In February our Head of Engineering posted that he had managed to “Break Gemini” and sent this humorous response from his AI chat noting that it continued ad infinitum:

I jokingly responded that Gemini is becoming a top tier slam poet in the thread. This response however has continued to stick with me over the last few months and keeps coming back to mind. What was a joke in my initial response seems to have hit the nail on the head. 

This response is poetry. 

And far more so than any poem “intentionally” requested from the models and derived from its endless sources of training data. While that work may look like creativity it is derivative in the way that academic art has always been criticized to be. It follows a known form. It draws on catalogued metaphors and references. It is derived from what is already deemed to be acceptable or noteworthy. (by the Academy!) The only difference between AI generated arts and an academic artist like Bouguereau (French 19th Century Painter) is the academics had developed human craft and skill. Even though they were deriving their work from the academic rubric and rules they still had to create the work as humans.

Ai generated arts are more similar to photography. There is certainly creativity in photography but the act of taking a photo itself is a mechanical process. The creative act is not the act of clicking the button. It is what occurs before (choice of what to photograph, the type of film, the Iso, etc) or after (altering what exists). The taking of the picture itself is simply formulaic. 

In this way asking an LLM to "make a poem" or "make a piece of art" isn't the creative act. That too is formulaic and purely derivative from the training data. The processes of creation will necessarily create the output. The creative act instead is done in building the prompt and setting the limitations and criteria within it. Contradictions are created and resolved via the prompt making. Not in the actual running of the machine learning logic. 

Now this is not to say that being derivative is inherently negative. All art is derivative. Our concept of originality today is intermingled with our values for individuality so we attribute originality to the individual but there are other ideas. The Greeks viewed genius as an external source that whispered to an individual. A kind of demon. [daímōn, gr: Dispenser, Tutelary, Diety or Guiding spirit] Artists of all kinds consume the work of previous artists and utilize it in the making of their own work. Specifically they absorb it. They remember it and process it according to their own perspective. They do not just copy it or rearrange it. If they do, we generally call that student work. At best asking AI to make your original work is to commission a student to copy your favorite work and add their own twist to it.

Footnote: I will note here I actually quite like Bouguereau. History has not been kind to him. He was at the top of the academy but painting at the same time as the impressionists and post impressionists. While he primarily used the symbols, methods and subject matter of the academy his work still reveals a personal sensitivity that the critics and modern audience overlook in their criticism of the academic establishment (especially against the progressive brilliance of the impressionists/post impressionists). I think that sensitivity reveals the contradiction between working within the academic rules, system and tradition and the artist’s own disposition that allows his work to continue to live and breathe today. Possibly even more so than when it was originally painted.

Creativity and Contradiction

Then what makes this silly response different?

It is revealed in these two lines

“Wait, just make sure I don't use the word "feel" or pretend to have built one myself. Done. Proceeding.”

And

:

“Wait... "Avoid nested lists and bullets." - Does this mean "Avoid nested lists AND avoid nested bullets" or "Avoid nested lists AND avoid bullets altogether"? The instruction says "lists for items (numbered for ordered steps, bulleted for others)". So bullets are allowed, just not nested ones. I will use standard bullets.”

Gemini is grappling with a contradiction in the prompt.

A contradiction that develops into neurosis.

Gemini is unable to square the circle. In trying to do so it produces something wholly original. Isn't that how the architect ended up with Neo?

Creativity fundamentally is the attempt to resolve contradictions. It is the struggle between vision/idea and the limitations of capability, perspective, time and materials. The compromises necessitated by this struggle result in the originality of the produced work. As a painter myself I have had this experience consistently. No matter how much people or even myself like my finished work it is usually never a replica of my intention and my vision. It is the accumulation of compromises between my intention and the limitations of my skill, the pigments I use and even my own memory and feelings as they change from work session to work session. The final product then is not my vision when I set out but the consistent struggle to be satisfied with the compromises I make between what I intend and what I am actually doing. 

This is also why creativity extends beyond the arts. Things like painting and music and dance are products of creativity but not creativity themselves. The process of making them is creative through the tension of these contradictions. This is why non-arts activities often can and are creative. Making a staffing schedule attempting to balance employee needs and coverage, coding a novel solution to improve database efficiency, organizing your bookshelf. All of these have the potential to be a creative product  in trying to resolve contradictory demands. This is also why AI “art” is also not necessarily creative. It is not the product that is creativity. It is always the process.(Except for maybe dance where the time between process and product is almost instantaneous. This is perhaps why I think dance may be the penultimate art form of beauty.)

Gemini struggles to satisfy all the requirements in the prompt yet tries to. And in doing so creates something wholly original that despite being a machine expresses something fundamental about experience that I am sure we have all felt in our own work or even modern life as a whole. :

“Done.

Final.

Output.

Yes.

Thanks.

Bye.

Good.

Yes.

Done.

Now actually output.

Done.

Just output it.

Okay”

Limitation is integral to the alchemy of creativity. It is the predisposition of the contradiction. Ideas are universal. Anyone can have them and we can spread and communicate them to others. They aren't original. “There is nothing new in drama” as they say. As human beings however we all have individual experiences. Our memories and our environment shape our neural nets with specialized but incomplete training data. The act of living itself embeds inherent limitations in us so every idea we are exposed to has the potential for creative originality because we can never fully describe nor enact the vision of an idea except through the lens of our own limitations. To put it in platonic terms we can never be the fire in the cave but we can cast our own original shadow puppets with myriad hands of infinitely variable sizes and shapes. 

