Project LION is a collaborative effort between talentDAO and labDAO that aims to build an open source AI-powered analytics system for assessing community health.
The initial goal of Project LION is to build a minimum-viable AI-powered community health toolkit. However, we also believe the underlying technology will have useful applications for social science.
Namely, the psychometric alignment of digital twins and their deployment in silico, to effectively run as organizational simulators. We’re excited about digital twins, but with the initial goal of this project in mind, we’ll be releasing a series of notebooks for analyzing DAO communication.
Building on labDAO infrastructure, our Jupyter notebook system will be able to run state-of-the-art transformer models in the browser. Extending the power of organizational psychology and data science to anyone building in the decentralized digital economy.
We believe the ethos of crypto extends to all new technology, including AI. Like crypto, AI has a place within the open source movement because it is fundamentally a technology that can help forward human progress, particularly when made open and accessible to all.
Here we outline our plan to build LION as a public good.
Our core objective is to train models that help evaluate the language patterns of an organization’s social network.  The first step is gathering data on the network we're evaluating. To do this, we'll build a data bridge from Discord to our model's training environment. This pipeline will send data through a secure database , followed by a pre-training data preparation process.
A connection to this database will be set up for the end user in the LION notebook system.
In the second phase of the project, we'll begin working on the parts of the system that improves upon classical NLP by leveraging transformers.
We’ll be adding a variety of heads to our analytical pipeline. Some initial examples will include classification and sentiment analysis, but we’ll also use empathic analysis and word embeddings to produce our measure of community voice. As we work through this approach, we’ll leverage talentDAO organizational scientists to help validate model outputs.
In phase three, we'll add a network graph feature that allows the user to visualize the informal structure of their organization or community. We'll then show how we can map psychometric properties onto the network using outputs from classification and other metrics derived from user activity and sentiment.
In addition to visualizing, we'll leverage network data science to gain deeper insight into the health of the network with metrics like connectivity and network centrality.
In the future, those using the talentDAO Health Survey will be able to map their survey responses onto the network.
In the fourth phase of this project, we'll build the first-ever LION digital twin. We'll demonstrate some simple experiments and begin to work through the psychometrics of these models to understand how they relate to their underlying corpora.
We'll demonstrate how fine-tuning can be applied to digital twins to create custom bots for different use cases.
After LION’s initial release, we hope to build a thriving open source community around the system to help improve and maintain it.
Our dream is to one day produce a front-end web application for exploring your organization’s social network by interfacing with an AI assistant.
We intend to document and experiment throughout this process. Ultimately, we’ll write a whitepaper with all the details of how we built it, and what we learned. We hope this process opens the door to a new way of working with social networks – a branch of social science that deploys AI agents to simulate organizations.
Social networks can be communities, organizations, etc. we refer to these interchangeably here and often use the term organization, despite some obvious differences in definitions because the technology has applications regardless of the defining social structure of the group.
Our database will dynamically update to prevent data from being stored longer than you need it to be. Over the long haul, we would like to add encryption.