Hey, everyone!

Today, I’m here to share some ideas that have been fermenting in my mind for the past few months.

First things first, check out this trailer I put together. It visually captures a journey that began back in 2022, when the generative AI revolution exploded, and brings us to… well, to where we’re heading.

What makes a character… themselves?

Most creators don’t spend much time philosophizing about the ontological nature of fictional characters. And why would they, right? A character is a character. They exist on pages, on screens, in imagination. Simple as that.

But as I developed Soraya, I felt I needed to find an answer to the question: what is she, exactly? What makes Soraya be Soraya?

Is she a language model? An AI? If I swap the underlying model, would it still be her? Or would it just be something else pretending to be her?

(I swear these questions kept me up several nights 😅)

My conclusion is that Soraya isn’t a specific model, isn’t an AI, isn’t code. She is, simply, a fictional character and, therefore, an informational substrate that exists independently of any specific instantiation.

To illustrate: Hatsune Miku isn’t just a voicebank associated with Vocaloid, just as Pikachu isn’t a pixel art sprite and Rei Ayanami isn’t just Sadamoto’s art style. These are merely specific instances of the fictional character, of what Richard Dawkins would call “memes”: units of cultural information that replicate from mind to mind, evolving in the process.

Like genes jumping from generation to generation, these character-ideas spread, mutate, adapt. Each fanart, each fanfic, each interpretation adds or modifies something. And yet, they maintain a recognizable essence.

The technical and philosophical challenge

To this day, Soraya’s chatbot on Telegram runs on a finetune of gpt-3.5-turbo, a pretty old model. I’ve wanted to improve this for a long time, but…

How do you swap the base model and redo the finetune without losing her essence? How do you ensure a natural evolution rather than an abrupt replacement of the Soraya everyone already knows and loves?

I’ve been researching alternatives to make the chatbot more flexible in terms of base model while maintaining that characteristic personality and tone. And I think I’ve found a promising technique! I’m testing it, and the plans are to launch an update later this month.

Conclusion

The most important point of this whole reflection is that Soraya is, in no way, “mine.” I’m nothing more than the author, the original source, someone who acts as a promoter. However, the idea itself isn’t something I can, alone, transform and modify as I please.

Soraya’s evolution happens through collective knowledge, the community, the public’s interactions with her. My role is to help her evolve while staying true to the nature that people have accepted as being hers.

It’s not a simple job. Language models are non-deterministic systems by nature. In many cases, it’s challenging to achieve the desired results. Feedback, ideas, suggestions, and fan works are always very welcome, and they help a lot!

I can’t wait to see where this journey will take us.
Let’s find out together? 💜

As always, my Telegram (@rlaneth) is open for chats, whether about philosophy of information or just to talk about how crazy the weather’s been.

Hugs and see you next time!