EDIT: LINK - https://reddit.com/r/SneerClub/comments/yodmju/i_really_thought_this_was_a_brilliant_satire_at/

The uber-cool techbro is in the comments giving us such gems as:

I stand by what I said. People will implement it soon, if they haven’t already. Aesthetic rating networks are a thing, and image generators are capable of combinatorial generalization, so it’s probably possible to use search (or maybe even gradient descent) to find images that are better than the ones in the training set (according to the metric), and then train it with those. The success of these techniques depends on the critic not being goodharted, so the results might be inferior to training it with human-curated data, but that is more expensive. Is there any flaw in this reasoning?

When asked what in the living hell is an “aesthetic rating network” he replied:

Take a set of prompts. Generate many images for each. Select the best ones according to the network as long as they are sufficiently realistic (according to the generator or other net) and still match the prompt. Finetune on those. Or something like that

My criteria for best image is whatever someone considers the best image. This varies between people, but models can take this into account. Other areas of art (all of them?) also follow the pattern of there being a data structure that people can prefer over others, and optimizing it is a problem that machines will eventually basically solve.

I’m not mistaken. People have preferences over trajectories reality can take. Part of that considers whether what they see is pretty (but obviously art is about more than that). If you want to solve art (or understand it properly at all) you need access to that rating function. You can do it by either studying the brain directly or by observing human behavior (like the score they give to an image) and fitting a model to reconstruct that part of their minds. I’m pretty sure the vast majority of artists don’t think about art this way, but that’s how you study it in mathematical terms.

:chefs-kiss:

I love techbrains. They’re so unused. Completely fresh and wrinkle-free.

  • wackywayneridesagain [none/use name]
    ·
    2 years ago

    I do machine learning optimization and other data engineering stuff for a living. Not for exploiting people thankfully, though probably automating away some jobs in a couple years' time.

    I think they're missing that art is ultimately consumed by humans. We can definitely use AI to simplify making art, by generating things that artists can manipulate. That's already been in use for years and is just accelerating a bit now.

    And we already train those models using the output of previous executions. Where the poster is wrong is whether this will be usable to replace human input completely. That's not realistic to me, at least not before virtually every human is dead.