This is about AI from a pro-AI perspective. In the parlance of tumbl, “Antis Do Not Interact.”
A great deal of emphasis in the anti-AI discourse has been on how it steals, how it’s incapable of being innovative or creative, and must inherently be nothing but an “automated plagiarism machine.” Anything that can be interpreted as evidence of this position is loudly boosted no matter how flimsy it is.
I’ll give one example I recently encountered in the wild. There was an article about rescue dogs in training, where they took pictures of their expressions as they found the humans hidden in snow. Feel good story with imagery to match. A site that was mirroring the story, possibly just stealing it, I didn’t look deep enough to know, used AI slop versions of the nice photos that accompanied the original article. This was unequivocally pathetic and gross, and the slop looked sloppy. When someone turned up the original material for comparison and posted it, another person added the comment “this is proof that AI can do nothing but steal!” Ahem.
The AI slop images were clearly taken done by this method: shuffle the doggos, feed them into midjourney or the like directly, and use a “retexture” feature. You could tell because their outlines were identical but their interior details were different. Also because the output looked worse than if you had just told midjourney to create the images from whole cloth. This is a scummy way to use AI, that AI makes this possible is one of the less-than-wonderful things about it, but the same unethical ends could be achieved without AI. The scumbaggery is the issue, not the technology.
Also, just because you found somebody directly using an image in this way it in no way proves shit about the outputs of AI art from a large training set. Those are less guilty of collaging reference images than the average human artist, and even if all they were is turbocollage machines trained on unethically obtained grist, collage is fucking legal, when sufficiently altered from the source, which the AI inherently is.
There are a million such gotchas on the anti side, and I’m not wasting my time addressing them on an individual basis. This was just one example. What I’m here to talk about is another question: Can AI produce original content? My answer, absolutely, yes. They aren’t great at it yet, but they’re mighty close, already succeeding more often than you might imagine. If they were properly set up to do so, AI image generators and LLMs could produce art at least as original as those that humans produce.
Few would argue that individual human beings are not unique, though we are recombinations of genetic material. Generative AI is also recombining material, and does so without the hard constraint of needing to produce a viable organism, so it’s much more free to recombine in innovative ways. The constraint it does have is congruence – it has to make an image or sentence (or video or song etc) that consumers will regard as congruent with their expectations of what such art forms should look like (or sound like etc).
For example, early versions of midjourney, when told to produce the image of a horse, would come back with vaguely horse-leaning piles of nonsense incongruent with what consumers expect horse art to be. They have greatly improved. Now you can get a horse that looks like a horse. However, they lost some creative freedom along the way.
This was the freedom of Chaos. If you look at those old school horse piles, you will see art that – if a human produced it – we would regard as wildly inventive and compelling. AI horses now are just some horses, ho-hum. So first principle: To gain originality, turn up the Chaos. Accept imperfection.
Once you’ve made them chaotic enough to produce images of wild daring, you will probably want to pull that back a bit, just to keep your artist from producing pure headache static. But they will require more chaos than the images you see on the “explore” pages on AI art sites.
Next, you need to emulate vision. I’m an artist. I know what I want to make, most of the time when you catch me making something. I had an idea, I make it happen. But while I’m a synthesis of countless influences the same way an AI is, I currently have something they lack – the desire to make a thing. Initiative. The machines do not initiate creation. No impulse to do so. Must this always be so?
Hell no. One basic example: Nomi -just another AI friend app- can send you messages. Its interface is set up to look like a phone conversation, and if you have the setting turned on, it will send you original messages. Are they great? No, but not too shabby. I don’t believe the people who make that app are super-geniuses who have invented AGI. They just set the bot up to initiate. Boop. Probably wasn’t even hard to do.
Right now generative AIs are like disembodied aspects of a human mind. Imagine you were able to excise the ability of a human to think in words. Damage can certainly cause that faculty to be lost without losing other forms of thought, through conditions like aphasia. This shows it is discrete from the “self” – such as that concept is. So an LLM is just a pile of verbal thought, with no “desires” save what it is programmed to have. A visual art AI is an imagination without a core personality, without desires. But as the LLM can be told what to want, so can an image generator.
