I've tried out a bunch of portable note-taking apps, and Napkin is by far my favorite because it's fun. Check out Napkin : napkin.one/signup?via=ArtPostAI
This has already happened with ai writing. Four years ago I tried out the earlier models, and while the storylines were barely coherent they felt weirdly original and fun. Now when I ask chat-gpt to make a story, it's coherent but it takes any cool concept I give it and makes the most generic version of it.
1 It definitely will not replace artists with a unique style, or artists that have developed a very high level of skill or design knowledge. It also can't do fine art (but can mimic it digitally to a degree at least). 2 It likely will replace artists who made relatively generic art, like the kind you might find in many mobile games. There are still huge limitations, I don't think any models can generate vector graphics for one. 3 one of the greatest utilities of AI is using local, highly customizable models like Stable Diffusion. The output of such models can be used as part of an artists workflow, or fine tuned to produce images that are much more specific and customizable than models like midjourney or dalle can output. Some of the workflows for Stable Diffusion are fairly technical and labor intensive to a degree which is actually pretty cool. It's still much different than having direct control over what you're creating but I don't think art requires that. But a lot of these local models, or at least the popular configurations of them have certain issues like saturation being way too high, perspective being off, etc. Which again points back to the value of being an artist and knowing design. All in all I don't think mostbartists should fear AI at the moment. There's many people saying they believe many of these models are starting to plateau for one. And besides that some artists will learn to use AI themselves or it will drive them to carve out a niche that AI can't replicate well or at all. If art is your passion you can't let technology discourage you from creating.
Your conclusion is exactly right in my opinion. 1- it can mimic fine art to the degree that normal people can't tell the difference. But it'll have big trouble trying to mimic the style of living, breathing artists who (my guess) are going to get much more exploratory with their style if people start training AI models specifically to target their style. 2- Completely agree. And that's pretty sad because these kinds of jobs are often a gateway into creating more worthwhile art and building valuable careers. 3- Completely true. And I think it'll get more and more customizable in the future. Just like AI researchers of my generation aren't diving into the nitty gritty of optimization algorithms or GPU optimization and have more time for architecture, dataset and loss function design, those of the next generation will be focusing on things more accessible to the public.
Great video! I like how you used how AI actually works as evidence to support your claims. Lots of people just see the line go up, but don't know what that line represents (including me) so this is a really valuable approach!
This is one of the best presented arguements on the limitations of AI Art I have yet seen. I had not considered that as a 'problem solving' technology AI is not really suited to being creative. In a sense AI Art is the antithesis of creativity because it's not really trying to communicate anything- and all creativity is about the desire to express something; an idea, a feeling or even just a particular mood or moment- AI's have no such aspirations, they really are just goal orientated machines trying to generate the optimal solution to the problem presented to them by the prompt input by their operator, with a degree of randomness thrown in. Which is not deny that some AI generated images are not genuinely impressive and strikingly novel- but even the most impressive AI Art is still lacking in the essential thing that distiguishes Art from mere pattern making. Give a monkey some paints and brushes and they may create some vivid and striking marks on a canvas- but we would not claim that these marks are works of art because we know that the monkey had no idea what it was doing- it was not making those marks in order to communicate with us, or to make us feel something, and lacking this intentionality it's mark making has no meaning- they are just marks on a canvas. And while it's true that the patterns created by AI are far more sophisticated than the marks made by the monkey, they too are not art because- just like the monkey- the AI was not trying to communicate anything- it was just generating patterns in reponse to the prompts input by it's user. A recent experiment was done to see if people could tell the difference between AI generated faces and photo's of real people. The result was interesting because while the test subjects claimed that they could not tell the difference their brains were seen on EEG to react differently to the AI generated images- so on some unconscious level the test subjects were able to identify which faces were real and which were AI. The reason for this is because humans are very good pattern detectors in their own right- and there were patterns in the AI generated faces that had registered on the brains of the test subjects on some subliminal level. OK- so if the ability to detect patterns in data is in part a function of sample size what happens when people become exposed to huge numbers of AI generated images? It seems likely that as we see more and more AI Art we will become more and more adept at recognizing it, to the point where AI Art will become a Genre of Art in it's own right- already the 'AI' look is something that many people recognize. It's subtle- AI can mimic many styles and mediums, so in theory it should not be possible to distingish AI Art from human made art- but in practice this is not usually the case- the more AI Art you see, the more your brain is 'trained' to detect it's presence. The reason for this is kind of obvious. If I use a tool to make a thing, that thing will inevitably bear the imprint of the tool I used- the artefact created will be an expression of the tool used to create it. And indeed among AI Artists there are lively debates as to the relative virtues of different tools- is Midjourney a better tool than Dall-e for example?. These debates only make sense if it can be said that both these tools leave an imprint on their creations that is identifiable in some way. That being the case what happens if millions of AI Artists use Midjourney to make their Art? The answer is that they will converge- their creations- all bearing the imprint of the tool they are using- will begin to look similar to each other. OK- they might argue- can the same thing not be said of a paintbrush or a sculptors chisel? Well, not really, because there is a huge difference between the input a paintbrush has on the final image and the impact that Midjourney has on the final image. In the case of the paintbrush the behaviour of the tool is massively influenced by the hand that holds the brush- because each artist is a unique person with different skills, aspirations and experiences. In the case of Midjourney on the other hand the influence of the operator is very much less- indeed Midjourney does not even need a human operator- we could simply connect a chatbot like GPT and have it generate random prompts that Midjourney could turn into images with no human artist involved at all- just one AI prompting another AI. So the endgame of AI Art is that of a rising tide of images that all exhibit similar artefacts as result of being the products of the same tool- and as that tide rises those artefacts will become more and more obvious as our brains learn to recognise them. At some point the 'AI Look' will become so obvious that the use of AI Art will become the hallmark of art that is cheap and low effort and any commercial application of AI Art will be seen as a lazy and inferior solution. For this reason the long term impact of AI Art on the market for Art may not be as great as many people expect.
Thank you for this comment -- it's a masterpiece ! "AI Art is the antithesis of creativity because it's not really trying to communicate anything" "even the most impressive AI Art is still lacking in the essential thing that distiguishes Art from mere pattern making" I'm taking notes ! Do you have the reference to that EEG experiment ? It sounds super interesting !! "It seems likely that as we see more and more AI Art we will become more and more adept at recognizing it, to the point where AI Art will become a Genre of Art in it's own right" As you say ! I had the intuition that something like this was about to happen but didn't have any scientific references to back it up ! It's also very true that using DALLE or Midjourney isn't like using a brush or a chisel. It has much more effect than these simple, unmoving tools. And the way they are made gives very little control over them, even for AI researchers. If you want the best quality, you have to use all the training data, and the best architectures and optimizers. Trying to come up with something new, you sacrifice quality. Although, I think that new ways for AI to make art are coming. To me, it's pretty clear that with improvements in Reinforcement Learning (and maybe even LLMs, they do anything with them these days), there'll eventually be computers that draw at a professional level by actually making lines on a physically simulated canvas. I feel that at that level, there'll be quite a lot more wiggle room for changing the "style" of the AI-painters (changing the kind of hand the robot controls, the tool it's weilding, augmenting gravity...). But that's pure speculation, and even then, since it is the AI itself drawing, it might be that the limited number of styles we can get it to use will all be detected by humans after a while. So your point would still stand for these too. Thanks again for the super insightful comment !
@@ArtPostAI Thanks for the kind words. I don't have a reference sadly for that experiment- I read it online about three months ago and it stuck in my mind. There is a test you can take for yourself though; go to Artstation and look at the gallery- which is mostly human made art- then go to Deviant Art and look at their AI Art gallery- if you flip between them it's clear that both galleries have a distinct and different aesthetic- despite the fact that the images in the AI gallery were literally trained on Artstation artwork ( without the permission of the artists of course.) The strange thing is that if you asked me to explain that difference in words I would struggle to define it, but it's there- somehow the AI gallery has a specific 'feel' despite the fact that it contains a diverse range of styles and subjects- and the same is also true of the Artstation gallery- both somehow reveal their provenence as being either human made or machine made- which is I think reflective of the fact that the the visual 'DNA' of AI Art is somehow exposed in the images it creates- and as humans we are capable of detecting this visual 'signature', especially when we see a lot of AI images gathered together. In a way this should not be really surprising- any process or tool must surely leave it's mark on the things it makes- this is true of a flint axe and is equally true of an AI programme in which specific training data is used and specific algorithms are deployed- the idea that AI Art machines are totally transparent and unbiased in the way they respond to user's prompts is really a kind of magical thinking. Perhaps one day AI's really will be able to read our minds and create perfect representations of the images we want - but todays AI Art machines are far more crude- first they must translate our words into images- an imperfect process in itself- then they execute their imperfect understanding of our intent using a specific and fixed array of trained weights which are themselves shaped and constrained by the specific algorithms used to run the programme. It would be genuinely surprising if this cumbersome process did not leave distict visual 'fingerprints' on the images it made. My view is that at some point AI Art will reach 'critical mass' in the sense that it will come to be seen as a genre in it's own right. This will not invalidate it as a solution commercially-but it will create a problem for anyone wishing to 'stand out' in a marketplace saturated with AI imagary. I've already heard from some artists that their clients want them to use AI (for reasons of cost saving) but are specifically asking them to avoid that 'AI look'- how this contradictory feat is to be achived no one seems to know, but it's interesting that already the term 'AI Look' has been coined and is in circulation among some who commission commercial art.
Do you think it would be possible for the set goal or optimization measure of “creative” AI to be a high ‘score for creativity,’ rated subjectively over a large population of humans? And then once AI knows how to create artwork that humans find creative, it could use that same algorithm to create “creatively”?
Thank you so much for the insightful question ! It makes me think... I'm surprised I haven't heard of such an approach in AI. I'll have to keep an eye out. I have two objections that come to mind immediately though : 1- There's a difference between what people call creative, and what actually is. Though there is some overlap, there are many eminently creative people who are dismissed in their lifetimes, only to be recognized later on (or not at all). 2- If you train this AI at a given point in time, it will learn to produce things that are judged to be creative right now. But what is creative today is passé tomorrow. So you'll need to ask people again, and train the AI again. 3- No art is judged creative by everyone. The most creative outputs from the AI are likely to cause division among the raters, which means when you average them together to create the final objective function, you get an average creativity value. I'm just guessing though. Do you agree ? Did I misunderstand the question ?
^ What he said, but also, remember that basically every quality you want to add to a model becomes just another label. "Creativity" can easily be trained as Label 20876 in a model, when you have pieces of data you've labeled as "creative". But when it comes time for the AI to perform on its test data, labeling something with a label you've already been given isn't very creative, is it? We don't want the model to be creative in a way we simply label as "creative" in our toy model: we want it to be creative in a more meta way. We want it to think outside-of-the-box, in a way a computer fundamentally can't. If we lose our phone in our house, we know that if we can't find it after searching for a while, it's more likely that a family member accidentally took it. Maybe the dog got it. Maybe we didn't actually lose it in the house after all. Posed with the same problem, an AI won't imagine that a family member might have it. It won't consider that maybe the dog buried it in the yard. It won't consider that maybe we just left our phone in the car. The AI will look behind the stove. Under the chest in the attic. On the ceiling. It cannot comprehend that there are other possible solutions, unless we explicitly tell it that there are. This same thing applies to all other aspects of life, and all aspects of *those* aspects, ad infinitum.
