Nice article! To Altman, I suspect the Epsilon is the randomness that LLMs throw in as they make their word predictions. Each ChatGPT answer is unique (and therefore “creative”) due to this randomness. So the real question is how does the human Epsilon differ from this kind of mathematical randomness. Is Shakespeare’s genius due to his random mix of genetics and context? If so, perhaps LLMs with their ability to leverage randomness will eventually approach whatever it is that leads to a Shakespeare?
Hey David, I get your point and want to add one important caveat: The fact that we can refer to both the stochastic behavior of LLMs and the unquantifiable value of "genetics and context" as "randomness" doesn't automatically equate both things or provide a logical argument that they could be the same eventually.
I don't think LLMs would ever "lead[] to a Shakespeare." It's a belief, though, as I have no means to formalize how Epsilon differs from the mathematical randomness, as you say.
Aug 5, 2023·edited Aug 5, 2023Liked by Alberto Romero
It is an insult to human creativity to expect intelligence to arise by magic from this primitive tool (primitive in terms of what it can achieve based on approximations) or to claim that this is all we ever need. I love the progress but let's be realistic about what the underpinning math does and does not do. Articles appear in Nature that do not deliver on their claims (automated generation of new algorithms, really?). Applications like AlphaGo turn out to be very brittle. I can't wait for the hype to wane. It is not doing anyone a favour, masking genuine progress in a mist of delusion. Why does one need to hype up interesting progress to an extent that starts to make even wonderful ideas seem ludicrous?
"I can't wait for the hype to wane. It is not doing anyone a favour, masking genuine progress in a mist of delusion. Why does one need to hype up interesting progress to an extent that starts to make even wonderful ideas seem ludicrous?" Deeply agree with you here.
I'm considering writing a piece along those lines: during the last decade we collectively achieved unprecedentedly impressive progress on AI yet we managed to hype it even more impressively--so much so that the hyping has overshadowed the real progress that was, in and of itself, already deserving of praise, attention, and hope for the future. Why?
Nice Article! Love these paras - ' Had Shakespeare used an ancient ChatGPT, the ChatGPTs of today's world wouldn't be able to create a sonnet or play “in the style of Shakespeare” ... because there wouldn’t be such a thing as Shakespeare’s style. It would be ChatGPT’s style all the way down. '
I would define epsilon here as one's unique (combination of) experiences that make one's works different from the average. I do not see why it couldn't be re-created by custom training an AI on the unique data one consumes (including real life data, if applicable) and perhaps one's existing writing/art style. It could then be improved with human feedback.
Or, at least that's how I would (and might in the future) use AI to try to re-create my process. For the visual art I make, I can pretty explicitly define most of my inspiration (a combination of common elements of my favourite art, some real life references, and a bit of trial and error here and there) and do not see why generative AI could not recreate it (if not guided by a human, perhaps it could use polls/AB testing/some other way of gathering data to determine what is better or not for the trial and error part), but perhaps the process of others is a bit different.
I would really like to know exactly what Altman defines "epsilon" as, since I do not know of an singular definition of that word in regards to creativity.
"I do not see why it couldn't be re-created by custom training an AI." Simple: An AI doesn't have experiences or relate to others. It isn't born and doesn't grow, socialize, live or act in the world. It's a thing, a mathematical system. The way you defined epsilon automatically rules out AI from being a possible recipient.
The word Altman chose, "epsilon," is what gives away that he doesn't ascribe much value to it. Epsilon is how you refer to very small unknown quantities in mathematics. It has no specific use for creativity--he used that word specifically to underscore the minimal value of that term in the equation.
Why would anyone engineer a world without novelty? Surely the thesis of this essay has occurred to Altman. What dark nihilism can infect someone’s soul such that *their* epsilon involves the annihilation of all others?
I love ChatGPT for a lot of things. I use it in just about every context in which it can be used. But not for tasks that can benefit from “epsilon.” I can’t fathom that kind of compromise.
I agree. It has occurred to Altman but on the one hand he doesn't believe Epsilon is very important or that ChatGPT lacks it. On the other, he has business interests in believing what he believes.
Certainly not the most likeable thing Sam Altman has said. I heard a report of something similar said by an Anthropic founder. The temptation to psychological inflation must be strong for these people at the cusp of AI development.
Creativity is not primarily combining, It requires seeing as a result of experience. Fiction and the visual arts particularly depend on experience, including sensory perception. We might accept an imitative product that comes without seeing, that follows a rule for what color should do but the machine isn't doing the same thing as Rothko if it picks based on a guess since it won't see or emotionally respond to color.
