I. Writing with LLMs

As part of my private writing project, I intend to use various AI language models (LLMs) to generate targeted philosophical texts that are inspired by Friedrich Nietzsche in terms of style and content. I'm not just interested in philosophically sounding texts with a few Nietzsche buzzwords, as is currently happening with simple prompts on ChatGPT. My goal is to obtain differentiated texts that are based on specific phases of work or types of texts (e.g. maxims, aphorisms or letters) in literary, content and contextual terms. To this end, I train AI models specifically with Nietzsche texts in order to capture stylistic and rhetorical peculiarities more precisely.

The following texts were generated using the ChatGPT-4O model using “Instruction Tuning”. This means that I have achieved more and more the desired result with selected examples and many prompts. This is often sufficient to generate individual and short text examples, while for systematic generation, a model is trained with larger, processed amounts of data through “fine-tuning”.² For laymen, the technical set-up of locally trained models is still associated with time-consuming imponderables and depends on powerful hardware. In general, it is advisable to carry out tests that are as manageable as possible in order to be able to understand the results and optimize them accordingly in a controlled manner.³

‍

II. Nietzsche-Aphorisms about Sorrento and Sicily

One obvious type of curiosity when imitating is to ask yourself what someone might have thought or said about something. I have often worked with landscape aphorisms because they can be formulated more openly and do not necessarily have to follow a concise thesis. In this way, I tried to generate an aphorism for the Italian city of Sorrento. I have searched three books that are close together in time for suitable passages on landscapes and hiking: the first volume of Human, All Too Human (April 14, 1878), Assorted Opinions and Maxims (March 12, 1879) and The Wanderer and His Shadow (December 18, 1879). The results were bumpy at first. I suppose because Sorrento is a specific city and therefore the description had to fit this concretion. The titles were also aphoristically ineffective. In addition, ChatGPT's talkativeness got in the way, which is why I intervened briefly a few times to reach the following result:

Surrentum ex umbra. — The south in retreat, a corner where even the light takes a break: cool, shady, yet quietly demanding. The rocks are almost weightless, leaning on — tired perhaps, or wise. Everything seems half said here. The wind whispers of the past and silent grottoes dream of sirens that have long since fallen silent. Here, where every thought is beguiled by limes and oranges, aromatic scents. Here where only the colors are clear — the thinking fables.

This result had to be optimized. I did like the basic mix of topics — southern Sorrento, the shady trees, the rocks... The contrast of clear sensual colors and beguiling spiritual thinking even gave the text some aphoristic punch line. However, the orientation was based on the aphorism “Et in Arcadia ego” from Wanderer and his shadow (link) too strong: Too often, the pattern “Here, where...” is used. In addition, the metaphor of the leaning rocks for Sorrento seemed inconclusive to me. Through suggested abbreviations and densifications as well as by searching for a title, I came to the following result:

Fugit lux, Surrentum device. — The South is retreating from itself. Here, where even the light stops — cool, shady, yet challenging. The rocks are half-high, straight and almost weightless: not falling, not defiant — but grown old, tired and clever. Everything is half-loud here, Half said. The wind whispers about the past. The caves dream of the sirens echoing. And in between: penetrating scents of lemon, salt, sun.

I really liked the stilimitation here, although the meaning of the content became leaner, especially in the final sentence — especially since a single aphorism must be coherent compared to a series of aphorisms. The two examples are therefore only intended to illustrate what an LLM can achieve through imitation and how this can be promoted with prompts, objectives and suitable materials.

‍

In another series of experiments, I asked ChatGPT to generate an aphorism about Sicily. Nietzsche did not write one of these either in Messina or anywhere else — and yet it seemed almost like a gap to me that there is no such thing in Nietzsche's works. The following aphorism has thus been generated over several stages of revision. For the “Instruction Training”, I also used examples of other Nietzsche locations, letters from Messina and a few excerpts from historical travel guides from Nietzsche's library:

Sicily. — On Sicily's soil, two powers are fighting for the wanderer's soul: there Mount Etna, a symbol of Dionysian fire, an everlasting passion — here the temples, heralds of Apollinan clarity, beauty and harmony carved in stone. Only those who have the courage to purify themselves in fire are able to climb the heights of pure knowledge and thus be truly human in harmony with the divine. Many people burn themselves up during this venture, burn up in excess of emotion — but who wanted to talk them out of their affirmation, which derived their right from existence?

It can be seen here that GPT-4o can work very well with pairs of terms: Apollonian powers, temples, recognition vs. Dionysian fire, volcano, feeling. However, the phases of work are mixed here, as the Nietzsche of 1878/79 no longer argues so strongly with Dionysian and Apollinian. Since there were no quotations from early Nietzsche in my instructional examples, it is clear that ChatGPT added some elements from Nietzsche's philosophy. This shows that LLMs tend to produce a generic or blending work phases, which they calculate based on their training data. This weakens the result from the point of view of a plausible imitation of the work. The final sentence, in which a new punchline should lie, was also always difficult. This was achieved reasonably well only after a few attempts.

