Modern AI is an interpolation engine. It can't extrapolate to anything new.
You understand this means stuff like Stable Diffusion is literally a fancy overpriced clipart program? It will have all the problems regular clipart has.
Basically what it does is scan the internet for pictures - forming a clipart portfolio - and then mix-and-matches the pictures. Fundamentally it's a limited, finite, exhaustible set of pictures.
It's higher quality due to lack of consent. Normally clipart collections are for cheaply finding mediocre pictures, and paying enough for quality pictures would make them not-cheap. AI grabs all sorts of expensive pictures for free. Thus AI art is...less-cheap. And that's the major difference.
Finding decent clipart takes so long it's quicker to just draw it yourself. Likewise, AI.
Clipart has a severe long-tail problem, with only a small number of decent pictures. This small set gets overused and overfamiliar. Likewise, AI.
GPT further has semantic problem. GPT can't understand meaning or grammar. It has to rely on the text it's copying knowing meaning and grammar. This means the designers' problem is trying to convince GPT to highly weigh meaningful text and discount random garble, without themselves understanding much of the difference.
All modern AIs rely on splicing, and it's a big issue with GPT due to nonlinear meaning. Since it can't understand meaning it can't tell where it's okay to splice and where it isn't.
I guess that really does nail the fact that consciousness lets you efficiently run Solomonoff Induction. GPT has to construct sentences by splicing every possible word onto the established sentence, and then trimming the possibilities the training data doesn't include. No wonder it needs to read through libraries that humans would take millennia to read through, and at the end can barely talk. Incredibly expensive way of doing it.
Let me repeat, because it's funny: GPT starts with a sentence fragment, then reads every book on the internet and picks the word these books would put there, then it reads every book on the internet again to pick the next word....
Well, it uses compression to try to avoid reading them all every time, and there's some pruning from the prompt. However, the point of the compression is to enable it to functionally read every book in the internet multiple times a second.
We must imagine Sisyphus happy, I guess.
apparently the early big imagesets and still many big services today run on armies of third worlders tagging pictures with words. the mechanical turk v3.
ReplyDeletethis means boorus/artstations are orders of magnitude better than the general dataset.
this also explains why ai art is highly rendered and soulless: those kinds of art get the tags.
Oh, so it's just cheating, like chess engines. The machine isn't thinking for itself, it's merely collating shuffling and regurgitating all these humans who think for it.
ReplyDeleteI wonder who is paying for these enormous expenses. It clearly isn't the end users.
Not only cheating at AI, it's cheating at accounting. It would be cheaper to hire an army of artists and have them draw to order, but then it would be impossible to pretend the initiative isn't losing money.
AI is highly soulless because the guys instructing the machine on art are soulless and don't know how to tag properly, if you see what I mean. Didn't hire artists (again, because accounting); garbage in, garbage out.
Chess engines play as well as their programmers do, but use strategies that their programmers couldn't carry out within the 2-minute clock. Not only is it a team vs. a solo, the team gets, effectively, days or years per turn. Plain cheating.
I could probably program a real AI myself. I can work out how thinking works because I'm not crippled by atheism. Further, if an Atheist works out how working out works, they have to explicitly set out to disprove it, to continue to uphold their atheism. Literally anti-AI, never mind artificial consciousness.
>I wonder who is paying for these enormous expenses.
ReplyDeletenothing wrong with the default of "everyone" via USD and other financial trickeries. the third worlders are paying for it: 'ai art causes inflation in argentina and starvation in sri lanka'.
"but why are the birthrates going down? it must be because men need to man up"
more directly, possibly, the ai companies. video became big because zuck? jobs? said video is the future, it's not actually because people are illiterate, it was some tech social media head announcing marching orders, and to this day companies are finding out video does not in fact have better returns than text. this is before hosting costs. youtube is now the homepage of the internet, but who pays for youtube...
https://youtu.be/5uwJzdPajQA
ReplyDelete"how do earnings contract at all when all prices are +10% to +20%" this is how. they're blowing all the money on silicon and electricity.