Philosophical AI

Title:

Philosophical AI

Proposal in one sentence:

Niche AI Model Trained on the podcast: ‘History of Philosophy Without Any Gaps’

Description:

I want to create a custom GPT chatbot supplemented with vectorized texts pulled from Peter Adamson’s podcast: History of Philosophy Without Any Gaps. It would be similar to the Huberman AI (https://huberman.rile.yt), which responds to prompts using Andrew Huberman’s podcast but instead would be trained to quote niche philosophical data.

I would be using huggingface and gradio to get a quick version of this up and running within a couple weeks, but I would hope to eventually switch to a separate domain after working out the core of the project.

I’ve already contacted Peter Adamson directly and have a clean dataset to start working on this problem, potentially making the process much cleaner and quicker than the Huberman AI project which had to use Whisper to transcribe every episode individually.

There are a few other philosophical datasets that I have vague ideas about implementing, like the Stanford Encyclopedia of Philosophy, or various lectures, novels, etc. Potentially I could expand past this original idea over a longer period.

Deliverable:

Interactive “History of Philosophy Without Any Gaps” AI website hosted on huggingface using gradio front end.

Database of vectorized philosophical texts.

Squad:

Parker Barrett (https://twitter.com/parkerminii) (Discord: minii#5947)

3 Likes

Hi Parker,

I feel stoked by your project. Philosophy and AI, what a wonderful (and rare) combo! Will your output be open source?

Once you built it, could the tool be easily re-used for a philosophy garden featuring other philosophers too?

Are you familiar with Stephen Reid’s turning a lot of podcast episodes with another philosopher, Daniel Schmachtenberger, into a knowledge graph? Could something like that be useful to your project?

2 Likes

Thank you! But my name is Parker just for reference.

Everything should be open source except for the data fed into the bot.
It should be extremely easy to re-use and add other texts or podcasts.

I just looked up the knowledge graph you’re mentioning, it’s definitely a useful thought! It seems necessary that the more we rely on AI for complex problem solving, the more we’ll want it to explain and quote. I wonder what visual system communicates this “proofing” best. I thought the Huberman AI was a good basis for understanding where knowledge is pulled from, but could be expanded by adding books with chapters, other podcasts, secondary sources, etc. As to create a general ‘academic quotation’ of information in an output. But it may also be necessary to explain the relation of ideas, like the knowledge graph you mention does visually. I’ll be thinking about this for a while, thank you!

Since I’m passionate about democratizing complex thought ecosystems by knowledge visualization, whenever you need a thinking partner to your process, I’d be available.

Cool project. Is the code behind the huberman.rile.yt and/or the training setup available to you? Maybe if you ask they will collaborate. If I understood you are going to fine-tune one of the GPT models, you might be able to fine-tune one of the instruct models via the OpenAI API. Do you know which model is behind the Huberman AI?

It is unclear to me why you would want to deliver the DB of vectorized texts. It might be better to keep the original text so future methods for vectorization can be used. I guess the vectorization is dependent on the model you are fine-tuning. It might be good to identify which model that will be.

I have not fine-tuned a LLM but from what I saw on the OpenAI API it looked like it expected a prompt and response in the fine-tuning data. Will you need to provide prompts?

Good luck with the project!