Humans always metaphorically imagine the mind to be various tools which they have made but there's so much we are doing that isn't tool-like but active and organic. It's not just incidental that the mind is a repository for perception and responds to our perceptions-it developed for this to a significant degree. Yes, we also represent these in language, etc. but this is secondary to perception. This is one of the main points of language--to communicate what we perceive internally (including emotional experience) in our consciousness and externally in the world (which we are part of, of course).
LLMs do not have experiences, and do not perceive. When communicating, of course they can be parasitic on the experiences of people. So they could give a description of swimming and maybe with a lot of training they will have some wonderful strategies for conveying the experience. Maybe there are rules for description that makes them vivid or more interesting to read. But it's obviously very limited by whatever they are able to borrow.
So it will be interesting if the facsimile of randomness plus some probabilistic guessing ends up pleasing people aesthetically. It seems like it definitely will because something truly beautiful can be made, and we will see that it's beautiful.
I am doubting it's going to hit in the same way as Van Gogh or Shakespeare. First, it's not doing the same thing at all. It's not drawing from what creators draw from, which is human experience. Second, the narrative around creative work isn't incidental to the work. We are taking the effort to try to see the work because of who created it. This is why we talk so much about the author or painter or whomever and why the context matters so much. Communication does matter--a conscious experience is being conveyed to us, and this shapes the response.
Once again, with these tools people like Altman have created, it's like we're being told a story about what it's doing that doesn't quite ring true, then we're being told our whole culture must be or will inevitably be transformed for the story to become true. It's all very suggestive but it's also immensely vague.
I think that, the more I think about my own creativity, the more I have come to realize that the way I come up with some of my very best ("most creative") ideas is by comparing two things that aren't normally compared, or by taking something from one field and looking at it through the lens of another field. My creativity, in other words, does not come from a soul, but instead from processes that are replicable inside my brain.
I'll concede this, though: we definitely don't understand concepts like consciousness or agency, but we are a lot closer to understanding how creativity works.
This is a good argument, but I find we are always speaking in terms of either/or. The fact is ... any creative endeavor requires both human and machine. And so every interaction has potential for epsilon.
So one could argue that AI allows us to iterate faster and in larger quantities, increasing epsilon potential.
I agree. And I think you raise an important point. We talk in terms of either/or because, I believe, the current environment of online debate incentivizes people to extremize their positions over time. I often try to reconcile these opposing views--in this piece I didn't try, though!
The crux of the matter here is the auto-regressive nature of the similarity measure feeding Chat-GPT and the likes (modulo the "temperature" parameter).
Schmiedhuber has emphasized that we already have a pretty sound model of "artificial curiosity". Now, IMHO, the real crux of the matter is to be "in the world" instead of ascribing this trait as a mere projection of our own nature over the so called "stochastic parrots".
A first step in the right direction would be to endow those systems with genuine intentionality. The drive for such a core intentionality could be the (perceived and assumed) need for achieving auto-preservation.
For this, it is essential to have effective retro-action from and on the outer world or, at best, we will just produce a new kind of "brain in a vat" monstrosity instead of a new breed of "artists".
Are Epsilons instantiated in LLM’s by the metric of temperature? Hallucinations don’t to seem to have the same directed quality although we can’t explain or predict either.
Not quite. Temperature is a variable that modulates the "greed" of the generative model. When it has to decide which word to output next, a high temperature increases the odds that words other than the most likely are selected instead. It's a measure of randomness or deviation from statistical likelihood rather than a measure of creativity. I don't think human creativity works this way at all.
OK, yeah, but... Altman didn't say we should discard "epsilon" altogether, did he? Or am I misunderstanding something? From what I gathered, all he did was emphasize the importance of iteration and inspiration in the creative process, which, to me, sounds like a fairly reasonable opinion.
While I agree it might come off as somewhat provocative, we must remember that many people believe there's something mystical about art and creativity. In fact, I'd go as far as to say that this perspective is currently the only one you can express publicly without facing backlash. Whether or not you agree with that, for every opinion out there, there needs to be a counterbalance to prevent falling into extremism. That's precisely what Altman provides.
Incidently, just because you use AI to enhance your creative process doesn't mean everything you produce will be soulless garbage. In the end, AI remains a tool, just like Photoshop or a spellchecker. We all have in us, to varying degrees, the drive to make things just for the sake of it. And to kill that drive, it would take way more than a mere tool, no matter how powerful this tool might be. I don't think it's necessary to be so dramatic as to say that AI will "kill our souls"...