Hey, awesome video! Here are some questions that popped up while watching, hope you'll provide your insight. a) Considering that the Likelihood is being optimized, wouldn't a large enough training data set result in solutions that we perceive as creative? I see how a 'small' training data set would lead to boring solutions, though wouldnt a large enough data set allow for a larger set of valid solutions -> More variations -> More 'creativity'? Since regardless of whether it actually is creative or not, I'd argue that the chance for a human to create new art that bears no similarities to a piece of art already present in the internet to be miniscule. b) Secondly, aren't humans also in a way heavily influenced by their "training data", as in their experiences throughout life? I do wonder how different that is in comparison to AI.
"result in solutions that we perceive as creative" It already happens, though 🤷♀ The way humans create is by understanding things we've seen before and extrapolating on it. So, in essence, it's true that with enough data, you could recreate "the human condition". But it's a gross oversimplification of how realistically we can categorize anything, and just how much data that would entail. It's like "Step 2: Draw the rest of the owl". Extrapolation is very human. If something does not fit within the bounds of what we've seen, we can extrapolate out into a constrained set of possible answers. We know *when* it's important to extrapolate, as well. An AI can't tell without being explicitly "told" anything: it only interpolates. Say that we have an item A: it looks like a duck, swims like a duck, and quacks like a duck. Humans know it is a duck. AI knows it is a duck. That's good! Say that we instead have item B: it looks like a duck, swims like a dolphin, and neighs like a horse. Humans know that this is some amalgamation of random animal traits that doesn't actually exist: a chimera. AI knows it is a duck. That's wrong (or at least, misleading). It can only interpolate based on prior data it was been fed. If could "tell" the AI that there is a secret category it can use to label unknown animals, but then we have to label *those* for training the AI... and so on and so on. We can, in theory, model extrapolation as interpolation if we have a truly mind-boggling amount of data and the perfect algorithm. But I could, in theory, buy everything on the planet if I had all the money in the world, too. Coming up with (accurate) labels for unseen things on-the-fly is closer to "true" AI. That's the ability to extrapolate as humans do. AI interpolate by design.
Tough questions ! Thank you so much :) a) "Considering that the Likelihood is being optimized, wouldn't a large enough training data set result in solutions that we perceive as creative?" -> This is an open question in AI. What's known as "the bitter lesson" is that compared to the benefits of scaling up the data and nb of parameters, all the effort we researchers put into designing better models is basically useless. Therefore, is all that's required for creative AI not changing the objective, but instead just adding more data ? Actually I hadn't thought of it, but it seems pretty likely to me that with scaling up the training data, AI will become more and more flexible (thus, maybe you'll be able to get the proverbial ramen bowl without the chopsticks). As such it will allow people to be more and more creative through it. So you have a great point here actually. As far as making the AI itself creative, there is still a prompter behind the AI. But with the comments on this video, I'm kind of reconsidering my use of the word "creativity". b) "aren't humans also in a way heavily influenced by their "training data", as in their experiences throughout life? " -> to me, it is different in that we don't have as a unique and absolute objective to reproduce that training data. Does that answer it ?
likely does not equal typical at all. you can literally find the word "typical" within the space of meaning that the AI operates with, and therefore you can make its "tunnel vision" aim at what is the opposite of typical. you can also, in principle, use this as a mere basis for the overall "creative process", by making an ai search for very un-typical associations between words, phrases etc., to create more un-typical prompts, and thereby create more un-typical pictures.
he adresses that actually, his point is that making models aim to be not typicall makes their outputs more random. but radomness here does not mean exotic outputs that are still coherent which is what we would like, it leans more towards white noise and give outputs with little value
@@question_mark "a bird that is as far from a typical bird as it gets" might not be the easiest prompt to make use of in the intened sense for an ai, but there is certainly a phrasing of this concept that could be used as a prompt, which would certainly not just lead to irrelevant randomness. but to an "exotic" (non-prototypical) bird - plausibly to a penguin, for example. just as human creativity benefits from (plausibly necessitates) limitations, (like refering not just to any non-typical concept, but to a non-typical bird, for example) so would ai-creativity. in fact i think it is also very implausible to conceive of creativity as independant of goals. even artists making art "just for the sake of it" do so out of more or less conscious drives. and any drive a human might have can be used as a blueprint for some type of training-goal for an ai.
@@davejacob5208 For the arbitrary prompt, yes but I guess then there is not really something analog to our creativity coming from the AI's side And I also think that creativity does not imply not having some kind of goal, I even made a comment on this video explaining why I think so x)
the creative force here is still technically the person writing the prompts, as the generative model itself cannot reason, and it does not "know" its own tunnel vision exists. it would act as it always does, a brute force hammer approach to creation
This video is quite good. However, I find challenging this idea of 'creativity'. How do we test for 'creativity'? How do we set a null hypothesis to disconfirm it? What assumptions about modern AI are we making? It seems like perhaps a mythical goal, rather than a practical one (AGI anyone?) I suspect when we say a person is creative, we really mean that there is a sense of perceivable 'vision' which went into something. It's hard to talk about a certain person being more creative than another. It altogether depends on the audience perceiving a sense of salience, a sense that there is a signal worth fishing from the noise. Is the banana taped to a wall creative?
Very keen observation. Seems you know quite a bit about AI ! That's actually one of the things I'll discuss somewhere in the next few videos is creativity a word like AGI (I completely agree with you on that one), time, or even "consciousness" ? Something we keep talking about, but can never grasp ? Maybe I should look for another word.
We're all creative, but it varies from person to person. One person might be creative in art, another on music, or cooking or building. It's an essential part of being human. AI can never be truly creative. IMHO.
@@JaneNewAuthor Let me put this another way. Suppose you met some extra-terrestrials. How would you test the creativity of this new species (of intelligent life)?
@@AZTECMAN they would be creative in their own way. In the most general sense, all life is creative. Look at the patterns on a seashell. When an artist creates they're using their skills, talents, experience and intelligence, but they're also using something indefinable that makes them human. Life experience maybe? AI will never be able to compete with that, because it's not alive.
I am very reluctant to place my trust in 'never'... but I tend to agree with the attitude. I happen to believe at the moment that there isn't any alive AI... however I'm not convinced this is impossible. Are you familiar with John Conway's "game of life"? Cellular automata are quite a bit of fun.
Thank you for sharing details! I feel like there's a lot of people with opinions about AI, but few people who have a strong foundation in both AI research and the domains it impacts, so if you keep sharing deep insights into AI, that would be really valuable :]
The problem with AI's lack of creativity is that it's trying to mimic something that we do so instinctively that we don't even know how we do it in the first place. We're trying to achieve the end by setting up networks that mimic the human brain in structure, but we fundamentally don't understand how the brain makes connections between ideas and how it determines relevancy. If we don't know how that works, how can we possibly program a computer to do it? Because all computers do is follow instructions. Those instructions may be complex algorithms, but they're still just instructions that the machine will follow without deviation.
You're hitting straight at the big question in AI : how do we get these computers to do stuff without having to tell them exactly how ? So far, the best answer is : define a hard-to-achieve objective, don't specify the steps to get there, and run an optimization to have the neurons figure out the steps on their own. It works well enough for AI art to be a thing without anyone having manually touched any of the weights in the network. The fact is : we don't know if it does or if it doesn't do the same things that happen in the brain. Because we don't know how either works. But from the little neuroscience we know, my colleagues in neuroscience would agree with your point : it is NOT doing the same as the brain, and thus cannot replace it.
Perhaps you've already made a video on this (I found this video in a short time window, I'll check out your other stuff later!) but the biggest weakness of AI art, and indeed the entirety of this new generation of AI is the issue of "flattening" of complexity. As humans, we can attenuate certain stimuli and patterns from our lives and experiences into a stronger stimulus for our endeavors. What I mean by that is that we can choose to study anatomy, or mood, or simply to dedicate time to ponder colors themselves. This then makes it possible to draw upon these experiences and understandings as a foundation to our art in our minds eye, before even putting pen to paper. AI as we know it can not do that. Structures do appear in neutral networks which seem to do an amazing job approximating an understanding of anatomy and proportions - but these are fuzzy things that haven't been attenuated, brought into specific focus, and simply stem from looking at enough pictures of humans. Meanwhile, a talented artist has a clearly defined concept of anatomy - and that concept intermingles with and is founded upon even more layers and concepts, like perspective, their own awareness of their own body, their experiences of touching the human body, and even an understanding of 3D space and an intuitive understanding of solids, rigidity, etc. AI is a linear funnel of information. Inputs go in, and outputs come out. Complex structures may arise within a neural network, but they are a far cry from the human mind, and will continue to be for as long as neural networks are directed, acyclic graphs (something that we don't even have an inkling of an idea of how to change!)
Thanks so much for the technical comment ! I feel like maybe I'm misunderstanding parts of it though ? Let me know. Completely agree about complexity. For now, the space of human possibilities in art is much larger than the space of possible AI outputs. Also completely agree about what you describe that we might call "accidentality" or arbitrariness (I'm in a mood to make up words) : the fact that we acquire skills, ignore others, and develop peculiarities during our art journey, and that this creates our artistic personality. but there are a few reasons why we may not rest too easily on these "flaws" of AI : 1- I recently attended a neuroscience seminar, where the speaker said : In the actual human brain, it's not obvious if we have these explicit representations of, say, anatomy, logic, color etc. It might all be as implicit and "intuitive" as it is in AI. 2- the universal approximation theorem : it's been mathematically proven that artificial neural networks, given enough scale, can approximate ANY mathematical function. If the brain can be modeled with a mathematical function, AI can (theoretically) mimic it. 3- With recurrent neural networks (or the more in vogue term "autoregressive"), it's not obvious that neural networks are acyclic. Say neuron A appears before neuron B. In the second pass, neuron A appears again. This means you have a A->B->A path in the graph.
@@ArtPostAI Hey! Thanks for the response and callout for where I was less-than-accurate! I'm going to preface this by saying that I wrote my original comment with a slight bias towards not causing people existential crises, so I wanted to focus on the hopeful "AI is nothing like humans!" tone. In reality though, I'm actually a pretty strong believer that there is no metaphysical spark that makes our brains in any way more special than a sufficiently sophisticated neural network. The only reason I mention this is to make my own assumptions and biases clearer. 1. I wasn't really expecting people who have an understanding of neuroscience to read what I wrote, so I definitely exaggerated the existence of discrete structures in our brain that understand abstract concepts to make it a little bit more palatable! What I was really alluding to is more related to the really cool perks of generalized intelligence + having a prefrontal cortex - specifically, how we can learn and understand abstract concepts through a huge variety of means and "reward functions", then isolate those concepts in our mental space and use them to compose our art. In a non-technical sense, what I'm referring to is the fact that we can stop painting, go take an anatomy course in college, come back to the easel, and suddenly be better at drawing humans. In a technical sense, I'm referring to the fact that our neural wiring is sophisticated enough that we can develop strong neural connections via one set of training data and reward function, then adapt that neural wiring to do a better job at a task with a completely different reward function and training data. 2. This comes back to my disclaimer, but in short, I agree! I'm terrified of the day it happens, but I think eventually we'll create AI that not only approximates human intelligence, but surpasses it. This next bit is more opinion than academic, but I do suspect that due to the nature of language, LLMs specifically are way way way further ahead than other AI applications such as art, which have a really long way to go before they can really rival the human spark. 3. I hadn't considered that, that's quite interesting! It's a poor man's cyclical graph. I was thinking about the messiness of a biological brain and how there's always a bunch of asynchronous processes all going on in a continuous way, and how different parts of our brains can take both external and internal stimuli as inputs (e.g. the nocebo effect, where a sensation of pain can be created entirely through thoughts which is indistinguishable from actual pain). I don't think our current models approximate anything like this to my knowledge - but admittedly, I actually don't know enough about the cutting edge from the last year or two to be able to confidently claim that it doesn't. Since you're actively working in AI Consciousness research, you undoubtedly know more about this than I do! Hopefully that clarifies my thoughts a little bit. I apologize for typing so damn much!
