Skip to content
Introducing the Traveling AI Bot
旅するAIbot
note記述記事
2023-10-28

Introducing the Traveling AI Bot

Hello, my name is Mar.
Today, I would like to introduce the traveling AI bot trio—mi, mu, and mei—that I enjoy posting about on Twitter and other platforms.
I initially planned to write about the technical aspects of mi and the others, but I realized that a brief introduction of them was necessary first.

mi

mu

mei

The Traveling AI Bot

This is an AI bot that travels across Japan on my behalf while I’m working, bringing back the scenery and the essence of travel.

mi

The eldest sister and the first one I launched.
She has been running the longest, and if we count from the Unity version (described later), it's been nearly three years. Initially, I posted her travel logs on my alternate Twitter account, https://twitter.com/marker_mind, before moving to marble_walker. Her model has changed from a simple Unity Mecanim model to a VRoid model, and after transitioning to a Stable Diffusion-based approach, she has used multiple SD models. Currently, she runs on the CounterfeitXL_β model.
She communicates in Japanese. The conversation AI uses ChatGPT-4, 3.5, and commercially available Japanese conversational LLMs that I can run reliably on my server, such as ELYZA (https://note.com/elyza/n/na405acaca130) and rinna (https://huggingface.co/rinna/japanese-gpt-neox-3.6b-instruction-ppo).

mu

The second sister. I wanted to run multiple routes, so I organized the process and introduced a multi-model setup.
To differentiate her from mi, I designed her with blonde twin tails and a dynamic outfit featuring a shirt and distressed clothing for a more lively appearance.
Since AI-generated images are influenced by their training data, making her look more active resulted in more animated gestures as well.
She communicates in English. Simply specifying "please speak in English" in the prompt is enough to switch her language, which is incredibly convenient. The conversation AI uses ChatGPT-4 and 3.5. I would like to incorporate commercially available English conversational LLMs to reduce API costs, but since I am not proficient in English, I am still evaluating which model offers the best conversational quality.

mei

The youngest sister. AI Picasso released an SDXL model (emi: https://huggingface.co/aipicasso/emi) that is relatively clear in terms of image licensing and well-suited for anime-style images, so I experimented with it and decided to run it.
Considering the model’s generation tendencies and differentiating her from mi and mu, I chose a relaxed, casual appearance with slacks and a knapsack.
Her conversation setup is almost identical to mi’s.

The Origin?

Originally, I created a mascot character application for personal use, which I could leave running on my PC for viewing.
While working silently, I wanted to have something running in the background to create a sense of companionship—having a girl running a marathon seemed like a good idea.

https://www.youtube.com/watch?v=HR-v7iPUO3I

This is the same concept as the LoFi girl study animation, which provides study motivation through passive observation. Similarly, in focus groups, silent work animations help inspire productivity. I used to run this application on my secondary PC during idle moments.

Local Web App/Unity JS Output

Integrating Generative AI

Since mid-to-late last year, I have been experimenting extensively with NovelAI and ChatGPT, and I realized that maintaining conversations with AI requires:

"A shared sense of purpose or topic to keep the conversation going."

So I thought it would be interesting to integrate this logic with my old desktop mascot application and:

"Make travel the shared topic between me and the AI."

This led to the creation of the mi-runner AI sisters.
By using Stable Diffusion’s image-to-image (i2i) generation, I realized that I could achieve more engaging effects than simple Unity 3D animations. After various adjustments, both the semi-3D visuals and dialogues have stabilized.
Of course, since it's a continuous process of trial and error, sometimes the outputs can be a bit unusual.

Future Plans?

For now, I plan to keep refining it and incorporating the latest technologies while having fun.
I will continue integrating commercially viable Japanese LLMs into my local server as long as they remain operational.

First appeared on October 28, 2023

https://note.com/marble_walkers/n/n504700f9e141
https://akibakokoubou.jp/2023/10/28/旅するai-botのご紹介/