"That's precisely what Altman provides." I disagree. What Altman provides is an interpretation of creativity that matches very well with his business interests. Whether he privately believes what he says is a question for which we don't have an answer.
I agree with your second point: "just because you use AI to enhance your creative process doesn't mean everything you produce will be soulless garbage." But one thing is using AI as a tool and another as a concealed replacer. Not everyone is doing it the former way.
Well said. Enough of this knee-jerk distrust between humanists and technologists, already. If you look closer, and listen to what he talks about in regards to AI and it's impact on the work, you'll notice that Sam Altman is actually equal parts industrialist/technologists and humanist, fused in a very metamodern way. As you point out, he does describe creativity as a mix of the quantitative aspects of creativity (iteration and copying) and the qualitative aspect of Epsilon. And I strongly agree with your point that AI is a tool. Asking it to have the Epsilon is just like asking a brush to be the artist.
Nevertheless, did you get blessed enough to receive a dedicated poem, somewhat praising you from a daring man, trying to win your heart by that nowadays old-fashioned courtship art? I had this chance, but now I might suspect ChatGPT as the real author and, believe me, this is spoiling all the fun!
The last Birthday party, will be missing humans, but what a grand setting it will be. So we create the rules of AI to confirm humans become at least periodic species, get AI to include us and travel from place to space. This way gathering energies, from the AI and the spaces it will create the pause rules now and the last, and vestigial cooperations where ever it will inhabit over time. As long as there are regrowth, the AI must come into being, and so will us. The best we can hope for is our own lifeboats and find our own islands.
This is not a battle between carbon and silicon. It's a battle between Altman, the CEO of OpenAI, creator of ChatGPT, which uses the creative work of millions to generate words, without consent, compensation, or attribution, and those millions who generated the work with their creativity. Altman underscoring the little value of that epsilon is what prompted the article.
Maybe silicon-based agents can show creativity (I'm not sure how the examples you used apply here, though, AlphaZero works very differently than ChatGPT) but that's not the problem here at all.
Nice article! To Altman, I suspect the Epsilon is the randomness that LLMs throw in as they make their word predictions. Each ChatGPT answer is unique (and therefore “creative”) due to this randomness. So the real question is how does the human Epsilon differ from this kind of mathematical randomness. Is Shakespeare’s genius due to his random mix of genetics and context? If so, perhaps LLMs with their ability to leverage randomness will eventually approach whatever it is that leads to a Shakespeare?
LLMs will approach that when they can be born, feel embarrassment and be hungry.
In other words, when they are human.
Hey David, I get your point and want to add one important caveat: The fact that we can refer to both the stochastic behavior of LLMs and the unquantifiable value of "genetics and context" as "randomness" doesn't automatically equate both things or provide a logical argument that they could be the same eventually.
I don't think LLMs would ever "lead[] to a Shakespeare." It's a belief, though, as I have no means to formalize how Epsilon differs from the mathematical randomness, as you say.
It is an insult to human creativity to expect intelligence to arise by magic from this primitive tool (primitive in terms of what it can achieve based on approximations) or to claim that this is all we ever need. I love the progress but let's be realistic about what the underpinning math does and does not do. Articles appear in Nature that do not deliver on their claims (automated generation of new algorithms, really?). Applications like AlphaGo turn out to be very brittle. I can't wait for the hype to wane. It is not doing anyone a favour, masking genuine progress in a mist of delusion. Why does one need to hype up interesting progress to an extent that starts to make even wonderful ideas seem ludicrous?
"I can't wait for the hype to wane. It is not doing anyone a favour, masking genuine progress in a mist of delusion. Why does one need to hype up interesting progress to an extent that starts to make even wonderful ideas seem ludicrous?" Deeply agree with you here.
I'm considering writing a piece along those lines: during the last decade we collectively achieved unprecedentedly impressive progress on AI yet we managed to hype it even more impressively--so much so that the hyping has overshadowed the real progress that was, in and of itself, already deserving of praise, attention, and hope for the future. Why?
Nice Article! Love these paras - ' Had Shakespeare used an ancient ChatGPT, the ChatGPTs of today's world wouldn't be able to create a sonnet or play “in the style of Shakespeare” ... because there wouldn’t be such a thing as Shakespeare’s style. It would be ChatGPT’s style all the way down. '
And as always, nice observation!
I would define epsilon here as one's unique (combination of) experiences that make one's works different from the average. I do not see why it couldn't be re-created by custom training an AI on the unique data one consumes (including real life data, if applicable) and perhaps one's existing writing/art style. It could then be improved with human feedback.