‍

III. Two Generated Nietzsche-Maxims from the Middle Phase

I consistently got better results when I settled on short forms and a style: be it letters, aphorisms, or maxims. With a selection of maxims from Assorted Opinions and Maxims (AO) I then had GPT-4o generate a new spell. Usually two or three so that I could choose one for further use. I liked the following two maxims:

Humans are nature that is ashamed — and culture that apologizes.

People flicker between instinct and virtue.

The titles of the maxims were once again difficult to generate with the same prompt. With a few inquiries and sample texts, I think it was then easy to deliver or correct them. In the following example, the first title “Windbreak” was replaced by “Individual”, which looks concise and appropriate:

Individually. — Some things fall not because they are weak, but because they are free.

This maxim is thought-provoking, makes sense and can be read extensively several times. In particular, the ambiguous “free standing” (standing unprotected, being alone, being free, etc.) invites different interpretations. The maxim is compatible with the game of separation (“individual”) and freedom (“free standing”), which has its price (e.g. free spirits in Human, All Too Human), and also a perfect fit for the Middle Nietzsche context.

Another procedure was to have the LLM combine original passages from AO into a new maxim. Due to the original material and the ambiguity of the original maxims, astonishingly good results can be generated here. I found this very successful:

Silent duty. — Anyone who works in the shadow of the big one doesn't know the brilliance, but the weight.

‍

IV. Conclusion and Outlook

My conclusion is: With suitable sample material and instructions, LLMs can aptly imitate Nietzsche's style in individual short texts. As complexity and volume of text increase, it quickly becomes more difficult to generate meaningful texts — for example, a series of 10 related aphorisms. The connection of work phases, styles, and contemporary contexts to Nietzsche's work significantly increases plausibility and makes the results more interesting. However, the coherence of work phases in particular is difficult to achieve due to the already trained, generic Nietzsche styles of prefabricated LLMs. This speaks for LLMs own finetunings. My technical skills and time options have so far been limited here: The Nietzsche generators trained by me through fine-tuning have so far proven to be poor compared to “instruction tuning” with leading models such as ChatGPT or Claude. However, these time-consuming fine-tuning sessions have helped me to understand better and better understand exactly what an LLM does and how I can guide it through prompts. In addition, you get to know philosophical works from a different perspective when you have them processed through an LLM — this learning effect should not be underestimated, especially for people who have previously dealt exclusively with qualitative and interpretive processes. As fine-tuning is becoming increasingly easier and more accessible, I expect a lot of potential here in the medium term.

At this stage of the technical development of LLMs, I am interested on the one hand in exploring the possibilities and on the other hand in the attitude of (human) readers, who always read and interpret through the lens of their author's ideas. I have therefore deliberately not yet addressed the opportunities and risks for Nietzsche scientific research. It is only possible to record for the moment: LLMs open up numerous opportunities to experiment with philosophical texts. Such experiments may seem pointless to some because they are not “original” quotations or because they do not consider calculated texts to be philosophically relevant. These reactions show how our ideas of authorship, originality, or human origin shape the philosophical idea of authorship. These formations move in the flow of time. With regard to pseudo-Aristotelian writings, pseudepigraphy or the anonymously published texts of the Enlightenment, I would not rule out that the close view of philosophical authorship and historical-critically edited “complete works” will change.

‍

Tobias Brücker has a doctorate in cultural studies and is head of HR personnel development at the Zurich University of the Arts. He has researched Nietzsche's working methods and published in 2019 the monograph On the road to philosophy. Friedrich Nietzsche writes “The Wanderer and His Shadow” ("Auf dem Weg zur Philosophie. Friedrich Nietzsche schreibt 'Der Wanderer und sein Schatten'"). He is interested in all facets of diets, authorship, and creativity techniques in philosophy and the arts.

‍

To the article image

“Nietzsche—Salomé” (AI-generated image, Tobias Brücker, 2024): Created with Midjourney based on the following historical photographs: Friedrich Nietzsche (ca. 1875, photograph by Friedrich Hermann Hartmann, public domain) and Lou Andreas-Salomé (c. 1897, studio Elvira Munich, public domain).

‍

Footnotes

1: For example on character.ai | AI Chat, Reimagined—Your Words. Your World.

2: The complex work involved in creating and fine-tuning a Nietzsche bot with various full texts from different phases of work is documented here: Building an Advanced Nietzsche AI Database | by Wayward Verities | Medium. The resulting “Nietzsche Reference Bot” makes it possible to interact with Nietzsche's full texts via chat and receive referenced answers, see here: https://chat.openai.com/g/g-F62wnKW8A-nietzsche-reference-bot.

3: Valuable insights into “instruction tuning” can be found in this experience report: I used AI to generate Nietzschean aphorisms | Towards AI