Automated Philosophy · Christopher Audretsch

Here’s an informal, note-y, working description of this project. Feel free to contact me if you have any ideas about this.

Fine-tuned LLMs

I fine-tune large base (non-chat) models, including Kimi-K2 (1T parameters), on a corpus of high-quality philosophy documents. Currently the models are fine-tuned with a LoRA on a corpus of ~5 mil tokens from ~6 authors including David Chalmers, Scott Alexander, Joe Carlsmith. Each document contains a header that goes ~~Title: [title] | Author: [author] | Date: [date]~~. You thus use a title, author, and date (probably current date) to prompt the fine-tuned model. The hope with this project is that this model can be used to generate actually interesting, original, valuable philosophy.

(Why? Claude is good at writing like Claude, but it’s significantly worse than its base model at writing in other voices and at imitating other writers. Ultimately, we don’t get a wide variety of philosophical ideas with Claude, and we succumb to a kind of assistant basin of ideas. Can a base model fine-tuned to write like David Chalmers and other great philosophers do better?)

(BTW, I’ve done some other fine-tunes which worked out pretty well. Namely, I’ve fine-tuned on high-quality Wikipedia articles, on David Foster Wallace, and on my own writing. To be clear, these are three separate fine-tunes. I think there could be utility to some of these… and at the very least they’re interesting and entertaining. I hope to play around with them more in the future. If I figure out the pipeline and judging, I could put these models in a pipeline!)

Update: While I did have some promising training runs with K2, I’ve been fine-tuning posttrained models (Kimi-K2.6) because there are no providers for large base models, and accessing GPUs for Kimi-K2 has been difficult (often there’s no availability) and expensive (I’m doing this all out of pocket — I should try to get funding). I use these fine-tuned posttrained models like a base model; there’s no User/Assistant structure.

Automated philosophy pipeline

I use a pipeline to write philosophy with those models. I generate a philosophy article in chunks à la loom, using a judge to pick the best completions. Once I have a completed article, I then simulate a dialectic. I have the LLM generate a critique of the first article in the style of a different author than the author of the first article. This can go back and forth additional times. Finally, I generate a synthesis article from the two articles.

Judge

The automated judges I’ve tried so far have been disappointing. I’ve tried… Claude Opus 5.0/5.5, GPT-5.6 Sol and 6 Astra, these models with various prompts, a judging process that uses multiple LLM instances and has some instances dedicated to specific things like factual correctness or interestingness or adherence to voice/author, having multiple LLM judges vote on the best completion, using the fine-tuned model itself with a prompt scaffold. Of course, in order to evaluate the results of this, I have to judge these texts myself! This is quite time and cognition consuming. At least it’s somewhat interesting in its own right, and also contributes to a dataset of “ground-truth”-judged articles which I can use to evaluate judges. (Though I’m just a guy who knows a lot about philosophy, AI, and writing. What if my judgements are part of the problem!!)

So this project basically consists of three parts:

Model
Judge
Pipeline

Each part has many possibilities, many variations. While each part if sufficiently good can shore up some amount of weakness in the other parts, I ultimately expect there won’t be good results unless both the model and judge are very good. More experimentation needed. Perhaps more theory is needed, too. This is all an ongoing project!