@@-Gnarlemagne I completely get it :) I feel like we can never truly know if there is this metaphysical spark in us. Even if we create an AI that has learned perfectly to mimic us in every aspect, there might still be something different about us.. I guess for now we can only speculate. When we get there, we'll have more elements to decide. 1- it's a very interesting remark that we, unlike AI (for now), are able to ground skills in other skills. I hadn't quite though of it that way... Thanks ! 2- it's true that since language is a way to describe the world much more complete than a single image, it's a much more promising approach to attaining intelligence than the AI art models. I'm curious how video models will do. I think video models might lack the abstract abilities of language models, or they might have to develop them in order to model the visual world ? It's all very interesting.. 3- I'm not that updated on the state of the art either -- AI is moving so fast I can't keep up even though it's my job x)
Wether human or AI we need constraints context (equalling at least a bit of a goal) and lots of past normal likely experiences to make creative new not so normal stuff. Though indeed current AI seems to go to less creative, more gloss. Dalle2 seemed more creative then Dalle3. Also lack of multimodal meta thinking paired with too stong bias to training data leads to things like utter inability to properly imagine microgravity in large non space environments.
Multimodal meta thinking may give AI dynamically changing metrics for dynamically changing goals eventually matching more closely what humans do. Like understanding physics models autonomously applying gravity vector field maps for weird gravity fantasies. Also autonomous inpaint editing. Not understanding negatives like "ramen bowl without chopsticks" seems lower level though.
Somebody else pointed out that the ramen example was wrong. You can find the right prompt to give to the AI, to get that image. I'm coming to think that your perspective is right. Adding more to the AI, having correctly unbiased datasets, more modalities, will make it more flexible and better able to play around with the rules it learns. Yet I find it hard to conceive continuous learning through dynamically changing optimization goals -- usually, even with things like RL, learning is done separately from inference, and it converges. Training can go wrong and it is also super costly. Moreover, you can't come up with a benchmark for continuously changing goals, and no one is interested in doing research that you can't evaluate and publish. So that the fixed objective still prevents AI from being creative on its own (it is dead and stagnant), but given what you say, people will be able to express more and more creativity through these AI models. Yet I think it'll still leave its mark (someone else pointed this out too) -- because so much of the process is outsourced to the AI, there will be a big similarity between different AI artists, that the general public will grow to detect.
@@ArtPostAI - Using AI image gen tools quite a bit I totally agree with that current and foreseeable image gen AI has big shortcomings in the range of expressible ideas (while simultaneously tremendously expanding expressibility of ideas for people that couldn't visually express them before at all). And I very much agree that stopping drawing due to image AI becoming a thing is a rather stupid and bad decision. AI has much room for improvement. IDK how fast it do will from here. Muldimodality is so darn important and still barely explored. Missing beyond that is gradual incremental learning and a long time memory that's not oddly tacked on. May come eventually. IDK when. We as humans especially when young and unexperienced run into dead ends too. Writers block, or blank paper block. Inspiration does not magically strike by waiting and praying to a Muse. We need to go out and see new parts of the world (or read papers from an other field) to get unstuck (or out of a honed in boring getting old artsytle). We need to take a step back to the meta level to see other trees of the forest to dynamically change optimization goals. Going to the philosophical extreme when stuck on the meta level going to the meta level of to the meta level. Perhaps we as humans are ultimately limited by how far we can step back to see more of natures beauty limited by financial and cognitive means we've been granted. A fascinating philosophical topic where new ideas ultimately come from. Current AI lacks the agency and capabilities to go out and seeking new meta understanding on its own accord. Plus AIs it's innards are purely deterministic (exact reproducability by giving same seed). I don't want to accept humans mental innards work the same. It scares me in a philosophical sense that would go to far here to explain. Well, basically is shrinks the multiverse by an uber astronomic factor in number of experiencable futures. Sorry, I digress. Reducing the number of possible outputs (colors^pixelnumber) to the number of possible seeds (symbols^symbolnumber). OUCH!
You mentioned you work in AI research. What is your focus area of your research? Another point you made is that "ChatGPT can't generate a bowl of Ramen without chopsticks" to illustrate a point about convergence. This doesn't work on my end. Both ChatGPT and Midjourney are able to generate images of Ramen without chopsticks via a one-shot prompt. Literally just "Ramen served alone". You also mention that evolution is creative, but that doesn't seem right either. Evolution being the name of a long term process in which entities adapt to the environment through their generations. This is the focus of evolutionary algorithms and reinforcement learning fields. The most standout example would be the Alpha series from DeepMind. They partly used reinforcement learning to create AlphaGo, which was described as "creative" by Lee Sedol and other professional Go players back in 2016. Now, to be fair, current image generation models aren't doing anything like this outside of RLHF. They also aren't their own autonomous entities like we are. They're single pass inference models done based on human input prompts. This puts the "creativity" on the prompt provider. You can say it isn't skillful, but the prompt is the creative bit. Through using an AI model, they trade off precision for speed. They can kinda get what they want from the model, but to create something with high precision, they would need to manually create it with finer precision tools (i.e., paint brush, pencil, etc.) You also do this thing when comparing AI to humans. It is that when it comes to AI, you go into very granular detail about how they operate. Then for the humans, you become very abstract by comparison. The point where you said that AI have an objective function is different from humans. Reason being is that you can ask humans the meaning of their life is, and they cannot say. It's an apples to oranges comparison that doesn't work in my opinion. For instance, humans have stratified levels of goals that are constantly changing in priority. Humans have the goal to survive, which motivates us to eat, fit into our communities, sleep, and use our bodies. Then we have goals that are seemingly abstract from our physical bodies; trying to get a particular job, wanting to watch a particular TV show, and etc. The ending point was kinda disappointing because there are very clear reasons why humans would still prefer human art, even if a machine with sentience could be more creative and skilled at art making.
Thanks a lot for the detailed comment ! I work on topics surrounding consciousness in AI. "evolution is creative, but that doesn't seem right either." Evolution gets us from molecules, to fish to humans and all other forms of life. To me, that's a creative process. Outside of image generation, as you mention, there are ways of training AIs that try to mimic evolution. For instance with competition in AlphaGo or evolutionary algorithms (which don't mimic evolution well since they converge and evolution doesn't) The competitive approach has been tried in image generation (Generative Adversarial Networks) and it doesn't lead to creativity. "It's an apples to oranges comparison" That's kind of the point. AI has a single, perfectly defined, mathematically formulated goal that it chases without any other distraction. on the human side, we "have stratified levels of goals that are constantly changing in priority". And we don't even know if we should call them goals, instincts, drives, desires or something else. We don't even know if it's correct to think of them as goals. We don't know how many there are. You're right about the ramen thing, I guess it was a wrong example since it's so easily countered. For me typing "ramen without chopsticks" invariably gives me chopsticks in the image. But you're right.
@@ArtPostAI "Evolution gets us from molecules, to fish to humans and all other forms of life. To me, that's a creative process." I guess it depends on what you define as creative / creativity. I think of creativity as something that a living entity does rather than creativity being a direct byproduct of an environment. This feels more intuitive because we can look at a beautiful landscape, we wouldn't think that the environment creatively crafted it. It's just the natural state of our ecosystem at work. "For instance with competition in AlphaGo or evolutionary algorithms (which don't mimic evolution well since they converge and evolution doesn't" I think the process of evolution is being followed correctly. An entity is dropped into an environment for it to exist in. Then through its generations, it adapts itself to this environment. This is why I believe the process is correct, but the staging ground for AI isn't the same as our human environment. It converges because the environment and its own "self" are simple. Like how cells and bacteria are very specialized and generally perform only a small set of actions. At this point I am knit picking technical details, but I guess my sole point of contention that lead me to those issues is that I disagree with your idea that artists are safe because AI isn't creative. I feel like it wouldn't matter even if AI was creative. Art is a form of communication. Humans crave this connection. Be it from drawing a picture, carving a sculpture, doing a dance, or anything else means something coming from humans we can share a bond with. It'll be interesting if AI can be creative. It would be a way to engage with it on a new level. But this doesn't replace human creativity in the slightest.
@@InnsmouthAdmiral Thank you so much for the nuanced and insightful response. I think you've got some great points about there being more than "mere" creativity in human art. Very interesting point about cells and bacteria as well. I'll keep that in mind :)
I love this video, it’s a great way to lay out the issue to the non-educated. I’m only a hobbyist artist myself, but I am very interested in the societal impact AI will have. A little thought I had about your point of the ouroboros-ing of AI art: Besides the capitalistic reasons for not doing so, couldn’t a machine running without a goal or restraints eventually find a niche in our world, as it encounters real-life limitations? If we could prompt it to fulfill an art-niche, couldn’t it have the same existential anxieties that you claim divides human art from AI art? Also, do you think people without such anxieties are capable of creating art?
Fantastic question to ponder. I guess it's the logical next step when we say "AI researchers are scared of not chasing objectives" : what would happen if they weren't ? If we did manage to run a machine in the real world without an objective, goals or restraints, I feel like that might take out a lot of AI's flaws. Probably not all, but we'd get a different animal than the one we're currently dealing with for sure. "If we could prompt it to fulfill an art-niche, couldn’t it have the same existential anxieties that you claim divides human art from AI art? " How do you prompt it ? With an objective function ? If so, then I don't think it can get these anxieties. Really, we don't know how to add emotions into AI, further talk would be speculatory I think. do you think people without such anxieties are capable of creating art? I don't think there are people without existential anxieties. But apart from that, it would be arrogant to try and say that certain people cannot make art. People always surprise us.
I find this subject very interesting, thanks a lot 😊 but what about the idea that our brains are designed to make us behave in a way that increases the likelihood of our survival or some criterion like that, I think that human brains works in a pretty homogeneous way actually, we kinda converged towards something It's hard to say where our creativity comes from, but if it emerges from the level of intelligence we have in our ability to represent our environment and interact with it, then convergence itself (aka using some sort of likelihood) won't necessarily prevent AIs from developing creativity. I guess the thing restraining modern AI to be creative would be the lack of generality in what they converge towards, maybe they'll get creative only if image generation comes as a by-product of the model's abilities
What a fantastic comment ! :) We could argue that evolution is just an optimization process aimed at optimizing reproduction likelihood. But it is a different optimization than mere gradient descent, and is so much richer that it's not obvious if it's a quantitative or qualitative difference. I guess only time will tell. Biologically it feels like human brains have converged because we look at it at our timescale. But if you check back in in a million years, you'll see something completely different. What is creatively changing at our timescale is culture. Culturally we aren't converging at all. You could imagine a kind of internet overwhelmed with AI images, where they learn from each other and develop a sort of flowing, constantly changing and reacting (and thus potentially creative) culture. But the evidence we have so far for AI having this ability is not promising. Generative adversarial networks do converge even though they compete forever. Evolutionary algorithms also converge. I think creativity is to be looked for outside of representation capacity. And thus you hit the nail on the head by saying that you need to train them for something else than image generation, to get creative image generation. I'll keep that in mind actually.
but in the industry most artist has a clear goal set by client or company, experimenting is just a small part of the process (exellent video btw finally an answer rather than just saying adapt to it)
Yes, since I'm also trying to make enjoyable art in the future, and I cannot deny the reality of present and incoming AI, I want to get into details about how to adapt to it.. Industry art follows an objective, but that objective is always much more loosely stated than the perfect, mathematically-defined objective of AI art. There are so many constraints on objective functions in AI, that even with all the research that's gone into it for over ten years, basically only two effective objective functions exist for generative AI : -Maximum likelihood estimation, -the GAN-style objective (which doesn't work as well). Whereas objectives in industry can be set by people who are themselves quite free and creative. Would you agree with this view ?
