Generative models are pushing the limits of what our machines can do. One of the sign of this push is perhaps the pressing need to redefine what differentiates us as human. A definition is needed from an identity standpoint: what defines us. But it also makes sense in the scientific and technical world: what is the next boundary?
As a side note this question is coming back so often in history, biology and artificial life. I guess wisdom suggests an answer: “in the long term: nothing”.
Turing tests
Turing tests have been nice tools to think about performance in artificial intelligence. Can a machine pass for a human? I am using “tests” as a plural, as they can have many flavors:
- For how long?
- At what task? E.g. chat, voice discussion, art generation
- With curation? Can a human curator pick the best answer from a limited set?
Curation is actually what today makes things looks a lot better that what they really are. We can still be thankful of the crowd to actually weed out the 50 - 90% garbage that is produced. But saying “I like this picture” is a task that we can agree is way simpler than making the picture.
Let me quote you “I, Robot”:
Spooner: “Can a robot write a symphony? Can a robot turn a canvas into a beautiful masterpiece”
Sonny: “Can you?”
The joke is on us. I can still not write a symphony, or decently paint or draw… But most of my tries at graphic designe are trumped by all the current generative models, Midjourney, Stable Diffusion and the likes are mind-blowing. Strangely enough before “I, Robot” was released, David Cope generated music through his EMI program. And it is amazing.
If our Turing test is “creating an art piece with curation”… those systems do pass the test. And if anything, such a test would demonstrate super human abilities, those systems are much faster than any human and they cover a range of skills (through different styles) that no human has, including experts.
The question is what is the boundary between what experts can and models can not achieve. For now.
Stochastic parrots
“Stochastic parrots” is a dismissive term that has been brought to my attention, perhaps even coined, through a paper from last year by Emily Bender et al. 2021 to discuss about text generative models such as BERT and GPT-3.
The idea behind this obscure jargon is that those models are learning what makes a sentence or a paragraph meaningful to us. It is in a sense a way of modeling a statistical distribution. That’s the “stochastic” part.
Once models have learnt this, you just run them with an input and a few random parameters, it generates new things. Those new things might feel novel for a human. Yet from the model standpoint they are just coming from the learnt distribution, that is just repeated in a different context, that’s the “parrot” part.
Is it really learning? In my understanding, yes.
Is it creativity though? It covers a lot of what we call as creativity.
What is left for humans
Van Gogh was a prolific painter, he produced about one painting every four day during ten years. What is considered of his contributions?
The paintings themselves? Style transfer models were already incredibly good at producing Van Gogh paintings. His style was particularly good for it, so the researchers showcased it a lot. Modern image generative models seem to produce better images as soon as you add “ArtStation” or “4k” or “dramatic” in your prompt. It has just learnt the meaningful difference between toddler drawings and mastery.
And that, leads us to the point of this essay.
Style …
All the above mentioned models are wonders when it comes to recombining meaning and style. They have first a good definition of what a given style is and how it can be connected to a meaning (or an input). But they are not able to come up with a new style.
They will show you a vision of anything you want from a news event, to a medieval fantasy scene in Dali’s style, but they will not invent a new style that will be recognized as aesthetically pleasing by humanity like Dali’s was.
So to go back to my initial definition, the key in our updated version of the Turing test is to tweak the task, “generate a style recognized by a community of experts as both novel and valuable”.
That is a test perhaps only a handful humans can succeed on the planet for a specific domain. A feat that is only achieve by a handful in each generation.
Perceiving technology as magic does not help us at understanding it … yet I could not resist to integrate a personification created by Stable Diffusion as a header of this article. Sorry .
The glitches you can notice on the texture are man made though.