This feels a bit like a "no true Scotsman" argument. Because the AI can't make interstitial decisions, what it's making isn't actually art? I suppose the internal experience of making art is truly...
This feels a bit like a "no true Scotsman" argument. Because the AI can't make interstitial decisions, what it's making isn't actually art? I suppose the internal experience of making art is truly the last refuge of the Artist.
I basically read the article and inserted the word “good” before every usage of the word “art.” By definition, yes, of course AI is making art already. But it is, by its very nature, reductive and...
I basically read the article and inserted the word “good” before every usage of the word “art.” By definition, yes, of course AI is making art already. But it is, by its very nature, reductive and bland, and therefore lacks legitimacy in my opinion (and it’s just not that good). In this same vein, I can record myself banging on a pan and screaming at the top of my lungs, and say it’s a piece of live performance art, but that doesn’t mean that anybody needs to validate its legitimacy or consider it good.
When I put this into a tech lens, the analogy I like to use is like when a software engineer is talking to somebody who just learned how to copy and paste a bash script into their CLI and the person says, “wow, this is all you do all day? This is easy.” AI “artists” are like script kiddies backing into art and trying to stake their claim because, “hey, this is art, too.”
Current (important disclaimer!) generative AI is pretty much a mathematical formalization of “generic”. It’s the most likely version of a text. There’s some random noise in there to keep it from...
Current (important disclaimer!) generative AI is pretty much a mathematical formalization of “generic”. It’s the most likely version of a text. There’s some random noise in there to keep it from getting stuck in loops and whatnot. But it is, by definition, predictable.
I’d make the claim that almost all art worth consuming has something about it that cannot be derived from previously released art, something truly new. “New” in the sense, of course, that it has been discovered by the artist via observing life and nature, going through the world with open eyes and paying attention to something no one has ever bothered to write down/paint/etc.
Once we figure out how to train AI with more real-world data, i.e. not just the text of billions of internet comments, all bets are off, though. I just don’t think that jump can come from adding more text parameters. It needs to learn how a baby learns, a microphone and camera pointed at the real world and probably would need a few hard-coded ways of reacting to visual stimuli, a sense of curiosity, etc.
This isn't really the case. In fact, the fact that the models in question here are black boxes and we thus can't predict what they'll do in response to new input is a big issue in the field. The...
But it is, by definition, predictable.
This isn't really the case. In fact, the fact that the models in question here are black boxes and we thus can't predict what they'll do in response to new input is a big issue in the field. The "generic" feeling from Gen AI products is heavily influenced by the guardrails the companies that run these things put on them (or, well, attempt to put on them) and is not an inherent feature of the model. Yes, it's mathematically predicting the most likely next word, but that needn't necessarily be generic. A lot of work went into training the model to make it more likely to be generic.
I agree that the AI isn't making "good" art on its own by my standards. But even with its current capabilities, artists are absolutely capable of making art of varying qualities using it. My main issue, though, is people making the claim that this somehow isn't art at all. It relies on an absolutely incoherent definition of art, imo, that's incredibly inconsistent with existing forms of art.
Well, the use of the word “predictable” is a little mushy here but I still think it fits. I remember the pre-GPT3 AI word salad and it was hilarious but largely useless. The “guardrails” you...
Well, the use of the word “predictable” is a little mushy here but I still think it fits. I remember the pre-GPT3 AI word salad and it was hilarious but largely useless. The “guardrails” you mention aren’t just for security or style, they are the sprinkle of “order” to the “chaos” in the source material needed to balance it out. I’m convinced we’re very close to a perfect balance between getting the most information (and variety) out of the training data while maintaining a sensible level of context awareness and logic. I don’t believe you can “tune” AI for creativity much further.
Gwern made a comment on this article I thought was pretty insightful: I find myself a bit frustrated when people associate AI capabilities with the corporate sanitized outputs of...
Gwern made a comment on this article I thought was pretty insightful:
He's also, broadly speaking, wrong. The blandness he refers to has little to do with 'taking the average of the choices'. Base model output is not bland! It is often wacky, bizarre, hilarious, or unexpected. See the samples people have been generating with the new Llama base model, or just again, go back to my old GPT-3 samples - samples Chiang should've seen since he's been pontificating regularly about LLMs this whole time. (In considerable part because there is usually no such thing as 'taking the average of the choices' in text: what is the 'average' of two likely predictions like 'a' or 'b'? There is no such thing as a letter 50% of the way in between 'a' and 'b', the sampling process has to pick one or the other, and then it's committed.)