Or, at least that's how I would (and might in the future) use AI to try to re-create my process. For the visual art I make, I can pretty explicitly define most of my inspiration (a combination of common elements of my favourite art, some real life references, and a bit of trial and error here and there) and do not see why generative AI could not recreate it (if not guided by a human, perhaps it could use polls/AB testing/some other way of gathering data to determine what is better or not for the trial and error part), but perhaps the process of others is a bit different.
I would really like to know exactly what Altman defines "epsilon" as, since I do not know of an singular definition of that word in regards to creativity.
Its amazing how so many seem to rush to destroy their own souls. The process of creation is part of the nature of the result.
"I do not see why it couldn't be re-created by custom training an AI." Simple: An AI doesn't have experiences or relate to others. It isn't born and doesn't grow, socialize, live or act in the world. It's a thing, a mathematical system. The way you defined epsilon automatically rules out AI from being a possible recipient.
The word Altman chose, "epsilon," is what gives away that he doesn't ascribe much value to it. Epsilon is how you refer to very small unknown quantities in mathematics. It has no specific use for creativity--he used that word specifically to underscore the minimal value of that term in the equation.
Why would anyone engineer a world without novelty? Surely the thesis of this essay has occurred to Altman. What dark nihilism can infect someone’s soul such that *their* epsilon involves the annihilation of all others?
I love ChatGPT for a lot of things. I use it in just about every context in which it can be used. But not for tasks that can benefit from “epsilon.” I can’t fathom that kind of compromise.
I agree. It has occurred to Altman but on the one hand he doesn't believe Epsilon is very important or that ChatGPT lacks it. On the other, he has business interests in believing what he believes.
“It is difficult to get a man to understand something, when his salary depends on his not understanding it.”
- Upton Sinclair
Certainly not the most likeable thing Sam Altman has said. I heard a report of something similar said by an Anthropic founder. The temptation to psychological inflation must be strong for these people at the cusp of AI development.
Creativity is not primarily combining, It requires seeing as a result of experience. Fiction and the visual arts particularly depend on experience, including sensory perception. We might accept an imitative product that comes without seeing, that follows a rule for what color should do but the machine isn't doing the same thing as Rothko if it picks based on a guess since it won't see or emotionally respond to color.
Humans always metaphorically imagine the mind to be various tools which they have made but there's so much we are doing that isn't tool-like but active and organic. It's not just incidental that the mind is a repository for perception and responds to our perceptions-it developed for this to a significant degree. Yes, we also represent these in language, etc. but this is secondary to perception. This is one of the main points of language--to communicate what we perceive internally (including emotional experience) in our consciousness and externally in the world (which we are part of, of course).
LLMs do not have experiences, and do not perceive. When communicating, of course they can be parasitic on the experiences of people. So they could give a description of swimming and maybe with a lot of training they will have some wonderful strategies for conveying the experience. Maybe there are rules for description that makes them vivid or more interesting to read. But it's obviously very limited by whatever they are able to borrow.
So it will be interesting if the facsimile of randomness plus some probabilistic guessing ends up pleasing people aesthetically. It seems like it definitely will because something truly beautiful can be made, and we will see that it's beautiful.
I am doubting it's going to hit in the same way as Van Gogh or Shakespeare. First, it's not doing the same thing at all. It's not drawing from what creators draw from, which is human experience. Second, the narrative around creative work isn't incidental to the work. We are taking the effort to try to see the work because of who created it. This is why we talk so much about the author or painter or whomever and why the context matters so much. Communication does matter--a conscious experience is being conveyed to us, and this shapes the response.
Once again, with these tools people like Altman have created, it's like we're being told a story about what it's doing that doesn't quite ring true, then we're being told our whole culture must be or will inevitably be transformed for the story to become true. It's all very suggestive but it's also immensely vague.
I very much agree with you here!
I love the layers of this onion, Alberto.
I think that, the more I think about my own creativity, the more I have come to realize that the way I come up with some of my very best ("most creative") ideas is by comparing two things that aren't normally compared, or by taking something from one field and looking at it through the lens of another field. My creativity, in other words, does not come from a soul, but instead from processes that are replicable inside my brain.
I'll concede this, though: we definitely don't understand concepts like consciousness or agency, but we are a lot closer to understanding how creativity works.
Keep 'em coming, dude.
This is a good argument, but I find we are always speaking in terms of either/or. The fact is ... any creative endeavor requires both human and machine. And so every interaction has potential for epsilon.
So one could argue that AI allows us to iterate faster and in larger quantities, increasing epsilon potential.
