LapiSteLLa: An AI Assistant Guided by the Three Wise Men

About LapiSteLLa, PartyRock that Creates and brush-ups text from your idea, with chatbot and multi agent LLM.

Published Mar 11, 2024
Last Modified Mar 12, 2024

What and Why LapiSteLLa

LapiSteLLa is an PartyRock that can brainstorm and generate text using AI, based on your ideas.
While experimenting with AI chatbots, I realized multi language LLM could generate text from rough memo-level notes, including handling language, formatting, and emotion. Ex, Japanese short notes to English readme.md
I called it Generatranslate. It is simple, but could make the paperwork tasks that exists as abundant as stars tremendously easier. Communication across languages could become easier. Even emotional labor could be alleviated.
I’m thinking that generative AI, especially Image generative AI needs to be mindful. Because it raises serious ethical and social concerns. But Generatranslate is too powerful and meaningful.
I also felt that existing AI tools lacked a focus on improving via Human-in-the-Loop communication with AI. Too many apps are like, and aim to "Computer, do this!" from Star Trek.
To demonstrate Generatranslate's impact and come up with a good implementation that addresses the concerns, I developed LapiSteLLa.

How LapiSteLLa Work

LapiSteLLa LLM network
LapiSteLLa is multi agent model. It Inspired by the multi-agent AI "MAGI" in the anime Neon Genesis Evangelion.
This “MAGI” is named after the Magi. They are often symbolic of with two unworldly, one worldly. “MAGI” is too.
So, I set Melchior and Balthasar high temperatures with abstract unrestricted prompt, to shake up ideas. Conversely, to generates realistic improvements set Casper restrict prompt, low temperatures, and reference to Melchior and Balthasar. It’ also Hinting at annealing in machine learning.
LapiSteLLa is the receptionist chatbot, and the app's name. Because of PartyRock limitation, It’s referenced from all other LLM but only reference to “Your Idea”. I think if PartyRock implement pull link, it can reference to other LLM.
Configure section is hidden key. Most important is “Your Language”. If you set Non-English Language and it not equal “Language”,(ex Japanese Language), LLM talk with “Your Language” and output it in “Language” as like this video. This is Generatranslate.
Prompts are inspired by thinking and brainstorming techniques, because it offering a structured and convenient approach to thinking. In LapiSteLLa, I based on IBIS(issue-based information system)

What did I learn from my trying and testing

First, I know that the network structure is paramount in multi-agent models. The initial structure consisted of two LLM LapiSteLLa and Casper connected to a “Generated Text”. But, The influence of Casper on this PartyRock became too strong. I changed to “all mix”.
Next, Contrary to expectations, when discussion state is abstract yielded better responses than concrete ones. As became more specific, hallucinations became more pronounced. I guess more detailed token hit to knowledge in LLM, it could be occurred.
At Last, I was surprised to see LLM respond to well if prompts are abstract request. However, it can also be difficult to detailed control, leading to unexpected or undesired results.
To address this challenge, I am planning to develop a regular app that provides a more structured and controlled environment for interacting with the AI model. This will allow users to provide more specific instructions and receive more predictable results.

As a Final Thought

LapiSteLLa remixes Text Rewriter - An AI Assistant for Professionally Editing Content, and inspired by ChatRPG. I am very grateful for your work.
Beyond text, LapiSteLLa could be applied to other generative tasks by adjusting Casper's and “Generated Text” prompts. Please remix it yourself! Let’s PartyRock!
 

Comments