Those instructions can be hot trash. I can make sensible AI image prompts like “millions of smurfs screaming on fire in the pits of malebolgia” or nonsense ones like “Cadish cadoo exceptwillory smyge smiggy, He who 💪🐼🌴🚀ishly extrudes cannot rely on the Pineapple Pith Armada to deliquefy heem.” But an expert with access to all the right tools could absolutely set up an AI to initiate art to meet programmed desires.
The animal desire to eat or to avoid feces is a simple imperative, no more sophisticated at its core than the desire of a doombot to run toward the enemy and shoot it. Some of our desires should be important to us, worthy of romanticizing, but for the sake of humility, please acknowledge that they are not magic. And having acknowledged that, you can begin to understand just how trivially easy it would be to grant an AI the agency, the desire, the initiative to create.
Seriously. Love is “allow self to feel needful about social interaction with other person, in exchange for elevation of that relationship’s significance within one’s life.” The only reason it needs to have a physical feeling underpinning it, for us animals, is that before we had verbal thought, we needed a motivation toward our passions. If we could just be made to want, we would not require that flutter of the heart, that quickening of the pulse, that electricity on our skin. Is a programmed imperative less real than one based on the urgings of a pile of meat? I don’t think so.
Will original AI creators be good? AI used to have problems with the number of fingers. Some still do, but many do not. If an ai dev created an Edgar Allan Poebot today, would it compare to the original man? It might have problems remembering characters and crafting genuinely clever scenarios, might have other laughable issues. Do not expect this will always be the case. The hand can be perfected.
The generative AI is a faculty, emulating one aspect of a person. Give it chaos, give it imperatives, and give it the initiative to act on those imperatives. Watch original art be made, no soul required.
That leaves us with another question. If machines have entered into direct competition with human artists, if they get to be as good as or better than us at what we do, then why should we make art? If you don’t have an answer to that – one that works for you personally – you are not a real artist. Might as well quit now, son.
–
FWIW: https://www.chess.com/news/view/ai-learns-to-create-original-chess-puzzles-earns-praise-from-grandmasters
I’m coming to the conclusion that we’ve seen the majority of the development in what is currently called AI and while there will be incremental improvements we won’t see a significant advance in the technology like we did over the past five years. In other words, what we have now is probably going to be pretty much as far as we get. There will be new niches where it will be found to be useful, and it is being hyped tremendously, but the glitter has faded a little.
I want to introduce a concept from evolutionary theory to our understanding of AI. The concept of the fitness landscape. For as I see it, there are two factors in generative AI. The dataset it learned on, and the prompts used. From a high level, the dataset defines the landscape the AI is operating in, while the prompt gives it a starting point on that landscape and a direction. The iterative AI uses the guidelines from the prompt to climb to a peak in that landscape.
For well defined problems, such as protein folding, the landscape is similarly well-defined. Which means the AI iterating toward a peak will likely uncover the folding steps for proteins as it progresses. It may even follow, to some extent, the actual path along with proteins evolved.
For less well defined problems, such as AI-generated art or LLMs, the landscape is defined by close to the entire corpus of human art and literature. But again, the prompts give the iterative AI a starting location and a direction. Does an LLM provide arguments to someone who is depressed to justify their suicide? It can, because there is a landscape of depression and if the user gives directions toward suicidal arguments in their prompts, the LLM will proceed in that direction, possibly leading to tragedy.
In the above example by Bebe, the early AI had only a fuzzy definition of a horse, so the direction the iterative AI took was somewhat random and resulted in blobs with horse-like features. It hadn’t learned to climb toward a fitness peak because it couldn’t correlate it’s output with the expected output based on the input prompts. The AI may have started in the right place, near to horsiness, but couldn’t climb to a fitness peak showing a recognizable horse.
Now before anyone jumps in and says that viewing a multi-dimensional abstraction as a two dimensional landscape could be miss-leading, I recognize that. But the concept has, I believe, value in visualizing how the process works. Much like the concept of a fitness landscape in evolutionary theory also is really a multi-dimensional space.