@@ArtPostAI I think it all comes down to the companies. Are they willing to spend more for creativity or do they prefer mediocre and cheap production.
What is this song in the end? Please don't tell me it is AI generated. The voice is similar to Ella Roberts voice...but is too in tune like it was auto tuned. If you think a lot of people thought Hatsune Miku was a person not a Vocaloid (Vocaloid is not human but it does not use AI) and many good Japanese composers began writing for it and using it like Yoneszu Kenshi and Ado...well everything is possible. UPDATE: I found the video where you said it was AI generated...ooooh the inunmanity! Machine men (and women) with machine minds and machine hearts...and voices.
“Any risks taken by moving out of the boring and typical, and generating a somewhat creative image is immediately understood by the model as a bad thing” Wow, that sounds like fundamentalist conservatives Thank you for the in-depth, insider perspective. It does make me feel better as an artist. As a professional artist, I still worry about companies who care more about money than creativity. But that’s a different story
Thank you so much :) Yes that's a different story. I'm curious if you have any details on how your experience of AI is as a professional artist ? Which kinds of people support it (I'm thinking ceos and executives ?) How do people feel about it ? How fast is it being adopted ? I feel like a lot of companies are being slow and deliberate with integrating AI because they're aware that it's a hype-ridden field. How fast is adoption ?
@ArtPostAI Hi hi, sorry, I completely forgot to reply to this Several companies have gotten into hot water for using, or supposedly using ai. Netflix released an anime where all the backgrounds were created with ai, and they didn’t even credit the person who generated it. In the credits, they are simply listed as “AI + Human.” Magic the Gathering got caught using ai, as well as the drawing tablet company, Wacom. Marvel’s been accused of using ai for at least one of its covers, though I’m not sure if that turned out to be real. As it stands, in the US, ai is not protected under copyright, because only humans are allowed to hold copyright (oddly, thanks to the peta monkey selfie. One of the only good things to come out of peta). However, ai is copyrightable in China, so I wouldn’t be surprised if we see a slew of ai produced media coming out of china in the next few years. Recently, Adobe has recently been extremely shady with their ToS, by forcing everyone who uses their software to give over the rights to all of their files. The thing is that this breaks NDA contracts. Adobe says it’s not a big deal because it’s for scraping data for their ai. But that does nothing to protect designs from being reproduced before a project is released. And with Adobe being the industry standard in entertainment, that has the potential to cause a lot of problems. And on the voice side of generative ai, open ai got into massive trouble for stealing Scarlet Johansson’s voice. Voice actors specifically got a garbage deal from the writers’ strike. Johansson is lucky she is famous enough to speak out about it, but the majority of people are not so lucky. We are going to see even more voices stolen and more jobs lost in the future. So yeah, this mass push for ai, being forced upon industry creatives is extremely detrimental to the entertainment industry :/
@@naiveghost9683 I don't think so, but maybe. What i think makes art valuable is not what is put into it; what makes art valuable is what comes out of it. An artist can put a lot of time, effort, and emotion into a piece of art, and yet have little to no effect on anyone. On the other hand another artist can draw one line in a unique way on a blank piece of paper in seconds or minutes that then goes on to have a very significant effect on people. Art is about inspiration, not hard work necessarily.
@@naiveghost9683 Certainly my opinion too. Seeing art such as Yuri Norstein's Tale of Tales, it is hard to argue that the sheer willpower of an artist invariably adds magic to a finished piece.
@@alexgonzo5508 Very interesting take. Art can be quick or slow, and each one has its advantages. Harding was an advocate for quick art for instance. I believe Pre-raphaelites, Hudson school river painters or classicists might have disagreed. As you mention, inspiration is separate from the amount of hard work. One might have inspiration for a new way to create art, that happens to be fast. Or, one might have the kind of inspiration that drives slow, deliberate and loving study of nature. Both are as valuable. But I believe that the awareness of the artist's emotion and sheer effort is part of the appeal of more elaborate works. All the while the appeal of more spontaneous art, is the mystery of chance and intuition.
Is a satellite image of some field in Iowa art? Without emotion AI will never be able to make art art doesn't just require creativity but it requires emotion add a more fundamental level to feed that creativity No matter how incredible a picture an AI generates maybe it is but a satellite image of a field in Iowa it will never be art
To me, art is communication. Taking an unconventional route to express things that could never be expressed or accepted if seen from the typical angle. Bypassing judgement by changing the medium. As such, AI art seems like a restraint on the artist's ability to communicate. Because of all its flaws. But each will have their own definition of art. Defining art, feels like a lost cause at this point right ? Ever since Duchamp's urinal, it feels impossible. So I guess your comment keenly his taps into the tough question of the value of art. What is art worth for the maker ? Making AI art is a cold, digital process, much less enticing than the journey of human art. For the viewer ? I think AI art and human art will both have their place in satisfying the viewer in the future.
I don't think AI art has meaning or value. Sure, they can make beautiful pictures, but the value of a piece of art for me isn't necessarily the art itself. It's the process someone underwent to make it. Creating is what's meaningful, and typing a prompt into a machine to get a pretty picture isn't creating. It's really not that different from googling a picture. Doesn't have as much impact.
"AI" and "Never" do not go together well. I guarantee you, there will be an art generator one day that is even more creative than humans. All we need is an LLM (which has already been proven to have creative abilities that are even greater than some humans) with an image generator built-in. Oh wait, that's exactly what GPT-4o is, and its image generator puts all previous image generators to shame, none of you have seen it yet because it was intentionally undermarketed. I get you're scared that you won't be able to monetize your art, but you will never be stopped from doing art, you'll just have to get a different job. Human art will always have value in the fact that it was human-made, it just isn't going to be very profitable. These AI tools give everyone the ability to do what you do, and even better. So many people have great ideas but no way to bring them to life, this gives them the power to create what they want with minimal effort. To me, that sounds pretty damn good, and if you think that we shouldn't have this ability just so you can profit off it, I don't think you deserve the money.
Thanks for the comment, it really helps me think :) I work in AI and use GPT-4o every day. It's doesn't feel any smarter than GPT4, it's just a little faster. And it certainly isn't creative, partly for reasons I cover in this video. GPT4 was already able to "see" and generate images. And having used both image generators, I still like Midjourney a lot more. Just another big marketing thing from OpenAI with everyone buying into the hype. The reason we think AI is creative is because we have completely abandoned the true meaning of creativity. In the scientific age, we have come to believe that creativity is just cleverness or rapidity of thought. That's not what it is. Writers I talk to say that GPT4 prose is obvious and ugly. Artists don't like the "AI look" that gets burned into images. You have to control it through text, which lacks the precision required for true expression. It requires very little practice, which takes away the process of mastery and all the richness it brings to the mind. It lacks the soul-filling enjoyability of real, disciplined art, which encourages you on a journey to a completely different way of seeing the world. It disconnects you from Nature (not just forests but the entirety of the world), which is mostly where artists get their inspiration It takes you out of your body, which is where intuition and immediacy are (talk to any painter or sculptor and they will tell you about the body) I have said much more on my channel, and will continue to do so I'm not saying that AI shouldn't exist because people want to make money from their art. I don't think AI will take that away, in part because of the reasons above. But the argument that AI art is a good to society is hard to defend. For one, people will get used to it and it'll become just another form of banal communication. People won't be creating art with it for the most part, because having access to Midjourney simply doesn't make you an artist, just like having cameras didn't make everyone artists. So let's assume that a multimodal LLM is all that's required to put artists out of their jobs. What people will make with it will be reminiscent of the uses people found for photography : sharing things with their friends, etc. A select few will make actual art with AI art models. They will find disruptive truths and communicate them eloquently through these glorified rendering engines. They will capture a new feeling, a new aspect of life, and make it flow through their images. But for the most part it'll just be a tool for expressing things that could be said through text, photos or even little sketches that everyone can do. In this (unlikely) situation, then we would have taken out the deepest, most exploratory form of expression that humans have (art) and replaced it with yet another way of making holiday pictures, but for imaginary holidays. With just a few people actually making something worthwhile out of it. Tldr : Trying to claim that AI will get creative is the wrong way of looking at things. Humans will always be behind the computer, prompting these things in one way or another. AI simply offers a new form of rendering. The question is : will this new form of rendering take over all other forms of art ? IMO, no. For many reasons I cover on the channel.
Nah mate you don't get it do you. They are going to train a text model on a prompt database of hand-picked pack of users who get great results from AI and feed it into an image model. They'll call it "creativity" and sell for profit, accelerating the commodification of real creativity even more. I don't get why they attack us so relentlessly. Maybe someone didn't manage to get into art school again. Maybe an artist slept with his wife. Something along those lines.
I do think there is rampant envy against artists in our society. Many people wanted to live the artist's life but didn't dare to try. Now those who are unhappy about their lives have a way to get back at those who make a living in a fulfilling manner
@@ArtPostAI Well, I believe creating AI is quite a fulfilling endeavor, somewhere in the middle between art, engineering and business. Art is theft, in essence, so stealing from ALL of the artists is quite a fun idea. Nonetheless, this looks more like a personal grudge, maybe contempt, not envy. That's what I'm trying to say))
This could hold water if it weren’t for the point about optimization. The AI is specifically supposed to be making images as similar to its references as possible. In this framework, the AI would be incentivized to mix images as little as possible. In the case it was presented with a prompt it didn’t have a reference for, it will simply default to the next best thing. For these reasons, the AI image generators we see today will become increasingly “perfect,” removing the conceptual value.
I've tried out a bunch of portable note-taking apps, and Napkin is by far my favorite because it's fun.
Check out Napkin :
napkin.one/signup?via=ArtPostAI
I like the technical stuff. I actually think that is what makes this channel unique. Pro human artist who also understands how AI works. Keep it up!
Great ! thanks so much !
This has already happened with ai writing. Four years ago I tried out the earlier models, and while the storylines were barely coherent they felt weirdly original and fun. Now when I ask chat-gpt to make a story, it's coherent but it takes any cool concept I give it and makes the most generic version of it.
Absolutely right. It's also described by frequent AI art model users.
1 It definitely will not replace artists with a unique style, or artists that have developed a very high level of skill or design knowledge. It also can't do fine art (but can mimic it digitally to a degree at least).
2 It likely will replace artists who made relatively generic art, like the kind you might find in many mobile games. There are still huge limitations, I don't think any models can generate vector graphics for one.