The blandness he's referring to is the effect of the preference-learning/instruction-tuning made infamous by ChatGPT, which has no simple analogy but has little to do with "the choices that other writers have made, as represented by text found on the Internet" (in fact, this is the opposite of correct, as making the outputs unlike 'text found on the Internet' is most of the point). The point of the tuning is to collapse the LLM, which starts off as an extraordinary mimic of every style, onto a narrow range of styles & vocabs - yielding the ChatGPTese we know all too well now.
I find myself a bit frustrated when people associate AI capabilities with the corporate sanitized outputs of ChatGPT/Dall-E/Bing Image/etc. Although it's not that surprising, considering that's the only way the average person will interact with AI.
If you're interested in playing around with more authentic AI, I'd suggest a non-finetuned version of Llama, or Stable Diffusion (although the newer versions of Stable Diffusion are a bit sanitized as well, 1.5 is probably the best for wackier outputs). Prompting will be harder, but you won't get outputs that sound/look like AI.
I agree with this, I think that people don't realize this with LLMs because the non-finetuned versions are less readily available and weren't trending as much as image models, plus maybe it's more...
I agree with this, I think that people don't realize this with LLMs because the non-finetuned versions are less readily available and weren't trending as much as image models, plus maybe it's more difficult for people to immediately see the creativity within text than within images.
Stable Diffusion is easy to run on common hardware and it's pretty easy to see the development. Each version was slightly more realistic and less creative than the last. I've spent many hours generating surreal images in Stable Diffusion 1.5 and the results it was routinely giving me were more creative and interesting than what a big portion of twitter/patreon artists produces. Of course, the model is too simple, which limits fidelity, but in many areas it really is surprisingly creative and interesting.
Midjourney seems to be able to keep some balance between fidelity and creativity even as new versions come out, so it probably is possible, but it requires conscious effort that may be perceived as less valuable than focusing on say "safety".
I found this article disappointing, which is why I didn't post it earlier. Here's a response from someone on Twitter that I thought was pretty good. I wonder why someone couldn't use AI for part...
I found this article disappointing, which is why I didn't post it earlier. Here's a response from someone on Twitter that I thought was pretty good.
I wonder why someone couldn't use AI for part of their artistic process? Random choices are sometimes a good creative input, if used selectively by someone with good taste.
The surreal word salad from earlier LLM's was more interesting than the usual output from today's mainstream AI chats. Most AI companies aren't really trying for creativity; they're trying to make AI safe for business.
Maybe we'll see interesting work from people messing around with open source LLM's that are tuned for artistic use?
It's not really that much of a change from existing paradigms insofar as art philosophy goes. Jackson Pollock would be surprised to hear that randomness is not useful in generating art, and Warhol...
It's not really that much of a change from existing paradigms insofar as art philosophy goes. Jackson Pollock would be surprised to hear that randomness is not useful in generating art, and Warhol would probably take issue with a stipulation against derivative works.
The person writing the article is arguing that it is probably not "art" to walk up to an AI and tell it to give you some art, but that hinges on the notion that art is a process of communication. I think that's a defensible definition of art, but most reasonable people should agree that art is not a concept that everyone sees equally, or which has solid, bright boundaries. If someone defines art as equivalent to beauty, then AI can certainly meet that standard every now and again. If someone defines art as mindful curation, then surely selecting AI outputs falls under that standard as much as presenting beautiful leaves and rocks from nature.
I don't mind if people don't like AI; there are reasons not to. Trying to say it's alien to the concept of art is demonstrably false, though. I know several artists that use it either for inspiration or as part of their creative process, and it remains a useful tool despite the assertions of those who prefer to take its inutility as a premise and work backward from there.
That tweet's author seems to have missed the point. Chiang clearly believes that AI can be used in the artistic process.
That tweet's author seems to have missed the point. Chiang clearly believes that AI can be used in the artistic process.
We can imagine a text-to-image generator that, over the course of many sessions, lets you enter tens of thousands of words into its text box to enable extremely fine-grained control over the image you’re producing; this would be something analogous to Photoshop with a purely textual interface. I’d say that a person could use such a program and still deserve to be called an artist.
I agree with what you’re saying, but I also don’t disagree with the statement being made in the article. I think it comes down to what you yourself already pointed out - I think that AI will, in...