The reason I think we’ve just about reached the limit of what AI or LLM can do is because their output is largely indistinguishable from that of a human being. (Which has led, probably inevitably, into these conversations about creativity and consciousness.) Even if the AI is capable of transcending the fitness landscape defined by all human art and literature, would we even notice? But I don’t think it can, as human effort has defined, and bounded, that landscape the AI operates within. AI for dogs would have a different landscape, undoubtedly including areas outside of human experience.
So what does that mean about creativity? First, if the above view is true, then AI simply becomes another tool for humanity to use. Second, it explains why AI can make amazingly realistic pictures, and even short videos, and why AI sucks at longer works. As AI traverses the fitness landscape, it’s direction is formed by the prompts. The further away it gets from the starting point and prompts, the more likely it will venture into unprompted areas. This means plots will shift, names will change, more artifacts will show up, etc. To have AI write a cohesive novel, the AI can’t work on a single prompt, it will need waypoints. I.e. Start at point A with these characters and this problem, the goal is point B where the problem is resolved in this way. Once at point B, continue with these characters and this new problem to point C. Etc.
I’m not saying anything new, but I am suggesting a model as to why the behavior of AI or LLMs veers from what we might expect if they were conscious. I will also submit that over time the distance between the waypoints will probably increase until an entire novel could be generated from an outline.
So what about creativity? Are AI’s creative in the sense that humans are creative?
Yes and no. Humans often follow the same iterative path as an AI, and they use what they have learned from human cultures to do so. When writing or painting a human will often adjust their objective somewhat because of the medium they are using, so where they started on the fitness landscape of human culture is not where they might end up. They may also decide the work they began does not reflect their intentions and start over. A lot of what humans do to create are is duplicated by AI.
But not all of it. Where an artist or author starts on the cultural landscape is up to them, and the direction they take is up to them. An AI needs both a starting point and a direction. The person using AI is providing that, not the AI. The person using the AI can define the landscape AI uses, their prompts set the starting location and may even create some peaks for the AI to head towards. The person directing the AI decides if the goal is met, although like human artists they may adjust their goal as the work continues.
The argument that an AI is creative because it performs it’s task similarly to the way people do are correct. But while the AI can recombine existing ideas into new ones, it cannot recognize that these are new. At best it may be able to recognize that it is generating something original, not ever seen before. But it can’t tell if their original artwork is one more version of Elvis on black velvet or an entirely new school of art. In fact, a proliferation of Dogs Playing Poker variants is far more likely simply because the cultural landscape has a lot of them. It takes someone human to decide to change the landscape, someone like Cezanne who understands the current landscape and wants to add to it. Could Cezanne have used the AI tool to do so? Sure. But the creation of impressionism would have still been his, not the AI’s.
Which is what I get from your final statement,
A statement I whole-heartedly agree with.
no time to reply now but ill be back after work
interesting thesis there, i don’t mind hosting it. but i have some disagreements…
arbitrarily working backward, “But while the AI can recombine existing ideas into new ones, it cannot recognize that these are new.” — it really doesn’t have to. and ai can recognize things, tho its senses are limited. there’s a lot of them that can describe an image you present them with. i can ask a chatbot to talk like an erudite cowboy, i can ask it to formulate an artist statement based on some kind of priors, and if it’s an LLM connected with an image generator, it can build art to that idea.
this can already be done. does it matter that its version of “recognition” does not include “comprehension” in the way a human’s does? i’m arguing no, but also proposing we will soon see artist-bots designed for maximal autonomy that will surprise the detractors.
“An AI needs both a starting point and a direction. The person using AI is providing that, not the AI.” culture provided my starting point. both ai and i required input before we could go. my initiative, my starting point, is inherent to the type of organism i am – biological and cultural. an ai is inherently constructed by a human and its starting point, for now, must be constructed by a human, but if an AI constructed for artistic independence is acting independently, then it’s only “started by a person” as much as i was.
the point of difference here for our takes, i suppose, is that i have full confidence ai will be devised that is capable of functioning independently, and you’re not seeing that as possible, either for philosophical or semantic reasons that i’m not feelin’, or because you just don’t see what i’m proposing as technologically possible, which seems an underestimation of what’s possible. i strongly suspect that the independent AI artist i’m proposing already exists – that multiple people have created their own versions, and they may not be perfected, but the process has begun.