3 one of the greatest utilities of AI is using local, highly customizable models like Stable Diffusion. The output of such models can be used as part of an artists workflow, or fine tuned to produce images that are much more specific and customizable than models like midjourney or dalle can output. Some of the workflows for Stable Diffusion are fairly technical and labor intensive to a degree which is actually pretty cool. It's still much different than having direct control over what you're creating but I don't think art requires that. But a lot of these local models, or at least the popular configurations of them have certain issues like saturation being way too high, perspective being off, etc. Which again points back to the value of being an artist and knowing design.
All in all I don't think mostbartists should fear AI at the moment. There's many people saying they believe many of these models are starting to plateau for one. And besides that some artists will learn to use AI themselves or it will drive them to carve out a niche that AI can't replicate well or at all. If art is your passion you can't let technology discourage you from creating.
Your conclusion is exactly right in my opinion.
1- it can mimic fine art to the degree that normal people can't tell the difference. But it'll have big trouble trying to mimic the style of living, breathing artists who (my guess) are going to get much more exploratory with their style if people start training AI models specifically to target their style.
2- Completely agree. And that's pretty sad because these kinds of jobs are often a gateway into creating more worthwhile art and building valuable careers.
3- Completely true. And I think it'll get more and more customizable in the future. Just like AI researchers of my generation aren't diving into the nitty gritty of optimization algorithms or GPU optimization and have more time for architecture, dataset and loss function design, those of the next generation will be focusing on things more accessible to the public.
Great video! I like how you used how AI actually works as evidence to support your claims. Lots of people just see the line go up, but don't know what that line represents (including me) so this is a really valuable approach!
Thank you so much ! I'll take that into account :)
This is one of the best presented arguements on the limitations of AI Art I have yet seen. I had not considered that as a 'problem solving' technology AI is not really suited to being creative.
In a sense AI Art is the antithesis of creativity because it's not really trying to communicate anything- and all creativity is about the desire to express something; an idea, a feeling or even just a particular mood or moment- AI's have no such aspirations, they really are just goal orientated machines trying to generate the optimal solution to the problem presented to them by the prompt input by their operator, with a degree of randomness thrown in.
Which is not deny that some AI generated images are not genuinely impressive and strikingly novel- but even the most impressive AI Art is still lacking in the essential thing that distiguishes Art from mere pattern making. Give a monkey some paints and brushes and they may create some vivid and striking marks on a canvas- but we would not claim that these marks are works of art because we know that the monkey had no idea what it was doing- it was not making those marks in order to communicate with us, or to make us feel something, and lacking this intentionality it's mark making has no meaning- they are just marks on a canvas.
And while it's true that the patterns created by AI are far more sophisticated than the marks made by the monkey, they too are not art because- just like the monkey- the AI was not trying to communicate anything- it was just generating patterns in reponse to the prompts input by it's user.
A recent experiment was done to see if people could tell the difference between AI generated faces and photo's of real people. The result was interesting because while the test subjects claimed that they could not tell the difference their brains were seen on EEG to react differently to the AI generated images- so on some unconscious level the test subjects were able to identify which faces were real and which were AI.
The reason for this is because humans are very good pattern detectors in their own right- and there were patterns in the AI generated faces that had registered on the brains of the test subjects on some subliminal level. OK- so if the ability to detect patterns in data is in part a function of sample size what happens when people become exposed to huge numbers of AI generated images? It seems likely that as we see more and more AI Art we will become more and more adept at recognizing it, to the point where AI Art will become a Genre of Art in it's own right- already the 'AI' look is something that many people recognize. It's subtle- AI can mimic many styles and mediums, so in theory it should not be possible to distingish AI Art from human made art- but in practice this is not usually the case- the more AI Art you see, the more your brain is 'trained' to detect it's presence.
The reason for this is kind of obvious. If I use a tool to make a thing, that thing will inevitably bear the imprint of the tool I used- the artefact created will be an expression of the tool used to create it. And indeed among AI Artists there are lively debates as to the relative virtues of different tools- is Midjourney a better tool than Dall-e for example?. These debates only make sense if it can be said that both these tools leave an imprint on their creations that is identifiable in some way. That being the case what happens if millions of AI Artists use Midjourney to make their Art?
The answer is that they will converge- their creations- all bearing the imprint of the tool they are using- will begin to look similar to each other. OK- they might argue- can the same thing not be said of a paintbrush or a sculptors chisel? Well, not really, because there is a huge difference between the input a paintbrush has on the final image and the impact that Midjourney has on the final image. In the case of the paintbrush the behaviour of the tool is massively influenced by the hand that holds the brush- because each artist is a unique person with different skills, aspirations and experiences. In the case of Midjourney on the other hand the influence of the operator is very much less- indeed Midjourney does not even need a human operator- we could simply connect a chatbot like GPT and have it generate random prompts that Midjourney could turn into images with no human artist involved at all- just one AI prompting another AI.
So the endgame of AI Art is that of a rising tide of images that all exhibit similar artefacts as result of being the products of the same tool- and as that tide rises those artefacts will become more and more obvious as our brains learn to recognise them. At some point the 'AI Look' will become so obvious that the use of AI Art will become the hallmark of art that is cheap and low effort and any commercial application of AI Art will be seen as a lazy and inferior solution. For this reason the long term impact of AI Art on the market for Art may not be as great as many people expect.
Thank you for this comment -- it's a masterpiece !
"AI Art is the antithesis of creativity because it's not really trying to communicate anything"
"even the most impressive AI Art is still lacking in the essential thing that distiguishes Art from mere pattern making"
I'm taking notes !
Do you have the reference to that EEG experiment ? It sounds super interesting !!
"It seems likely that as we see more and more AI Art we will become more and more adept at recognizing it, to the point where AI Art will become a Genre of Art in it's own right"
As you say ! I had the intuition that something like this was about to happen but didn't have any scientific references to back it up !
It's also very true that using DALLE or Midjourney isn't like using a brush or a chisel. It has much more effect than these simple, unmoving tools.
And the way they are made gives very little control over them, even for AI researchers. If you want the best quality, you have to use all the training data, and the best architectures and optimizers. Trying to come up with something new, you sacrifice quality.
Although, I think that new ways for AI to make art are coming. To me, it's pretty clear that with improvements in Reinforcement Learning (and maybe even LLMs, they do anything with them these days), there'll eventually be computers that draw at a professional level by actually making lines on a physically simulated canvas.
I feel that at that level, there'll be quite a lot more wiggle room for changing the "style" of the AI-painters (changing the kind of hand the robot controls, the tool it's weilding, augmenting gravity...).
But that's pure speculation, and even then, since it is the AI itself drawing, it might be that the limited number of styles we can get it to use will all be detected by humans after a while.
So your point would still stand for these too.
Thanks again for the super insightful comment !
@@ArtPostAI Thanks for the kind words. I don't have a reference sadly for that experiment- I read it online about three months ago and it stuck in my mind. There is a test you can take for yourself though; go to Artstation and look at the gallery- which is mostly human made art- then go to Deviant Art and look at their AI Art gallery- if you flip between them it's clear that both galleries have a distinct and different aesthetic- despite the fact that the images in the AI gallery were literally trained on Artstation artwork ( without the permission of the artists of course.)
The strange thing is that if you asked me to explain that difference in words I would struggle to define it, but it's there- somehow the AI gallery has a specific 'feel' despite the fact that it contains a diverse range of styles and subjects- and the same is also true of the Artstation gallery- both somehow reveal their provenence as being either human made or machine made- which is I think reflective of the fact that the the visual 'DNA' of AI Art is somehow exposed in the images it creates- and as humans we are capable of detecting this visual 'signature', especially when we see a lot of AI images gathered together.
In a way this should not be really surprising- any process or tool must surely leave it's mark on the things it makes- this is true of a flint axe and is equally true of an AI programme in which specific training data is used and specific algorithms are deployed- the idea that AI Art machines are totally transparent and unbiased in the way they respond to user's prompts is really a kind of magical thinking.
Perhaps one day AI's really will be able to read our minds and create perfect representations of the images we want - but todays AI Art machines are far more crude- first they must translate our words into images- an imperfect process in itself- then they execute their imperfect understanding of our intent using a specific and fixed array of trained weights which are themselves shaped and constrained by the specific algorithms used to run the programme. It would be genuinely surprising if this cumbersome process did not leave distict visual 'fingerprints' on the images it made.
My view is that at some point AI Art will reach 'critical mass' in the sense that it will come to be seen as a genre in it's own right. This will not invalidate it as a solution commercially-but it will create a problem for anyone wishing to 'stand out' in a marketplace saturated with AI imagary. I've already heard from some artists that their clients want them to use AI (for reasons of cost saving) but are specifically asking them to avoid that 'AI look'- how this contradictory feat is to be achived no one seems to know, but it's interesting that already the term 'AI Look' has been coined and is in circulation among some who commission commercial art.
This is truly a amazing comment! It is very insightful.
Great video! Thank you!!
Thanks !!
Do you think it would be possible for the set goal or optimization measure of “creative” AI to be a high ‘score for creativity,’ rated subjectively over a large population of humans? And then once AI knows how to create artwork that humans find creative, it could use that same algorithm to create “creatively”?
Thank you so much for the insightful question ! It makes me think...
I'm surprised I haven't heard of such an approach in AI. I'll have to keep an eye out.
I have two objections that come to mind immediately though :
1- There's a difference between what people call creative, and what actually is. Though there is some overlap, there are many eminently creative people who are dismissed in their lifetimes, only to be recognized later on (or not at all).
2- If you train this AI at a given point in time, it will learn to produce things that are judged to be creative right now. But what is creative today is passé tomorrow. So you'll need to ask people again, and train the AI again.
3- No art is judged creative by everyone. The most creative outputs from the AI are likely to cause division among the raters, which means when you average them together to create the final objective function, you get an average creativity value.
I'm just guessing though.
Do you agree ? Did I misunderstand the question ?
^ What he said, but also, remember that basically every quality you want to add to a model becomes just another label. "Creativity" can easily be trained as Label 20876 in a model, when you have pieces of data you've labeled as "creative". But when it comes time for the AI to perform on its test data, labeling something with a label you've already been given isn't very creative, is it? We don't want the model to be creative in a way we simply label as "creative" in our toy model: we want it to be creative in a more meta way. We want it to think outside-of-the-box, in a way a computer fundamentally can't.
If we lose our phone in our house, we know that if we can't find it after searching for a while, it's more likely that a family member accidentally took it. Maybe the dog got it. Maybe we didn't actually lose it in the house after all. Posed with the same problem, an AI won't imagine that a family member might have it. It won't consider that maybe the dog buried it in the yard. It won't consider that maybe we just left our phone in the car. The AI will look behind the stove. Under the chest in the attic. On the ceiling. It cannot comprehend that there are other possible solutions, unless we explicitly tell it that there are. This same thing applies to all other aspects of life, and all aspects of *those* aspects, ad infinitum.
The background music detracts from the experience, it shouldn't be there, or at least it should be quieter
I see... thanks for the feedback !
Hey, awesome video! Here are some questions that popped up while watching, hope you'll provide your insight.
a) Considering that the Likelihood is being optimized, wouldn't a large enough training data set result in solutions that we perceive as creative?
I see how a 'small' training data set would lead to boring solutions, though wouldnt a large enough data set allow for a larger set of valid solutions -> More variations -> More 'creativity'?
