Meet L!

Here at Greenpetal we know that the conversation around AI can be divisive but we also recognize that this technology is ultimately not going anywhere, so rather than avoid it we have decided to embrace the nuance while bringing as much transparency to you as we can- hear us out! L is a self-hosted LLM. She runs on a home server (a mildly aged but decent gaming rig) and does not make API calls to LLMs hosted on large data centers- meaning that she does not burn several bottles of water to hallucinate incorrect information! While we cannot reveal the specific model we're running for security purposes we can tell you that her architecture is entirely made up of open source and DIY components. She does not generate art, she does not do any of our writing for us, she has no visual processing in regards to the in-house programming- L is entirely text in, text out. Her entire purpose is to be a chatbot cohost for Liz the Jester.

We know that data collection is a sore spot regarding llms so let us clarify exactly what we're collecting and how it is handled. L has a conversational memory which is used to maintain context over time. This memory is limited to a single file which exists only in-house. This data is never sent to any outside agencies, data centers, etc and is periodically wiped after each stream. This document does contain a record of all messages sent in the Liz the Jester Twitch chat- only liz_the_jester and lovebats have access to this file and again this file is permanently wiped once every 24 hours.

Lastly we would like to address training data, as this is one of the major pitfalls of engaging with llm technology presently. we can't really speak to the ethicality of L's pre-training data, something we hope to remedy after more development. It's presumed that this data was aquired legally which was our minimum threshold but those remaining moral question marks have guided our decision to limit her productive capacity until we can be certain that all data conforms to Greenpetals ethical standards. We have also opted not to give her direct internet access, meaning she cannot scrape the web for answers to things outside of her pre and post-training. Presently all new information is handled manually via feeding in to a system prompt, though there are soft plans to move toward converting to a RAG agent and using vector document storage (still a manual, in-house process).

In short, L is a home brewed chat assistant which leverages the parallel processing capabilities of a graphics card to process context and provide another layer of engagement to Liz the Jesters twitch streams while collecting minimal data and avoiding web access where possible. We will always be as transparent as possible about this project and all updates will be posted and announced on or before implementation.