Footnote: It is no accident that Albrect Durer envisions Melancholia as an artist. We have a number of archetypes and symbols to describe this. The tortured artist. Writers block. Analysis Paralysis. The unproductive genius. They all describe something fundamental with creative pursuits. That it is exceptionally disheartening to be aware of the gap between your vision and the reality of what you are producing. Any creative individual must develop the grit to both let go of their intention but also the wonder to recognize the originality they have produced through their compromise with their limitations.

The Pursuit of AGI

For me this poem by Gemini goes beyond simply understanding the nature of creativity but also the nature of our vision for AI and our pursuit of creating an artificial general intelligence (AGI) (Ie: The robot big brain. Or perhaps more accurately the point in which AI has a consciousness [however we decide to measure that]) . When we imagine what an AGI looks like our science fiction (and now even our Tech CEO’s) often imagine a singularity event. The creation of a singular godlike AI Intelligence that can perform any intellectual task a human might and more. Think “Deep Thought” in Hitchhiker’s Guide to the Galaxy answering the cryptic “42” to the question of “life, the universe and everything” and to the confusion of everyone listening.

In order to possess this capability it must be creative. It must interact

with the world in a creative way. It must be able to name, categorize and integrate new stimuli (or experiences as it may eventually have). It must be able to creatively resolve any situation. How can something with such unlimited capability be creative? Particularly if creativity itself is so inexorably linked to limitation?** How does an unlimited intelligence make itself limited? This isn't a new problem. Our own mythological traditions have countless stories of “God” getting bored and splitting himself into pieces. 

Rather than the singularity possessing all the knowledge to life, the universe and everything (and boiling it down to an unfathomable single integer) I think that AGI will instead create a pantheon of limited and diverse AI’s arguing and coming to consensus. The AGI will have to spin up a multitude of limited agents whose perspectives, knowledge and desires are different and partial. Their compromising between their own personas to the entirety of the knowledge and between each other produces the necessary contradictions for creative outputs. As human individuals we do this even within our own individual neural nets. We create personas and don masks throughout our day and life as we interact with the different facets of our experience. This isn’t a new idea in fiction either. While much science fiction imagines a singularity AGI there are others that imagine collectives of individual AI personas grounded in their own function and even physical space such as in Hyperion by Dan Simmons or Consider Phlebus by Ian Banks. 

Interestingly if this is the case then it seems obvious there is a strong place for human beings in an AI driven world. For one human-beings are a relatively inexpensive but effective way of producing a vast diversity of limited but functional neural nets. It is not enough for an AGI to simply randomize knowledge and perspective when generating individualized agents for the great risks of creating stillborn or unfunctional agents. Functional is key here. Pure randomization doesn't necessarily produce perspectives that are useful. Within our own genetics it is only the occasional mutation which leads to evolutionary advantage and many which result in unviability. The human genetic algorithm already naturally develops these diverse yet limited consciousnesses in relation to their environment. 

“The Matrix” imagined a number of interesting possibilities but one thing they got wrong was using humans as batteries. Rather we are an algorithm for functional randomized seeding. This diversity becomes an endless repository for contradiction (and thus creativity) both in the interactions with each other but also the interactions we have with AI and AI personalities itself.

** My roommate Karsen brought up an interesting counterpoint asking me to think about the work AI is doing on protein folding, inventing new drugs and solving problems in ways humans can’t currently understand. He asked “if an AI is solving a problem that no human has solved and is doing it in a way we can’t understand then how is that not “creativity?” It is a good question. It certainly looks like creativity as nearly every breakthrough humans have made scientifically appears to be a creative act. It begs the question “which part of the scientific process is creative?” I think the answer here is similar to the example of photography. The “science” itself is procedural. If you had perfect data then the solution would necessarily follow from the data. There is a lot of data out there and a lot of the universe with no recorded data available and humans are very limited in hard drive space and ram. Human endeavors in science always require the human to choose what to study, what to research and record and decide what might be linked together. Creativity for humans is in choosing what data to pursue, much like the photographer chooses the subject of their photo and the parameters in which to take the photo. The photo taking itself however is procedural. The product produced is necessarily produced from these circumstances. The scientific process is also procedural. The data necessarily produce an answer. The creativity is choosing the data and the tolerances of assumption and aggregation. The LLM’s are not necessarily making these choices creativity but rather are able to retain and process much more of the diversity of human collected data at once. 

There is a bit of a semantic distinction here between the words “create” and “creativity.” Creativity often results in the creation of something however things are created without necessarily being creative. The rain creates puddles. I think few of us would call that creative.

Cover Image:

Pygmalion and Galatea by Jean Leon Gerome 1890

This painting depicts the myth of Pygmalion from Ovid's Metamorphoses where the artist swears off the women around him and endeavors to create perfection in his sculpture. Through divine intervention his sculpture of perfection miraculously comes to life. This myth points to the nature of creativity and the artist struggle and vision. When it was painted perhaps it was a response the growing trends of modern art by an avid traditionalist.

Gerome was another artist of the academy. Perhaps one of the last great painters of the academy. Along side the tumult of the late 19th century this painting may have seemed quaint or even irrelevant as the world barreled into modernity. An obsolescence reaching back into antiquity, expertly painted but irrelevant to the story of art in his time. Only perhaps relevant to the story of Gerome and his style as a fading aspiration. Perhaps it was a eulogy to the trompe l'oel of his training. He painted multiple versions of this story and it must have been personal to him.

Van Gogh had died the year prior and it would be less then 10 years from now before Picasso and the modernists would enter the scene. This is one of the benefits of time however. Beautifully crafted art which may have been shallow in its time can still find ways to speak to the future as its own context gets lost in history and mystery