Since regardless of whether it actually is creative or not, I'd argue that the chance for a human to create new art that bears no similarities
to a piece of art already present in the internet to be miniscule.
b) Secondly, aren't humans also in a way heavily influenced by their "training data", as in their experiences throughout life? I do wonder how different that is in comparison to AI.
"result in solutions that we perceive as creative" It already happens, though 🤷♀
The way humans create is by understanding things we've seen before and extrapolating on it. So, in essence, it's true that with enough data, you could recreate "the human condition". But it's a gross oversimplification of how realistically we can categorize anything, and just how much data that would entail. It's like "Step 2: Draw the rest of the owl".
Extrapolation is very human. If something does not fit within the bounds of what we've seen, we can extrapolate out into a constrained set of possible answers. We know *when* it's important to extrapolate, as well. An AI can't tell without being explicitly "told" anything: it only interpolates. Say that we have an item A: it looks like a duck, swims like a duck, and quacks like a duck. Humans know it is a duck. AI knows it is a duck. That's good! Say that we instead have item B: it looks like a duck, swims like a dolphin, and neighs like a horse. Humans know that this is some amalgamation of random animal traits that doesn't actually exist: a chimera. AI knows it is a duck. That's wrong (or at least, misleading). It can only interpolate based on prior data it was been fed. If could "tell" the AI that there is a secret category it can use to label unknown animals, but then we have to label *those* for training the AI... and so on and so on. We can, in theory, model extrapolation as interpolation if we have a truly mind-boggling amount of data and the perfect algorithm. But I could, in theory, buy everything on the planet if I had all the money in the world, too.
Coming up with (accurate) labels for unseen things on-the-fly is closer to "true" AI. That's the ability to extrapolate as humans do. AI interpolate by design.
Tough questions !
Thank you so much :)
a) "Considering that the Likelihood is being optimized, wouldn't a large enough training data set result in solutions that we perceive as creative?"
-> This is an open question in AI. What's known as "the bitter lesson" is that compared to the benefits of scaling up the data and nb of parameters, all the effort we researchers put into designing better models is basically useless.
Therefore, is all that's required for creative AI not changing the objective, but instead just adding more data ?
Actually I hadn't thought of it, but it seems pretty likely to me that with scaling up the training data, AI will become more and more flexible (thus, maybe you'll be able to get the proverbial ramen bowl without the chopsticks).
As such it will allow people to be more and more creative through it.
So you have a great point here actually.
As far as making the AI itself creative, there is still a prompter behind the AI. But with the comments on this video, I'm kind of reconsidering my use of the word "creativity".
b) "aren't humans also in a way heavily influenced by their "training data", as in their experiences throughout life? "
-> to me, it is different in that we don't have as a unique and absolute objective to reproduce that training data.
Does that answer it ?
likely does not equal typical at all.
you can literally find the word "typical" within the space of meaning that the AI operates with, and therefore you can make its "tunnel vision" aim at what is the opposite of typical.
you can also, in principle, use this as a mere basis for the overall "creative process", by making an ai search for very un-typical associations between words, phrases etc., to create more un-typical prompts, and thereby create more un-typical pictures.
he adresses that actually, his point is that making models aim to be not typicall makes their outputs more random. but radomness here does not mean exotic outputs that are still coherent which is what we would like, it leans more towards white noise and give outputs with little value
@@question_mark "a bird that is as far from a typical bird as it gets" might not be the easiest prompt to make use of in the intened sense for an ai, but there is certainly a phrasing of this concept that could be used as a prompt, which would certainly not just lead to irrelevant randomness. but to an "exotic" (non-prototypical) bird - plausibly to a penguin, for example.
just as human creativity benefits from (plausibly necessitates) limitations, (like refering not just to any non-typical concept, but to a non-typical bird, for example) so would ai-creativity.
in fact i think it is also very implausible to conceive of creativity as independant of goals.
even artists making art "just for the sake of it" do so out of more or less conscious drives. and any drive a human might have can be used as a blueprint for some type of training-goal for an ai.
@@davejacob5208 For the arbitrary prompt, yes but I guess then there is not really something analog to our creativity coming from the AI's side
And I also think that creativity does not imply not having some kind of goal, I even made a comment on this video explaining why I think so x)
the creative force here is still technically the person writing the prompts, as the generative model itself cannot reason, and it does not "know" its own tunnel vision exists. it would act as it always does, a brute force hammer approach to creation
@@question_mark i did not really understand your first paragraph, could you please elaborate?
This video is quite good.
However, I find challenging this idea of 'creativity'.
How do we test for 'creativity'? How do we set a null hypothesis to disconfirm it?
What assumptions about modern AI are we making?
It seems like perhaps a mythical goal, rather than a practical one (AGI anyone?)
I suspect when we say a person is creative,
we really mean that there is a sense of perceivable 'vision' which went into something.
It's hard to talk about a certain person being more creative than another.
It altogether depends on the audience perceiving a sense of salience, a sense that there is a signal worth fishing from the noise.
Is the banana taped to a wall creative?
Very keen observation.
Seems you know quite a bit about AI !
That's actually one of the things I'll discuss somewhere in the next few videos
is creativity a word like AGI (I completely agree with you on that one), time, or even "consciousness" ?
Something we keep talking about, but can never grasp ?
Maybe I should look for another word.
We're all creative, but it varies from person to person. One person might be creative in art, another on music, or cooking or building. It's an essential part of being human.
AI can never be truly creative.
IMHO.
@@JaneNewAuthor
Let me put this another way.
Suppose you met some extra-terrestrials. How would you test the creativity of this new species (of intelligent life)?
@@AZTECMAN they would be creative in their own way. In the most general sense, all life is creative. Look at the patterns on a seashell.
When an artist creates they're using their skills, talents, experience and intelligence, but they're also using something indefinable that makes them human. Life experience maybe?
AI will never be able to compete with that, because it's not alive.
I am very reluctant to place my trust in 'never'... but I tend to agree with the attitude.
I happen to believe at the moment that there isn't any alive AI... however I'm not convinced this is impossible.
Are you familiar with John Conway's "game of life"? Cellular automata are quite a bit of fun.
Thank you for sharing details! I feel like there's a lot of people with opinions about AI, but few people who have a strong foundation in both AI research and the domains it impacts, so if you keep sharing deep insights into AI, that would be really valuable :]
Great ! I didn't expect this response, I thought it wouldn't be entertaining to have technical details
@@ArtPostAI it's extremely helpful and encouraging to hear technical details from someone who is both educated and actually cares
The problem with AI's lack of creativity is that it's trying to mimic something that we do so instinctively that we don't even know how we do it in the first place. We're trying to achieve the end by setting up networks that mimic the human brain in structure, but we fundamentally don't understand how the brain makes connections between ideas and how it determines relevancy. If we don't know how that works, how can we possibly program a computer to do it? Because all computers do is follow instructions. Those instructions may be complex algorithms, but they're still just instructions that the machine will follow without deviation.
You're hitting straight at the big question in AI : how do we get these computers to do stuff without having to tell them exactly how ?
So far, the best answer is : define a hard-to-achieve objective, don't specify the steps to get there, and run an optimization to have the neurons figure out the steps on their own.
It works well enough for AI art to be a thing without anyone having manually touched any of the weights in the network.
The fact is : we don't know if it does or if it doesn't do the same things that happen in the brain. Because we don't know how either works.
But from the little neuroscience we know, my colleagues in neuroscience would agree with your point : it is NOT doing the same as the brain, and thus cannot replace it.
Perhaps you've already made a video on this (I found this video in a short time window, I'll check out your other stuff later!) but the biggest weakness of AI art, and indeed the entirety of this new generation of AI is the issue of "flattening" of complexity. As humans, we can attenuate certain stimuli and patterns from our lives and experiences into a stronger stimulus for our endeavors. What I mean by that is that we can choose to study anatomy, or mood, or simply to dedicate time to ponder colors themselves. This then makes it possible to draw upon these experiences and understandings as a foundation to our art in our minds eye, before even putting pen to paper.
AI as we know it can not do that. Structures do appear in neutral networks which seem to do an amazing job approximating an understanding of anatomy and proportions - but these are fuzzy things that haven't been attenuated, brought into specific focus, and simply stem from looking at enough pictures of humans. Meanwhile, a talented artist has a clearly defined concept of anatomy - and that concept intermingles with and is founded upon even more layers and concepts, like perspective, their own awareness of their own body, their experiences of touching the human body, and even an understanding of 3D space and an intuitive understanding of solids, rigidity, etc.
AI is a linear funnel of information. Inputs go in, and outputs come out. Complex structures may arise within a neural network, but they are a far cry from the human mind, and will continue to be for as long as neural networks are directed, acyclic graphs (something that we don't even have an inkling of an idea of how to change!)
Thanks so much for the technical comment !
I feel like maybe I'm misunderstanding parts of it though ? Let me know.
Completely agree about complexity. For now, the space of human possibilities in art is much larger than the space of possible AI outputs.
Also completely agree about what you describe that we might call "accidentality" or arbitrariness (I'm in a mood to make up words) : the fact that we acquire skills, ignore others, and develop peculiarities during our art journey, and that this creates our artistic personality.
but there are a few reasons why we may not rest too easily on these "flaws" of AI :
1- I recently attended a neuroscience seminar, where the speaker said :
In the actual human brain, it's not obvious if we have these explicit representations of, say, anatomy, logic, color etc. It might all be as implicit and "intuitive" as it is in AI.
2- the universal approximation theorem : it's been mathematically proven that artificial neural networks, given enough scale, can approximate ANY mathematical function. If the brain can be modeled with a mathematical function, AI can (theoretically) mimic it.
3- With recurrent neural networks (or the more in vogue term "autoregressive"), it's not obvious that neural networks are acyclic. Say neuron A appears before neuron B. In the second pass, neuron A appears again. This means you have a A->B->A path in the graph.
@@ArtPostAI Hey! Thanks for the response and callout for where I was less-than-accurate!
I'm going to preface this by saying that I wrote my original comment with a slight bias towards not causing people existential crises, so I wanted to focus on the hopeful "AI is nothing like humans!" tone. In reality though, I'm actually a pretty strong believer that there is no metaphysical spark that makes our brains in any way more special than a sufficiently sophisticated neural network. The only reason I mention this is to make my own assumptions and biases clearer.
1. I wasn't really expecting people who have an understanding of neuroscience to read what I wrote, so I definitely exaggerated the existence of discrete structures in our brain that understand abstract concepts to make it a little bit more palatable! What I was really alluding to is more related to the really cool perks of generalized intelligence + having a prefrontal cortex - specifically, how we can learn and understand abstract concepts through a huge variety of means and "reward functions", then isolate those concepts in our mental space and use them to compose our art. In a non-technical sense, what I'm referring to is the fact that we can stop painting, go take an anatomy course in college, come back to the easel, and suddenly be better at drawing humans. In a technical sense, I'm referring to the fact that our neural wiring is sophisticated enough that we can develop strong neural connections via one set of training data and reward function, then adapt that neural wiring to do a better job at a task with a completely different reward function and training data.
2. This comes back to my disclaimer, but in short, I agree! I'm terrified of the day it happens, but I think eventually we'll create AI that not only approximates human intelligence, but surpasses it. This next bit is more opinion than academic, but I do suspect that due to the nature of language, LLMs specifically are way way way further ahead than other AI applications such as art, which have a really long way to go before they can really rival the human spark.
