Key insights of today’s newsletter: In his new role as the CEO of the Microsoft AI devision, Mustafa Suleyman gave TED talk in Vancouver, Canada. In his talk, Suleyman suggested we need new words and metaphors to think more effectively about AI. He also brought a metaphor of his own: AI as a new digital species, making the case that AI is different from all technologies that came before.
I'd push even more strongly against Suleyman's claim that AI is "creative". It's not just that it was trained on a vast collection of human-made images; it's that training consists of the model "learning" to reproduce what's in the training data. The whole goal is to imitate what it's been fed!
Now, the usual response is along the lines of "that's what humans do, too". And, of course, human creativity involves some amount of imitation. But obviously there's more, otherwise all art would be... I don't know, cave drawings? In the last couple hundred years we've seen so many art movements that were not, in fact, derivative of what came before: abstract expressionism, surrealism, cubism, impressionism, etc. In music, we've gone from classical to modernist to jazz, blues, rock'n'roll, country, hip-hop, heavy metal, punk, electronica, etc.
There is no generative AI technology in existence that can perform what is fundamental to human creativity: making art that is, at least in part, genuinely new.
Thanks for the compliment, Ben! That means a lot! And I agree with your points.
Most tech leaders like to borrow these words to make the technology sound more impressive. It’s also easier to say its creative than to say “it mimics creative works based on a dataset of images that it has seen before, thus creating the illusion of creativity”. Such a mouthful.
Granted, but I'm looking at creativity in more than just art. fiction is creativity too. I think of hallucination like that. It is at it's base, creativity, not an error.
We are who we are now, as humans, because of everything that came before, especially the stuff important enough to remember or write down.
Interesting! So, while hallucinations are an intriguing way to look at creativity, it think there's two parts to human creativity that make it distinct: intentionality and learning.
When we make something, we make something with a reason, and this reason behind the artwork or a piece of fiction, is the why. AI models lack this intentionality. Their creativity is a product of randomness.
Also, when we create, we learn something through the act of creation. This learning part, where the act itself is giving something back to us, is lacking when it comes to AI. If an AI draws a thousand images, it has learned nothing new in the process of doing that, whereas a person painting a thousand paintings has.
Having been through image diffusion, and now messing about with LLM's I can understand where you are coming from. The real effect of AI in my life so far is to get me interested in humans again having given up as a teenager. I do think you're right about expressly human creativity. In as much as there is understanding involved. A child with felt tips is going to produce something more objectively human than any current "AI" model.
However, "AI" in the shape of Alpha Fold has already produced a glossary of compounds that can be produced, and that has been spun out to see if any of the compounds amount to more than "Grey Goo"
Meanwhile there has been almost no progress made in 40 years in Cancer research, (I was lucky to get the kind that submits to monoclonal antibodies.) Similarly with physics and maths, there have been some advances, some of the great challenges have been attempted, but in quantum theory not much has been achieved, again in 40+ years. At least on the order of Einstien, Bohr, Boltzman & Turing. This is also "creativity" as far an Novel theory goes. String Theory was the great white hope for a while but it's not empirically testable.
Thus I have hope than with access to the world's knowledge (peer reviewed papers) we may be able to get a model to conjecture based on existing research about what may be possible. You can see this in action, courtesy of Claude 3, with WorldSim and Websim. This is now with today's tech that has neither a World Model, Reasoning or understanding.
I just found your blog, love it so far.
I'd push even more strongly against Suleyman's claim that AI is "creative". It's not just that it was trained on a vast collection of human-made images; it's that training consists of the model "learning" to reproduce what's in the training data. The whole goal is to imitate what it's been fed!
Now, the usual response is along the lines of "that's what humans do, too". And, of course, human creativity involves some amount of imitation. But obviously there's more, otherwise all art would be... I don't know, cave drawings? In the last couple hundred years we've seen so many art movements that were not, in fact, derivative of what came before: abstract expressionism, surrealism, cubism, impressionism, etc. In music, we've gone from classical to modernist to jazz, blues, rock'n'roll, country, hip-hop, heavy metal, punk, electronica, etc.
There is no generative AI technology in existence that can perform what is fundamental to human creativity: making art that is, at least in part, genuinely new.
Thanks for the compliment, Ben! That means a lot! And I agree with your points.
Most tech leaders like to borrow these words to make the technology sound more impressive. It’s also easier to say its creative than to say “it mimics creative works based on a dataset of images that it has seen before, thus creating the illusion of creativity”. Such a mouthful.
"Today’s AI is entirely contingent on these precious sources of human creativity; without them, it wouldn’t exist."
You could say the same about us too.
Today’s human is entirely contingent on these precious sources of our creativity; without them, we wouldn’t exist.
Think carefully, though. You say that, but is that actually true? Without having seen an artwork in your whole life, you can paint and be creative.
In the abstract yes, but we are no longer cave painting to the Gods. Since the enlightenment we have been standing on the shoulders of Giants.
When a child is coloring in a coloring book it is infinitely more creative than an AI model.
Granted, but I'm looking at creativity in more than just art. fiction is creativity too. I think of hallucination like that. It is at it's base, creativity, not an error.
We are who we are now, as humans, because of everything that came before, especially the stuff important enough to remember or write down.
Interesting! So, while hallucinations are an intriguing way to look at creativity, it think there's two parts to human creativity that make it distinct: intentionality and learning.
When we make something, we make something with a reason, and this reason behind the artwork or a piece of fiction, is the why. AI models lack this intentionality. Their creativity is a product of randomness.
Also, when we create, we learn something through the act of creation. This learning part, where the act itself is giving something back to us, is lacking when it comes to AI. If an AI draws a thousand images, it has learned nothing new in the process of doing that, whereas a person painting a thousand paintings has.
Having been through image diffusion, and now messing about with LLM's I can understand where you are coming from. The real effect of AI in my life so far is to get me interested in humans again having given up as a teenager. I do think you're right about expressly human creativity. In as much as there is understanding involved. A child with felt tips is going to produce something more objectively human than any current "AI" model.
However, "AI" in the shape of Alpha Fold has already produced a glossary of compounds that can be produced, and that has been spun out to see if any of the compounds amount to more than "Grey Goo"
Meanwhile there has been almost no progress made in 40 years in Cancer research, (I was lucky to get the kind that submits to monoclonal antibodies.) Similarly with physics and maths, there have been some advances, some of the great challenges have been attempted, but in quantum theory not much has been achieved, again in 40+ years. At least on the order of Einstien, Bohr, Boltzman & Turing. This is also "creativity" as far an Novel theory goes. String Theory was the great white hope for a while but it's not empirically testable.
Thus I have hope than with access to the world's knowledge (peer reviewed papers) we may be able to get a model to conjecture based on existing research about what may be possible. You can see this in action, courtesy of Claude 3, with WorldSim and Websim. This is now with today's tech that has neither a World Model, Reasoning or understanding.
My children draw and sing without having been taught anything.
But your children benefit from 35,000 to 100,00 years of evolutionary genetic adaption of the human cerebellum. All "AI" has is Norbert Wiener :P
Birds are also able to sing and build bowers innovatively with materials without said cerebellum.
The missing ingredient is love.
Birds have been on earth for approximately 150 Million years. "AI" has been a thing you could "train" since the 1990's.
AI is far more likely to just kill us all, but not before first making us miserable.