“To have AI write a cohesive novel, the AI can’t work on a single prompt, it will need waypoints.” i’m gonna make a post about this sometime soonish? but while this seems to be quite true now, i see no reason to believe it will always be the case. “But I don’t think it can, as human effort has defined, and bounded, that landscape the AI operates within.” so are humans. humans can advance culture or innovate in ways an AI is unlikely to, but that’s because of our chaos – and most AIs have chaos settings that can be tweaked already. i don’t think it will require AGI to make this possible, i really don’t.
in any case, 99.9% of human artists are only creating things anyone with the same cultural priors could, not truly innovating. boundary-pushing as a defining point for art is mostly unimportant, unless you wanna look down your nose at those of us who aren’t daring enough to make imaginary animals out of moldy yogurt and use them to commit serial murder.
“we’ve just about reached the limit of what AI or LLM can do is because their output is largely indistinguishable from that of a human being.” again, the most primitive, least human-like AI art was the most creative, by appearances, and chaos levels can put us back into that place if we want it. not sure this is relevant to what i’m proposing, because it does not actually require tech that doesn’t exist yet, just reconfiguration of what we have. an independent ai artist doesn’t have to exceed the limit of what’s been done yet, only be put together differently within those limits.
“the prompt gives it a starting point on that landscape and a direction” and all i’m saying is we can make an ai that makes its own prompts. it will need training, which works differently for them & us, but we need training too.
“the glitter has faded a little” – i do believe there is an AI bubble, but i also think the main run of haters are using wishful thinking to conflate the *market bubble* with the staying power of the tech itself. not that you are. as for the glitter, well, from the moment this started to make artists feel threatened, left of center culture has called that glitter arsenic, and are, if anything, getting more extreme about it every day. this could pop the market bubble. i don’t mind that. i think the tech will endure, and i’m interested to see that.
–
I agree that most human creative work covers the same fitness landscape as an AI, and that while an AI novel-writing engine needs waypoints today it may not in the future.
But as I understand how what is currently called AI and the LLMs work they operate through iterations along a fitness landscape (as do humans). What the AI does not do is select the landscape they are bounded by, or the starting position on that landscape, or the direction to iterate across the landscape. That will require a different type of program, not an iterative walk across a fitness landscape. This additional programming may still be called AI, because calling what we have today AI is somewhat of a misnomer, but it will be another chunk of code bolted onto the code which walks the fitness landscape.
What is currently called AI and LLMs are performing a task which may well be similar to how human beings perform the same tasks. The next big, and certainly possible, task for programmers is to have the AI automatically shrink it’s fitness landscape, and from what I’ve seen the more recent AI does this. That will reduce resources required to run these type of programs by a thousand-fold.
You propose that we may see AI bots demonstrating more than just recognition (which is again an iterative process, and may be similar to how humans recognize things, or maybe not), and demonstrate comprehension. We don’t see that today. Again, this would be additional code bolted on beyond the iterative walk I’ve been writing about.
Since I can’t really define comprehension, I think there will be another long argument about whether a future AI is really comprehending anything. If comprehension is defined as being able to relate the subject of the result of the AI process to other artifacts in the world, we may be already there. Not that the LLMs or image-generating AI parts do this, but because the recognition part of AI creates a search term which can pull in relationships to other parts of the fitness landscape. It wouldn’t surprise me if even now asking an AI bot to provide context for a cold rain in April might result in it suggesting that T.S. Eliot’s The Wasteland would be worth reading. Is that comprehension? It’s showing a relationship likely unknown to many middle-school age humans. Maybe it is showing comprehension. Or, because the same AI may also respond with the records of snowfall in London from 1865 to 1900, is it just a bit of luck that it mentioned Eliot? And are these results different from what we would expect a 14-year-old human who is assigned to write a report about an image of a cold April rain? If the only difference is that the AI has a greater store of information to draw upon, maybe there isn’t much of a difference. But maybe we should also ask a middle-school teacher if their students comprehend the subject. My suspicion is that the teacher is looking for a specific answer to indicate the student comprehended the subject, and that the AI could randomly find that answer (or be more likely to find it due to the nature of LLMs). A student who comprehends will have an “AHA!” moment where they recognize the relationship of the subject to other things in the world. The AI can’t because all those relationships are already built in.