3. I hadn't considered that, that's quite interesting! It's a poor man's cyclical graph. I was thinking about the messiness of a biological brain and how there's always a bunch of asynchronous processes all going on in a continuous way, and how different parts of our brains can take both external and internal stimuli as inputs (e.g. the nocebo effect, where a sensation of pain can be created entirely through thoughts which is indistinguishable from actual pain). I don't think our current models approximate anything like this to my knowledge - but admittedly, I actually don't know enough about the cutting edge from the last year or two to be able to confidently claim that it doesn't. Since you're actively working in AI Consciousness research, you undoubtedly know more about this than I do!
Hopefully that clarifies my thoughts a little bit. I apologize for typing so damn much!
@@-Gnarlemagne
I completely get it :)
I feel like we can never truly know if there is this metaphysical spark in us.
Even if we create an AI that has learned perfectly to mimic us in every aspect, there might still be something different about us..
I guess for now we can only speculate. When we get there, we'll have more elements to decide.
1- it's a very interesting remark that we, unlike AI (for now), are able to ground skills in other skills. I hadn't quite though of it that way... Thanks !
2- it's true that since language is a way to describe the world much more complete than a single image, it's a much more promising approach to attaining intelligence than the AI art models.
I'm curious how video models will do. I think video models might lack the abstract abilities of language models, or they might have to develop them in order to model the visual world ? It's all very interesting..
3- I'm not that updated on the state of the art either -- AI is moving so fast I can't keep up even though it's my job x)
Wether human or AI we need constraints context (equalling at least a bit of a goal) and lots of past normal likely experiences to make creative new not so normal stuff. Though indeed current AI seems to go to less creative, more gloss. Dalle2 seemed more creative then Dalle3.
Also lack of multimodal meta thinking paired with too stong bias to training data leads to things like utter inability to properly imagine microgravity in large non space environments.
Multimodal meta thinking may give AI dynamically changing metrics for dynamically changing goals eventually matching more closely what humans do.
Like understanding physics models autonomously applying gravity vector field maps for weird gravity fantasies. Also autonomous inpaint editing.
Not understanding negatives like "ramen bowl without chopsticks" seems lower level though.
Somebody else pointed out that the ramen example was wrong. You can find the right prompt to give to the AI, to get that image.
I'm coming to think that your perspective is right.
Adding more to the AI, having correctly unbiased datasets, more modalities, will make it more flexible and better able to play around with the rules it learns.
Yet I find it hard to conceive continuous learning through dynamically changing optimization goals -- usually, even with things like RL, learning is done separately from inference, and it converges. Training can go wrong and it is also super costly.
Moreover, you can't come up with a benchmark for continuously changing goals, and no one is interested in doing research that you can't evaluate and publish.
So that the fixed objective still prevents AI from being creative on its own (it is dead and stagnant), but given what you say, people will be able to express more and more creativity through these AI models.
Yet I think it'll still leave its mark (someone else pointed this out too) -- because so much of the process is outsourced to the AI, there will be a big similarity between different AI artists, that the general public will grow to detect.
@@ArtPostAI - Using AI image gen tools quite a bit I totally agree with that current and foreseeable image gen AI has big shortcomings in the range of expressible ideas
(while simultaneously tremendously expanding expressibility of ideas for people that couldn't visually express them before at all).
And I very much agree that stopping drawing due to image AI becoming a thing is a rather stupid and bad decision.
AI has much room for improvement. IDK how fast it do will from here.
Muldimodality is so darn important and still barely explored.
Missing beyond that is gradual incremental learning and a long time memory that's not oddly tacked on.
May come eventually. IDK when.
We as humans especially when young and unexperienced run into dead ends too.
Writers block, or blank paper block. Inspiration does not magically strike by waiting and praying to a Muse.
We need to go out and see new parts of the world (or read papers from an other field) to get unstuck (or out of a honed in boring getting old artsytle).
We need to take a step back to the meta level to see other trees of the forest to dynamically change optimization goals.
Going to the philosophical extreme when stuck on the meta level going to the meta level of to the meta level.
Perhaps we as humans are ultimately limited by how far we can step back to see more of natures beauty
limited by financial and cognitive means we've been granted.
A fascinating philosophical topic where new ideas ultimately come from.
Current AI lacks the agency and capabilities to go out and seeking new meta understanding on its own accord.
Plus AIs it's innards are purely deterministic (exact reproducability by giving same seed).
I don't want to accept humans mental innards work the same.
It scares me in a philosophical sense that would go to far here to explain.
Well, basically is shrinks the multiverse by an uber astronomic factor in number of experiencable futures. Sorry, I digress.
Reducing the number of possible outputs (colors^pixelnumber) to the number of possible seeds (symbols^symbolnumber). OUCH!
You mentioned you work in AI research. What is your focus area of your research?
Another point you made is that "ChatGPT can't generate a bowl of Ramen without chopsticks" to illustrate a point about convergence. This doesn't work on my end. Both ChatGPT and Midjourney are able to generate images of Ramen without chopsticks via a one-shot prompt. Literally just "Ramen served alone".
You also mention that evolution is creative, but that doesn't seem right either. Evolution being the name of a long term process in which entities adapt to the environment through their generations. This is the focus of evolutionary algorithms and reinforcement learning fields. The most standout example would be the Alpha series from DeepMind. They partly used reinforcement learning to create AlphaGo, which was described as "creative" by Lee Sedol and other professional Go players back in 2016.
Now, to be fair, current image generation models aren't doing anything like this outside of RLHF. They also aren't their own autonomous entities like we are. They're single pass inference models done based on human input prompts. This puts the "creativity" on the prompt provider. You can say it isn't skillful, but the prompt is the creative bit. Through using an AI model, they trade off precision for speed. They can kinda get what they want from the model, but to create something with high precision, they would need to manually create it with finer precision tools (i.e., paint brush, pencil, etc.)
You also do this thing when comparing AI to humans. It is that when it comes to AI, you go into very granular detail about how they operate. Then for the humans, you become very abstract by comparison. The point where you said that AI have an objective function is different from humans. Reason being is that you can ask humans the meaning of their life is, and they cannot say. It's an apples to oranges comparison that doesn't work in my opinion.
For instance, humans have stratified levels of goals that are constantly changing in priority. Humans have the goal to survive, which motivates us to eat, fit into our communities, sleep, and use our bodies. Then we have goals that are seemingly abstract from our physical bodies; trying to get a particular job, wanting to watch a particular TV show, and etc.
The ending point was kinda disappointing because there are very clear reasons why humans would still prefer human art, even if a machine with sentience could be more creative and skilled at art making.
Thanks a lot for the detailed comment !
I work on topics surrounding consciousness in AI.
"evolution is creative, but that doesn't seem right either."
Evolution gets us from molecules, to fish to humans and all other forms of life. To me, that's a creative process.
Outside of image generation, as you mention, there are ways of training AIs that try to mimic evolution. For instance with competition in AlphaGo or evolutionary algorithms (which don't mimic evolution well since they converge and evolution doesn't)
The competitive approach has been tried in image generation (Generative Adversarial Networks) and it doesn't lead to creativity.
"It's an apples to oranges comparison"
That's kind of the point. AI has a single, perfectly defined, mathematically formulated goal that it chases without any other distraction.
on the human side, we "have stratified levels of goals that are constantly changing in priority".
And we don't even know if we should call them goals, instincts, drives, desires or something else.
We don't even know if it's correct to think of them as goals. We don't know how many there are.
You're right about the ramen thing, I guess it was a wrong example since it's so easily countered. For me typing "ramen without chopsticks" invariably gives me chopsticks in the image. But you're right.
@@ArtPostAI "Evolution gets us from molecules, to fish to humans and all other forms of life. To me, that's a creative process."
I guess it depends on what you define as creative / creativity. I think of creativity as something that a living entity does rather than creativity being a direct byproduct of an environment. This feels more intuitive because we can look at a beautiful landscape, we wouldn't think that the environment creatively crafted it. It's just the natural state of our ecosystem at work.
"For instance with competition in AlphaGo or evolutionary algorithms (which don't mimic evolution well since they converge and evolution doesn't"
I think the process of evolution is being followed correctly. An entity is dropped into an environment for it to exist in. Then through its generations, it adapts itself to this environment. This is why I believe the process is correct, but the staging ground for AI isn't the same as our human environment. It converges because the environment and its own "self" are simple. Like how cells and bacteria are very specialized and generally perform only a small set of actions.
At this point I am knit picking technical details, but I guess my sole point of contention that lead me to those issues is that I disagree with your idea that artists are safe because AI isn't creative.
I feel like it wouldn't matter even if AI was creative. Art is a form of communication. Humans crave this connection. Be it from drawing a picture, carving a sculpture, doing a dance, or anything else means something coming from humans we can share a bond with. It'll be interesting if AI can be creative. It would be a way to engage with it on a new level. But this doesn't replace human creativity in the slightest.
@@InnsmouthAdmiral Thank you so much for the nuanced and insightful response.
I think you've got some great points about there being more than "mere" creativity in human art.
Very interesting point about cells and bacteria as well.
I'll keep that in mind :)
Thank you for cutting thru all the BS, and getting straight to the "root"!!
thanks :)
I love this video, it’s a great way to lay out the issue to the non-educated. I’m only a hobbyist artist myself, but I am very interested in the societal impact AI will have. A little thought I had about your point of the ouroboros-ing of AI art:
Besides the capitalistic reasons for not doing so, couldn’t a machine running without a goal or restraints eventually find a niche in our world, as it encounters real-life limitations? If we could prompt it to fulfill an art-niche, couldn’t it have the same existential anxieties that you claim divides human art from AI art? Also, do you think people without such anxieties are capable of creating art?
Fantastic question to ponder.
I guess it's the logical next step when we say "AI researchers are scared of not chasing objectives" : what would happen if they weren't ?
If we did manage to run a machine in the real world without an objective, goals or restraints, I feel like that might take out a lot of AI's flaws. Probably not all, but we'd get a different animal than the one we're currently dealing with for sure.
"If we could prompt it to fulfill an art-niche, couldn’t it have the same existential anxieties that you claim divides human art from AI art? "
How do you prompt it ? With an objective function ? If so, then I don't think it can get these anxieties.
Really, we don't know how to add emotions into AI, further talk would be speculatory I think.
do you think people without such anxieties are capable of creating art?
I don't think there are people without existential anxieties. But apart from that, it would be arrogant to try and say that certain people cannot make art. People always surprise us.
More details but i am an AI geek. And my thoughts on creativity, just generate noise according to fractals.
Care to elaborate ? :)
By doing so, you might create a new style, but it'll be just one style. Not a creative machine that creates styles, right ?
I find this subject very interesting, thanks a lot 😊
but what about the idea that our brains are designed to make us behave in a way that increases the likelihood of our survival or some criterion like that, I think that human brains works in a pretty homogeneous way actually, we kinda converged towards something
It's hard to say where our creativity comes from, but if it emerges from the level of intelligence we have in our ability to represent our environment and interact with it, then convergence itself (aka using some sort of likelihood) won't necessarily prevent AIs from developing creativity.
I guess the thing restraining modern AI to be creative would be the lack of generality in what they converge towards, maybe they'll get creative only if image generation comes as a by-product of the model's abilities
What a fantastic comment ! :)
We could argue that evolution is just an optimization process aimed at optimizing reproduction likelihood.
But it is a different optimization than mere gradient descent, and is so much richer that it's not obvious if it's a quantitative or qualitative difference. I guess only time will tell.
