues. It’s a cascading problem.

    I consider nerfs a necessary evil. It’s absurd to ask developers to always buff weapons and give them so much work when they could be developing actual additions to the game. Sometimes, a weapon really does need a nerf.