#ai topic

[tiredfoxtf]

No, but this is actually crazy in the most dystopian sense.

[luwha]

THIS ISN'T A ROAST. It's some thoughts and backgrounds on a thing that's been eating my brainworms for a bit.

haha ok so, the last few years I've been hearing a lot about how AI can be useful and a tool, which i refuse to believe it is (and if so, the damage it causes to artists FAR OUTWEIGHT any benefit).

So, there's this youtuber that covers, like, tech and stuff. When generative AI for images and text came up , they were rightfully poking at the ethical dilemma, and copyright infringement, but STILL made a soft argument about how useful it can be on pitch ideas, concepts, and other things that quick and cheap art could be used (also fuck whoever did this jobs before i guess lol)

BUT THEN an app finally did the same for music, generating indistinguishable quality music. The said youtuber makes music, usually vibe check music to hear on video's background and such.

but see

this time they were frustrated, and made a video about the ethical concern, how unfair it is, how it destroys the effort and passion of creating and all, and did NOT make an argument about how it could be useful, because now they knew that despite the benefit, if it goes off, their dreams is over.

A lot of people are quick to jump into "this can be useful" for generative tools, and very rarely i see artists, especially digital artists, agreeing with this. And most that do either have no stakes on their careers, aka they're already well off, or have no careers to begin with.

Serious artists KNOW how bad it is, no matter what. The youtuber had no stakes in visual art. They doesn't make them, they don't really profit or benefit from making them.

When it came to a thing they DO understand and make, now they've got to understand why it's bad. Why the benefits are not good enough to justify the disaster it makes on artist's life, on the humankind development as a whole.

idk what the point is here, i just really got pissed that at the point, they've been really kinda "this is bad but this exist, who knows, could be good if it's not unethical etc", to go fully into "this is bad. like, real bad for everyone here and music industry".

Wish that understanding came from their kindness, not AI coming with a baseball bat to their knees. Makes me real sad that it came for them too.

[cetaceanhandiwork]

the conversation around generative neural networks is a dumpster fire in a dozen different ways but I think the part that disproportionately frustrates me, like on an irrational pet peeve level, is that nobody in that conversation seems to understand automata theory

back before most of these deep learning techniques were a twinkle in a theorist's eye, back when computing was a lot less engineering and a lot more math, computer scientists had worked out the math of different "classes" of computer system and what kinds of problems they could and couldn't solve

these aren't arbitrary classifications like most taxonomy turns out to be. there's qualitative differences. you can draw hard lines: "it takes class X or above to run programs with Y trait", and "only class X programs or below are guaranteed to have Y trait". and all of those lines have been mathematically proven; if you ever found a counterexample, then we'd be in "math is a lot of bunk" territory and we'd have way bigger things to worry about

this has nothing to do with how fast/slow the computer system goes; it's about "what kinds of program can it run at all". so it includes emulation and such. you can emulate a lower system in a higher one, but not vice versa

at the top of this heap is turing machines, which includes most computers we'd bother to build. there's a lot of programs that it's been mathematically proven require at least a turing machine to run. and this class of programs includes a lot of things that humans can do, too

but with this power comes some inevitable restrictions. for example, if you feed a program to a turing machine, there's no way to guarantee that the program will finish; it might get stuck somewhere and loop forever. in fact there's some programs that you straight up can't predict whether they'll ever finish even if you're looking at the code yourself

these two are intrinsically linked. if your program solves a turing complete problem, it needs a turing machine; nothing less will do. and a turing machine is capable of running all such programs, given enough time.

ok. great. what does any of that jargon have to do with AI?

well... the important thing to know is that the machine learning models we're using right now can't loop forever. if they could loop forever they couldn't be trained. for any given input, they'll produce an output in finite time

which means... well, any program that requires a turing machine to run, or even requires a push-down automaton to run (a weaker type of computer system that can get into infinite loops but that you can at least check ahead of time if a program will get stuck or not), can't be emulated by these systems. they've got to be in the next category down: finite state machines at most - and thus unable to compute, or emulate computation of, programs that inhabit a higher tier

and there is a heck of a lot of stuff we conceptualize as "thinking" that doesn't fit in a finite state machine

...I suspect it will some day be possible for a computer program to be a person. I am absolutely certain that when that day comes, the computer program who's a person would require at least a turing machine to run them

what we have right now isn't that. what we have right now is eye spots on moths, bee orchids, mockingbirds. it might be "artificial intelligence", depending on your definition of "intelligence", but prompt it to do things that we've proven only a turing machine can do, and it will fall over

and the reason I consider this an "irrational pet peeve" and not something more severe? is because this information doesn't actually help solve policy questions! if this is a tool, then we still need to decide how we're going to allow such tools to be built, and used. it's not as simple as a blanket ban, and it's not as simple as letting the output of GNNs fully launder the input, because either of those "simple" solutions are rife for abuse

but I can't help but feel like the conversation is in part held back by specious "is a GNN a people" arguments on the one hand, and "can a GNN actually replace writers, or is it just fooling execs into thinking it can" arguments on the other, when the answer to both seems to me like it was solved 40 years ago

[tiredfoxtf]

A message to artists against AI. From @/dannyphantomexe on YouTube.

[cetaceanhandiwork]

Have you seen the AutoGPT framework? That adds a scaffolding to LLMs so that they can run indefinitely with a memory store, would that be Turing complete?

so there's three caveats here:

it's been 20 years since I actually studied this topic, and have forgotten like 95% of what I've learned

a wrapper program that runs an FSM in a loop with extra input from an oracle (the internet) is a lot harder to reason about than an FSM on its own

I'm found the GitHub for this but I'm not gonna read that many lines of code for free

all that said, my initial skim of the AutoGPT codebase is that the way it's implemented is making some pretty extraordinary assumptions about GPT's ability to generate sensible results for the kind of prompts it uses.

it's essentially trying to break down work into bite sized pieces by telling the text generator to:

deepdream a bureaucracy of specialized task runners

handle a user request by writing delegated tasks for members of the bureaucracy

execute those tasks as though you're the recipient member of the bureaucracy

repeat until "done"

there's a couple different ways that this can go wrong

first off, it's not clear to me whether the above hierarchical breakdown of work is being done in a way that's allowed to loop "until done". if not, we're back at FSMs.

second off, it's not clear to me whether the text generator can generate subtasks competently, the way AutoGPT is requesting them, without already being turing complete. I see a lot of hay being made in the prompts about "explain your reasoning", as context to be passed along to future invocations in order to produce more meaningful results, but answering that prompt requires an amount of introspection where I'd be surprised if an FSM was capable of generating a real answer, instead of some "sounds normal" mimic handwave. and if these output fields are garbage, then the proof by induction falls apart that the preserved context is making outputs better and not worse; you'd get something that maybe has all the physical organs to emulate a turing machine, but miswired such that it'll never actually succeed at computing anything beyond the sum of its parts.

[bitcoitus]

[khuantru]

Man.. I don't know..

the use of A.I is a complex augment.

It's good & bad. 🤖🖥️

(is a user an artist / illustrator if they use AI tool. The same can be said about music. I'm starting to see & hear AI music in all forms of genre. Are they artists as well?)