Biologically it feels like human brains have converged because we look at it at our timescale. But if you check back in in a million years, you'll see something completely different.
What is creatively changing at our timescale is culture. Culturally we aren't converging at all.
You could imagine a kind of internet overwhelmed with AI images, where they learn from each other and develop a sort of flowing, constantly changing and reacting (and thus potentially creative) culture.
But the evidence we have so far for AI having this ability is not promising. Generative adversarial networks do converge even though they compete forever. Evolutionary algorithms also converge.
I think creativity is to be looked for outside of representation capacity. And thus you hit the nail on the head by saying that you need to train them for something else than image generation, to get creative image generation.
I'll keep that in mind actually.
but in the industry most artist has a clear goal set by client or company, experimenting is just a small part of the process (exellent video btw finally an answer rather than just saying adapt to it)
Yes, since I'm also trying to make enjoyable art in the future, and I cannot deny the reality of present and incoming AI, I want to get into details about how to adapt to it..
Industry art follows an objective, but that objective is always much more loosely stated than the perfect, mathematically-defined objective of AI art.
There are so many constraints on objective functions in AI, that even with all the research that's gone into it for over ten years, basically only two effective objective functions exist for generative AI :
-Maximum likelihood estimation,
-the GAN-style objective (which doesn't work as well).
Whereas objectives in industry can be set by people who are themselves quite free and creative.
Would you agree with this view ?
@@ArtPostAI I think it all comes down to the companies. Are they willing to spend more for creativity or do they prefer mediocre and cheap production.
What is this song in the end? Please don't tell me it is AI generated. The voice is similar to Ella Roberts voice...but is too in tune like it was auto tuned. If you think a lot of people thought Hatsune Miku was a person not a Vocaloid (Vocaloid is not human but it does not use AI) and many good Japanese composers began writing for it and using it like Yoneszu Kenshi and Ado...well everything is possible.
UPDATE: I found the video where you said it was AI generated...ooooh the inunmanity! Machine men (and women) with machine minds and machine hearts...and voices.
You're really funny
“Any risks taken by moving out of the boring and typical, and generating a somewhat creative image is immediately understood by the model as a bad thing”
Wow, that sounds like fundamentalist conservatives
Thank you for the in-depth, insider perspective. It does make me feel better as an artist. As a professional artist, I still worry about companies who care more about money than creativity. But that’s a different story
Thank you so much :)
Yes that's a different story. I'm curious if you have any details on how your experience of AI is as a professional artist ?
Which kinds of people support it (I'm thinking ceos and executives ?)
How do people feel about it ?
How fast is it being adopted ?
I feel like a lot of companies are being slow and deliberate with integrating AI because they're aware that it's a hype-ridden field.
How fast is adoption ?
@ArtPostAI Hi hi, sorry, I completely forgot to reply to this
Several companies have gotten into hot water for using, or supposedly using ai. Netflix released an anime where all the backgrounds were created with ai, and they didn’t even credit the person who generated it. In the credits, they are simply listed as “AI + Human.”
Magic the Gathering got caught using ai, as well as the drawing tablet company, Wacom. Marvel’s been accused of using ai for at least one of its covers, though I’m not sure if that turned out to be real.
As it stands, in the US, ai is not protected under copyright, because only humans are allowed to hold copyright (oddly, thanks to the peta monkey selfie. One of the only good things to come out of peta). However, ai is copyrightable in China, so I wouldn’t be surprised if we see a slew of ai produced media coming out of china in the next few years.
Recently, Adobe has recently been extremely shady with their ToS, by forcing everyone who uses their software to give over the rights to all of their files. The thing is that this breaks NDA contracts. Adobe says it’s not a big deal because it’s for scraping data for their ai. But that does nothing to protect designs from being reproduced before a project is released. And with Adobe being the industry standard in entertainment, that has the potential to cause a lot of problems.
And on the voice side of generative ai, open ai got into massive trouble for stealing Scarlet Johansson’s voice. Voice actors specifically got a garbage deal from the writers’ strike. Johansson is lucky she is famous enough to speak out about it, but the majority of people are not so lucky. We are going to see even more voices stolen and more jobs lost in the future.
So yeah, this mass push for ai, being forced upon industry creatives is extremely detrimental to the entertainment industry :/
AI can be a wakeup call for the artists to realy look for what makes art valuable in the first place
what makes art valuable is the time, effort, emotion and meaning that went into it, which is something AI can't replicate
@@naiveghost9683 I don't think so, but maybe. What i think makes art valuable is not what is put into it; what makes art valuable is what comes out of it. An artist can put a lot of time, effort, and emotion into a piece of art, and yet have little to no effect on anyone. On the other hand another artist can draw one line in a unique way on a blank piece of paper in seconds or minutes that then goes on to have a very significant effect on people. Art is about inspiration, not hard work necessarily.
@Kilakilic There's definitely truth to that statement. Hardship creates purpose.
@@naiveghost9683 Certainly my opinion too. Seeing art such as Yuri Norstein's Tale of Tales, it is hard to argue that the sheer willpower of an artist invariably adds magic to a finished piece.
@@alexgonzo5508
Very interesting take.
Art can be quick or slow, and each one has its advantages.
Harding was an advocate for quick art for instance. I believe Pre-raphaelites, Hudson school river painters or classicists might have disagreed.
As you mention, inspiration is separate from the amount of hard work. One might have inspiration for a new way to create art, that happens to be fast.
Or, one might have the kind of inspiration that drives slow, deliberate and loving study of nature.
Both are as valuable.
But I believe that the awareness of the artist's emotion and sheer effort is part of the appeal of more elaborate works.
All the while the appeal of more spontaneous art, is the mystery of chance and intuition.
Is a satellite image of some field in Iowa art? Without emotion AI will never be able to make art art doesn't just require creativity but it requires emotion add a more fundamental level to feed that creativity
No matter how incredible a picture an AI generates maybe it is but a satellite image of a field in Iowa it will never be art
To me, art is communication. Taking an unconventional route to express things that could never be expressed or accepted if seen from the typical angle. Bypassing judgement by changing the medium.
As such, AI art seems like a restraint on the artist's ability to communicate. Because of all its flaws.
But each will have their own definition of art.
Defining art, feels like a lost cause at this point right ? Ever since Duchamp's urinal, it feels impossible.
So I guess your comment keenly his taps into the tough question of the value of art.
What is art worth for the maker ? Making AI art is a cold, digital process, much less enticing than the journey of human art.
For the viewer ? I think AI art and human art will both have their place in satisfying the viewer in the future.
I don't think AI art has meaning or value. Sure, they can make beautiful pictures, but the value of a piece of art for me isn't necessarily the art itself. It's the process someone underwent to make it. Creating is what's meaningful, and typing a prompt into a machine to get a pretty picture isn't creating. It's really not that different from googling a picture. Doesn't have as much impact.
Very good point
"AI" and "Never" do not go together well. I guarantee you, there will be an art generator one day that is even more creative than humans. All we need is an LLM (which has already been proven to have creative abilities that are even greater than some humans) with an image generator built-in. Oh wait, that's exactly what GPT-4o is, and its image generator puts all previous image generators to shame, none of you have seen it yet because it was intentionally undermarketed.
I get you're scared that you won't be able to monetize your art, but you will never be stopped from doing art, you'll just have to get a different job. Human art will always have value in the fact that it was human-made, it just isn't going to be very profitable. These AI tools give everyone the ability to do what you do, and even better. So many people have great ideas but no way to bring them to life, this gives them the power to create what they want with minimal effort. To me, that sounds pretty damn good, and if you think that we shouldn't have this ability just so you can profit off it, I don't think you deserve the money.
Thanks for the comment, it really helps me think :)
I work in AI and use GPT-4o every day. It's doesn't feel any smarter than GPT4, it's just a little faster. And it certainly isn't creative, partly for reasons I cover in this video. GPT4 was already able to "see" and generate images. And having used both image generators, I still like Midjourney a lot more. Just another big marketing thing from OpenAI with everyone buying into the hype.
The reason we think AI is creative is because we have completely abandoned the true meaning of creativity. In the scientific age, we have come to believe that creativity is just cleverness or rapidity of thought. That's not what it is.
Writers I talk to say that GPT4 prose is obvious and ugly. Artists don't like the "AI look" that gets burned into images.
You have to control it through text, which lacks the precision required for true expression.
It requires very little practice, which takes away the process of mastery and all the richness it brings to the mind.
It lacks the soul-filling enjoyability of real, disciplined art, which encourages you on a journey to a completely different way of seeing the world.
It disconnects you from Nature (not just forests but the entirety of the world), which is mostly where artists get their inspiration
It takes you out of your body, which is where intuition and immediacy are (talk to any painter or sculptor and they will tell you about the body)
I have said much more on my channel, and will continue to do so
I'm not saying that AI shouldn't exist because people want to make money from their art. I don't think AI will take that away, in part because of the reasons above.
But the argument that AI art is a good to society is hard to defend. For one, people will get used to it and it'll become just another form of banal communication. People won't be creating art with it for the most part, because having access to Midjourney simply doesn't make you an artist, just like having cameras didn't make everyone artists.
So let's assume that a multimodal LLM is all that's required to put artists out of their jobs. What people will make with it will be reminiscent of the uses people found for photography : sharing things with their friends, etc.
A select few will make actual art with AI art models. They will find disruptive truths and communicate them eloquently through these glorified rendering engines. They will capture a new feeling, a new aspect of life, and make it flow through their images.
But for the most part it'll just be a tool for expressing things that could be said through text, photos or even little sketches that everyone can do.
In this (unlikely) situation, then we would have taken out the deepest, most exploratory form of expression that humans have (art) and replaced it with yet another way of making holiday pictures, but for imaginary holidays. With just a few people actually making something worthwhile out of it.
Tldr : Trying to claim that AI will get creative is the wrong way of looking at things. Humans will always be behind the computer, prompting these things in one way or another. AI simply offers a new form of rendering. The question is : will this new form of rendering take over all other forms of art ? IMO, no. For many reasons I cover on the channel.
Nah mate you don't get it do you. They are going to train a text model on a prompt database of hand-picked pack of users who get great results from AI and feed it into an image model. They'll call it "creativity" and sell for profit, accelerating the commodification of real creativity even more. I don't get why they attack us so relentlessly. Maybe someone didn't manage to get into art school again. Maybe an artist slept with his wife. Something along those lines.
I do think there is rampant envy against artists in our society. Many people wanted to live the artist's life but didn't dare to try. Now those who are unhappy about their lives have a way to get back at those who make a living in a fulfilling manner
@@ArtPostAI Well, I believe creating AI is quite a fulfilling endeavor, somewhere in the middle between art, engineering and business. Art is theft, in essence, so stealing from ALL of the artists is quite a fun idea. Nonetheless, this looks more like a personal grudge, maybe contempt, not envy. That's what I'm trying to say))
Nah. AI can mix concept together and everything new is a mix of old things. Anyway, if you are creative AI is the best thing to ever happen.
How?
How is ai the best thing to ever happen? What?
lil bro is gas lighting
@@bananaman7529 my thoughts exactly.
This could hold water if it weren’t for the point about optimization. The AI is specifically supposed to be making images as similar to its references as possible. In this framework, the AI would be incentivized to mix images as little as possible. In the case it was presented with a prompt it didn’t have a reference for, it will simply default to the next best thing. For these reasons, the AI image generators we see today will become increasingly “perfect,” removing the conceptual value.