Comments by "H33t3" (@H33t3Speaks) on "The Impending AI Model Collapse Problem" video.

  1. Definetly L M A O.
    2
  2. Basically.
    1
  3. @MrMichiel1983 We're still going to run into a problem with Baysian averages, which is really how all the Neural Networks that run these things work. It becomes an average, of an average of an average. The answer I think is to determine a Degeneracy rating, IE, how many iterations behind the original data the generated data is. This relies on the same logic that dictates why embedded systems programmers abhor recursion. Weird, completely undefinable program behavior is guaranteed on a large enough time scale for a device with a recursive algorithm. We also wind up with an a priori/a posteriori knowledge problem. If the model effectively cannot distinguish what is real and what is not... if it's just a mill for information much in the same way a microprocessor is nothing more than a mill for electrons, then all it is going to do is produce more informational phenomena that will need to be authenticated for veracity by the very same algorithm that de facto cannot discern truth from falsehood: it is merely averaging information and producing an x-hat out of said field. In sum, how can it unlearn garbage if it doesn't know it's learned anything? You may as well just reset the entire model can call it a day.
    1
  4. I’ve called it “Bitrot” for 20 something years. Digital stuff DOES degrade.
    1