I agree with what you’re saying, but I also don’t disagree with the statement being made in the article. I think it comes down to what you yourself already pointed out - I think that AI will, in time, certainly assist with the creation of art. But I’m of the belief that it will not, can not, make good bespoke art. I also think the examples given in the response you linked are all pretty lukewarm examples that prove the point we likely both agree on.
Sampling is a valid part of making music, but unless you’re a freak of nature like Girl Talk, one can’t really consider sampling to be art in and of itself.
“Synthesizers” is so immensely broad, it’s like saying “digital art.” Which, again, I’d maintain that AI generative art is an inch deep and infinitely wide at this point, so sure, you can supplement your art with it, but to claim that its utility comes anywhere close to the leaps and bounds music took forward upon the invention of synthesizers is frankly silly.
Rap? That’s an entire genre of music. Apples to oranges, in my opinion. It also requires lots and lots of talent or effort (often both) to write and perform rap well, and there’s not currently any sort of AI toolset analogous to something as all-encompassing as an entire genre of music.
The list goes on, but my point is that folks try to make this black-and-white argument that if you push back against AI at all, you’re just as bad as the repressed townsfolk in Dirty Dancing. Whereas I’m saying there’s a lot of gray area here, and while I think AI tooling could bring us some really cool advancements in how we approach art in the future, to try to put it in the same category as anything else being mentioned in the “cultural disrupters” category that Seb is mentioning is not an accurate representation of either side of that spectrum.
This feels a bit like a "no true Scotsman" argument. Because the AI can't make interstitial decisions, what it's making isn't actually art? I suppose the internal experience of making art is truly the last refuge of the Artist.
I basically read the article and inserted the word “good” before every usage of the word “art.” By definition, yes, of course AI is making art already. But it is, by its very nature, reductive and bland, and therefore lacks legitimacy in my opinion (and it’s just not that good). In this same vein, I can record myself banging on a pan and screaming at the top of my lungs, and say it’s a piece of live performance art, but that doesn’t mean that anybody needs to validate its legitimacy or consider it good.
When I put this into a tech lens, the analogy I like to use is like when a software engineer is talking to somebody who just learned how to copy and paste a bash script into their CLI and the person says, “wow, this is all you do all day? This is easy.” AI “artists” are like script kiddies backing into art and trying to stake their claim because, “hey, this is art, too.”
Current (important disclaimer!) generative AI is pretty much a mathematical formalization of “generic”. It’s the most likely version of a text. There’s some random noise in there to keep it from getting stuck in loops and whatnot. But it is, by definition, predictable.
I’d make the claim that almost all art worth consuming has something about it that cannot be derived from previously released art, something truly new. “New” in the sense, of course, that it has been discovered by the artist via observing life and nature, going through the world with open eyes and paying attention to something no one has ever bothered to write down/paint/etc.
Once we figure out how to train AI with more real-world data, i.e. not just the text of billions of internet comments, all bets are off, though. I just don’t think that jump can come from adding more text parameters. It needs to learn how a baby learns, a microphone and camera pointed at the real world and probably would need a few hard-coded ways of reacting to visual stimuli, a sense of curiosity, etc.
This isn't really the case. In fact, the fact that the models in question here are black boxes and we thus can't predict what they'll do in response to new input is a big issue in the field. The "generic" feeling from Gen AI products is heavily influenced by the guardrails the companies that run these things put on them (or, well, attempt to put on them) and is not an inherent feature of the model. Yes, it's mathematically predicting the most likely next word, but that needn't necessarily be generic. A lot of work went into training the model to make it more likely to be generic.
I agree that the AI isn't making "good" art on its own by my standards. But even with its current capabilities, artists are absolutely capable of making art of varying qualities using it. My main issue, though, is people making the claim that this somehow isn't art at all. It relies on an absolutely incoherent definition of art, imo, that's incredibly inconsistent with existing forms of art.
Well, the use of the word “predictable” is a little mushy here but I still think it fits. I remember the pre-GPT3 AI word salad and it was hilarious but largely useless. The “guardrails” you mention aren’t just for security or style, they are the sprinkle of “order” to the “chaos” in the source material needed to balance it out. I’m convinced we’re very close to a perfect balance between getting the most information (and variety) out of the training data while maintaining a sensible level of context awareness and logic. I don’t believe you can “tune” AI for creativity much further.