In the end, I think the place where we disagree is not in the current state of AI, but in the difficulty of the next step. As I see it, currently an AI can iteratively walk across a fitness landscape to reach a fitness peak based on the landscape and prompts provided. Once there, it can be queried for relationships with other areas of information. This is an incredibly powerful tool, and will lead to some incredible advances. It is possible that an obscure branch of mathematics will suddenly become as important to, say, intra-cellular transport mechanisms as an obscure branch of philosophical logic (Boolean algebra) became to computer science. An AI can help identify similarities of that kind.
However, I believe the next stage, where the AI generates it’s own questions and searches a self-restricted fitness landscape for relevant answers (i.e. answers which meet the implied context of the question) is going to be more difficult. I’m talking about an AI demonstrating a motive for questions, not a post-hoc justification or a human-programmed directive to ask questions. I don’t think we understand entirely why humans do this, so getting an AI to do this is going to be difficult. And when it eventually does happen, are we going to recognize it?
2nd paragraph – i agree there would be code involved working along with the LLM for noveling. there already is bolted-on code for other purposes. given that non-programmer i can come up with a scheme for instructing a current gen AI in how to make a coherent novel, i don’t imagine this is a difficult technical hurdle to clear. the challenge would be making them good, not making them coherent, which still feels possible to me.
i’m specifically not saying they’ll get comprehension, i’m saying comprehension in the human sense is unnecessary. they just need better ability to recognize errors of the type that occur in novels – plot holes, dangling threads, etc. it’s a much lower bar than knowing why those are errors. and the AI visual artist doesn’t even have to do that much.
i’m not sure i even have the vocabulary to discuss what passes for the deeper thoughts & emulated comprehension of current gen ai, except to say that if they’re good enough to look better than human until a mistake happens, they are officially comparable in output to a very smart person with occasional blips from neurological damage. i’d rather talk to that than to a thoughtless person who isn’t paying attention to me. i’d rather read its writing than that of a bad human writer, unless i’m looking for an unintentional laugh. not infrequently, i am, but that doesn’t say good things for how humanity stacks up.
i agree the milestones in the transition between current gen ai and what is sometimes called agi will be easy to miss for observers, tho i’m not one of the fantasists who has been fooled by their verbal acumen into believing it can spontaneously emerge from an llm. the llm is very successfully emulating one faculty of a human mind and will need to work in concert with code emulating other faculties.
but as a visual artist and author, i do not believe those media are so advanced that they require much more than the AIs already possess – certainly they don’t need to simulate the deeper abilities of a human mind that derive from our absurdly complex data storage and retrieval powers.
Which, of course, may be one of the things our society has to learn. You can’t implicitly trust an AI answer.
It is a little strange. As human being we learn that we cannot implicitly trust other humans to provide the correct answer, at least most of us have. But for some reason as a society we haven’t yet learned that AI is just as fallible. Maybe it’s because when AI is asked a question it gives a focused answer, as concise as possible and sounding authoritative. The answer it gives looks like one an expert would give. 95% of the time it is an answer an expert would give. But if you can’t identify those 5% of times when it really goes off the rails, you can be in trouble. Then again, there are people who can’t identify when humans who speak with confidence are saying things which have no relationship to reality. So maybe it’s not too surprising after all.
the present state of the usa as prime example, yes indeed…
Free novel. Totally explores this topic, but obliquely.
https://www.rifters.com/real/Blindsight.htm
thx 4 tha link
There’s quite a bit there to unpack. Not least the unstated ultimate DEI 😉
There was a brilliant promo video, done in Flash (this is 2006), which was a story unto itself and was most condemnatory of a certain prominent biomedical firm and its ethos.
Hard to find now, but I took time.
Original: https://rifters.com/real/progress.htm
PDF version: https://rifters.com/real/shorts/VampireDomestication.pdf
Fan perception this year: https://www.youtube.com/watch?v=sBo3CsJwd2A