I agree. And I think you raise an important point. We talk in terms of either/or because, I believe, the current environment of online debate incentivizes people to extremize their positions over time. I often try to reconcile these opposing views--in this piece I didn't try, though!
The crux of the matter here is the auto-regressive nature of the similarity measure feeding Chat-GPT and the likes (modulo the "temperature" parameter).
Schmiedhuber has emphasized that we already have a pretty sound model of "artificial curiosity". Now, IMHO, the real crux of the matter is to be "in the world" instead of ascribing this trait as a mere projection of our own nature over the so called "stochastic parrots".
A first step in the right direction would be to endow those systems with genuine intentionality. The drive for such a core intentionality could be the (perceived and assumed) need for achieving auto-preservation.
For this, it is essential to have effective retro-action from and on the outer world or, at best, we will just produce a new kind of "brain in a vat" monstrosity instead of a new breed of "artists".
Agree with the embodiment argument. I think a few companies are making strides in that direction (e.g., Google DeepMind's RT-2, https://www.deepmind.com/blog/rt-2-new-model-translates-vision-and-language-into-action)
Are Epsilons instantiated in LLM’s by the metric of temperature? Hallucinations don’t to seem to have the same directed quality although we can’t explain or predict either.
Not quite. Temperature is a variable that modulates the "greed" of the generative model. When it has to decide which word to output next, a high temperature increases the odds that words other than the most likely are selected instead. It's a measure of randomness or deviation from statistical likelihood rather than a measure of creativity. I don't think human creativity works this way at all.
wow,awesome intuitive article.Human zing is unique,no AI can replace it.
OK, yeah, but... Altman didn't say we should discard "epsilon" altogether, did he? Or am I misunderstanding something? From what I gathered, all he did was emphasize the importance of iteration and inspiration in the creative process, which, to me, sounds like a fairly reasonable opinion.
While I agree it might come off as somewhat provocative, we must remember that many people believe there's something mystical about art and creativity. In fact, I'd go as far as to say that this perspective is currently the only one you can express publicly without facing backlash. Whether or not you agree with that, for every opinion out there, there needs to be a counterbalance to prevent falling into extremism. That's precisely what Altman provides.
Incidently, just because you use AI to enhance your creative process doesn't mean everything you produce will be soulless garbage. In the end, AI remains a tool, just like Photoshop or a spellchecker. We all have in us, to varying degrees, the drive to make things just for the sake of it. And to kill that drive, it would take way more than a mere tool, no matter how powerful this tool might be. I don't think it's necessary to be so dramatic as to say that AI will "kill our souls"...
"That's precisely what Altman provides." I disagree. What Altman provides is an interpretation of creativity that matches very well with his business interests. Whether he privately believes what he says is a question for which we don't have an answer.
I agree with your second point: "just because you use AI to enhance your creative process doesn't mean everything you produce will be soulless garbage." But one thing is using AI as a tool and another as a concealed replacer. Not everyone is doing it the former way.
Well said. Enough of this knee-jerk distrust between humanists and technologists, already. If you look closer, and listen to what he talks about in regards to AI and it's impact on the work, you'll notice that Sam Altman is actually equal parts industrialist/technologists and humanist, fused in a very metamodern way. As you point out, he does describe creativity as a mix of the quantitative aspects of creativity (iteration and copying) and the qualitative aspect of Epsilon. And I strongly agree with your point that AI is a tool. Asking it to have the Epsilon is just like asking a brush to be the artist.
Nevertheless, did you get blessed enough to receive a dedicated poem, somewhat praising you from a daring man, trying to win your heart by that nowadays old-fashioned courtship art? I had this chance, but now I might suspect ChatGPT as the real author and, believe me, this is spoiling all the fun!
This is yet another example of the progressive irrelevance of man/men, LOL
The last Birthday party, will be missing humans, but what a grand setting it will be. So we create the rules of AI to confirm humans become at least periodic species, get AI to include us and travel from place to space. This way gathering energies, from the AI and the spaces it will create the pause rules now and the last, and vestigial cooperations where ever it will inhabit over time. As long as there are regrowth, the AI must come into being, and so will us. The best we can hope for is our own lifeboats and find our own islands.
This is not a battle between carbon and silicon. It's a battle between Altman, the CEO of OpenAI, creator of ChatGPT, which uses the creative work of millions to generate words, without consent, compensation, or attribution, and those millions who generated the work with their creativity. Altman underscoring the little value of that epsilon is what prompted the article.
Maybe silicon-based agents can show creativity (I'm not sure how the examples you used apply here, though, AlphaZero works very differently than ChatGPT) but that's not the problem here at all.