Gwern made a comment on this article I thought was pretty insightful:
I find myself a bit frustrated when people associate AI capabilities with the corporate sanitized outputs of ChatGPT/Dall-E/Bing Image/etc. Although it's not that surprising, considering that's the only way the average person will interact with AI.
If you're interested in playing around with more authentic AI, I'd suggest a non-finetuned version of Llama, or Stable Diffusion (although the newer versions of Stable Diffusion are a bit sanitized as well, 1.5 is probably the best for wackier outputs). Prompting will be harder, but you won't get outputs that sound/look like AI.
I agree with this, I think that people don't realize this with LLMs because the non-finetuned versions are less readily available and weren't trending as much as image models, plus maybe it's more difficult for people to immediately see the creativity within text than within images.
Stable Diffusion is easy to run on common hardware and it's pretty easy to see the development. Each version was slightly more realistic and less creative than the last. I've spent many hours generating surreal images in Stable Diffusion 1.5 and the results it was routinely giving me were more creative and interesting than what a big portion of twitter/patreon artists produces. Of course, the model is too simple, which limits fidelity, but in many areas it really is surprisingly creative and interesting.
Midjourney seems to be able to keep some balance between fidelity and creativity even as new versions come out, so it probably is possible, but it requires conscious effort that may be perceived as less valuable than focusing on say "safety".
I found this article disappointing, which is why I didn't post it earlier. Here's a response from someone on Twitter that I thought was pretty good.
I wonder why someone couldn't use AI for part of their artistic process? Random choices are sometimes a good creative input, if used selectively by someone with good taste.
The surreal word salad from earlier LLM's was more interesting than the usual output from today's mainstream AI chats. Most AI companies aren't really trying for creativity; they're trying to make AI safe for business.
Maybe we'll see interesting work from people messing around with open source LLM's that are tuned for artistic use?
It's not really that much of a change from existing paradigms insofar as art philosophy goes. Jackson Pollock would be surprised to hear that randomness is not useful in generating art, and Warhol would probably take issue with a stipulation against derivative works.
The person writing the article is arguing that it is probably not "art" to walk up to an AI and tell it to give you some art, but that hinges on the notion that art is a process of communication. I think that's a defensible definition of art, but most reasonable people should agree that art is not a concept that everyone sees equally, or which has solid, bright boundaries. If someone defines art as equivalent to beauty, then AI can certainly meet that standard every now and again. If someone defines art as mindful curation, then surely selecting AI outputs falls under that standard as much as presenting beautiful leaves and rocks from nature.
I don't mind if people don't like AI; there are reasons not to. Trying to say it's alien to the concept of art is demonstrably false, though. I know several artists that use it either for inspiration or as part of their creative process, and it remains a useful tool despite the assertions of those who prefer to take its inutility as a premise and work backward from there.
That tweet's author seems to have missed the point. Chiang clearly believes that AI can be used in the artistic process.
I agree with what you’re saying, but I also don’t disagree with the statement being made in the article. I think it comes down to what you yourself already pointed out - I think that AI will, in time, certainly assist with the creation of art. But I’m of the belief that it will not, can not, make good bespoke art. I also think the examples given in the response you linked are all pretty lukewarm examples that prove the point we likely both agree on.
Sampling is a valid part of making music, but unless you’re a freak of nature like Girl Talk, one can’t really consider sampling to be art in and of itself.
“Synthesizers” is so immensely broad, it’s like saying “digital art.” Which, again, I’d maintain that AI generative art is an inch deep and infinitely wide at this point, so sure, you can supplement your art with it, but to claim that its utility comes anywhere close to the leaps and bounds music took forward upon the invention of synthesizers is frankly silly.
Rap? That’s an entire genre of music. Apples to oranges, in my opinion. It also requires lots and lots of talent or effort (often both) to write and perform rap well, and there’s not currently any sort of AI toolset analogous to something as all-encompassing as an entire genre of music.
The list goes on, but my point is that folks try to make this black-and-white argument that if you push back against AI at all, you’re just as bad as the repressed townsfolk in Dirty Dancing. Whereas I’m saying there’s a lot of gray area here, and while I think AI tooling could bring us some really cool advancements in how we approach art in the future, to try to put it in the same category as anything else being mentioned in the “cultural disrupters” category that Seb is mentioning is not an accurate representation of either side of that spectrum.
Mirror: https://archive.is/fQ2UU