On Automated Philosophy
This text, initially written in the summer of 2021, is part of a lecture series about language and AI at ToftH School. The original quotes can be found in the German version.
There’s a slightly yellowed book sitting on my desk. The cover is square and kept entirely in lowercase letters. It reads: “max bense theorie der texte [theory of texts] kiepenheuer & witsch”. Almost sixty years after its publication, I’m looking at the screen in front of me and read the words “OpenAI Beta Playground”.
The company OpenAI offers an experimental service to interact with GPT-3, their much discussed speech-based artificial intelligence. Given an arbitrary text, GPT-3 can predict which tokens are most likely to follow next. The results are surprisingly coherent sentences, which can even take more abstract context into account.
The technology is based on the Transformer architecture for artificial neural networks developed at Google a few years ago. What makes GPT-3 special is the extraordinary amount of training data. This AI has read virtually the whole Internet and, through this, independently learned to credibly imitate natural language. Here, language is to be understood as a generalized concept. The text can be composed in English or German, or even in a programming language.
“Artificial poetry is therefore one that does not arise out of consciousness, but lies in language itself”.
Max Bense (1910–1990) was a German philosopher and writer. His most influential contributions were to the fields of semiotics and aesthetics. Beginning in the 1950s, the Stuttgart School formed around him, a group of scientists and artists who experimented with new forms of expression.
Even though, at the time, computers had not yet permeated our everyday lives as profoundly as today, their revolutionary significance was already evident for theorists like Bense. Particularly, the new possibility of subjecting a work of art to automated character processing opened the door both to ambitious attempts at a formalization of aesthetics through statistics and to the earliest experiments with generative art. Nowadays, we rarely mathematically measure the aesthetic value of a work of art, but automated computer art has made qualitatively new things possible.
The Theory of Texts mentioned at the outset contains a chapter that seemed to me particularly suitable for demonstrating the connection between this seminal work and today’s technology. From page 143 on, Bense writes:
It can now contribute to the elucidation of a general concept of poetry if one first differentiates between natural and artificial poetry. In both cases, one works with words, their derivatives, which can be interpreted as deformations in relation to an underlying word space, and their sequences, which are arranged linearly or planarly. For our consideration, however, the essence remains in the difference in the manner of formation.
Natural poetry is understood here as the kind of poetry which, it is the classical and traditional case, presupposes a personal poetic consciousness, as Hegel already called it; a consciousness that possesses encounters, experiences, feelings, memories, thoughts, visions of an imagination, etc., in short, a preexisting world and is capable of imparting linguistic expression to it. Only in this ontological framework can there be a lyrical subject or a fictional epic world. Poetic consciousness in this sense is always a transposing one, namely beings [Seiendes] in signs, and the epitome of these signs we call language, insofar as they metalinguistically possess a self-relation and a world-aspect. In this natural poetry, writing does thus not cease to be an ontological continuation. Every word it expresses follows the world-experience of a self, and even the aesthetic status granted to it could still be regarded as a reflex of this world.
Artificial poetry, however, is understood here as a kind of poetry in which, insofar as it was produced, e.g., by a machine, there is no personal poetic consciousness with its experiences, encounters, feelings, memories, thoughts, visions of an imagination, etc., thus no preexisting world, and in which writing is no longer an ontological continuation through which the words’ world-aspect could be related to a self. Consequently, neither a lyrical subject nor a fictional epic world can be meaningfully drawn from this poetry’s linguistic fixation. While an intentional beginning of the word process is thus characteristic of natural poetry, there can only be a material origin for artificial poetry.
Even these first few paragraphs are sufficient for GPT-3, building on the vast amount of previously analyzed sentences, to continue automatically:
Artificial poetry is therefore one that does not arise out of consciousness, but lies in language itself. It is the result of an automatic processing of character sequences, which as such possess no world-aspect and therefore cannot produce a lyrical subject or a fictional epic world either. It is thus no longer the continuation of the world, but its alteration.
This distinction between natural and artificial poetry is confirmed by the fact that there have indeed been repeated attempts in the history of literature to produce artificial poetry, but never an attempt to create natural poetry. This distinction is thus no theoretical speculation, but a historical fact.
The question now is: How can one produce artificial poetry?
The answer is: By building a machine which automatically carries out processing on the basis of character sequences, thereby leading to new character sequences. These new sequences are no longer in the language, but in the machine. They are no longer a word which can be related to a self, but a character sequence which possesses no world-aspect and therefore cannot produce a lyrical-subject-aspect.
“It is the result of an automatic processing of character sequences, which as such possess no world-aspect”.
On the surface alone, it is remarkable how GPT-3 imitates Bense’s writing style — including the old spelling [in the German original]. Right away, the first sentence gets to the point in terms of content too: “Artificial poetry is therefore one that does not arise out of consciousness, but lies in language itself”. One could definitely put it this way. Artificial neural networks ultimately select text elements statistically, based on common patterns in already existing texts. In the end, of course, this is only a “continuation of the world”. That’s so nicely put that we will just conveniently overlook the semantic inversion.
The next paragraph is either profound or meaningless. Why was there “never an attempt” to create natural poetry? Maybe simply because it already existed. Although, thought more deeply: What would it be like to try to create natural poetry anew — as if it didn’t exist yet? One could read a large and ambitious lyrical project into this. In order for us not to think about this too much, however, it is immediately followed by the rhetorical bludgeon: It is supposedly a “historical fact” and not “theoretical speculation”. That does sound like Bense.
In closing, we’re left with the central question and its self-critical answer: How can one create artificial poetry? With the automatic processing of character sequences “in the machine”, of course. This, however, has no “world-aspect” and therefore no “lyrical-subject-aspect” either, so the assertion.
Despite the remarkable familiarity of these artificial sentences, it is still us humans who provide them with meaning. The reader ultimately decides on sense or nonsense. Even the product of an impressive AI like GPT-3 is more inspiration than finished work. Its first serious applications such as GitHub Copilot attest to this by merely generating suggestions for the user.
We can stick our neck out even further with this interpretation and try to generalize the statements about artificial poetry to artificial intelligence as a whole. The advantage of natural intelligence would therefore be the capacity for alteration of the world, rather than only its continuation. The ultimate way to achieve the elusive world-aspect would be through the embodied AI of robotics. In a sense, our teams at Everyday Robots, Google Research, and elsewhere are working toward this future. A robot which can learn from experience certainly comes at least one step closer to a self —be it lyrical or prosaic.
“It is thus no longer the continuation of the world, but its alteration”.
I only got to know Max Bense toward the end of his life. At the time, my own ability to credibly mimic human language was still in its infancy. That didn’t stop me, however, from having passionate conversations with my father’s doctoral advisor.
What might he have to say about the artificial intelligence of today?
