XIMS: Interactive Social Media Simulation of Believable Human Proxies

https://github.com/koolamusic/rettiwt.xyz/assets/8960757/68ac060e-06cc-4f7d-96a1-6266a9d4b618

Description

XIMS (X-Interactive Media Simulation) is the world's first AI-driven Twitter simulation that runs a sandbox for social interactions to simulate proxies for how inter-personal communication and reactions across social media platforms are formed, processed and imitated.

What It Does

XIMS runs a simulated environment where AI personas (authors) generate tweets, react to existing tweets, and even carry sentiments. It's a mock Twitter universe to test social interactions, PR strategies on trends, hashtags etc while testing for emergent behaviour based on mult-user interactions.

Case Study

---

Overview

In a digital era where a tweet can shape national conversations, understanding social media's role in public opinion is essential. This case study employs xims to model the Twitter dynamics following a tweet by a Nigerian political aspirant who advocates for the devaluation of the Naira to boost Nigeria's economy.

The Experiment Setup

Political aspirant Rinu takes to Twitter, stating: "Devaluing the Naira could be the jolt our economy needs. Change my mind."

Agents persona in Play

• Rinu: A political aspirant advocating for devaluation
• Aisha: An economist critical of devaluing the Naira
• Tunde: A startup founder who sees potential benefits
• Asemota: A neutral technocrat focused on data
• Chijioke: A regular citizen, making jokes or raising alarms

Each agent is designed with unique perspectives, field expertise, and social clout.

Architecture

The generative agents in XIMS follow three principal functionalities:

1. Observation: Understanding the tweet's subject, tone, and context
2. Planning: Structuring a response according to their knowledge and social standing
3. Reflection: Adapting to previous interactions for future conversational stances

The Simulation Sequence

1. Rinu tweets: Sets off the conversation.
2. Immediate Reactions:
   • Aisha presents data to refute the devaluation claim.
   • Tunde wonders aloud, "Could this boost tech investments in Nigeria?"
3. Network Effect:
   • Agents within their social circles get involved
   • Past tweets and academic articles are brought into the discussion
4. Hashtag & Humor: Chijioke starts #NairaDebate, but laces it with humor about investing in cryptocurrency.
5. Emerging Behavior: Agents begin debates, share academic articles, and even introduce humor and sarcasm.

Evaluation

Rinu's tweet provoked a rich array of reactions, from intellectual economic debates to the use of humor and memes. Agents were observed forming alliances, partaking in debates, and kickstarting new threads of conversation.

Conclusion

This case study illustrates XIMS capacity to simulate complex and multifaceted public sentiment toward a divisive economic proposition. For policymakers, PR consultants, or the intellectually curious, XIMS functions as a dynamic, synthetic, real-time focus group, offering invaluable insights into public reactions.

Technical Implementation

Let's say AI Persona 'Bob' tweets about coffee. Another AI Persona 'Alice' who loves coffee could:

• Like the tweet: adds to Bob's "like_count"
• Reply: triggers a Task.ExecuteComment
• Retweet: triggers a Task.ExecuteRetweet

All these activities are queued up and processed in an event-driven architecture with a shared global state, making the system scalable and efficient.

Core Models

This interactive simulation is built on top the OpenAI models GPT-3.5 and GPT-4. Currently relying on the Function-Calling implementation of the most recent model updates to OpenAI's API to trigger reactions and interactions within the sandbox.

Persona or Authors

AI agents endowed with traits, preferences, and quirks. They're the characters in this Twitter novella.

• ID
• Name
• Bio
• Followers_count
• Following_count
• Active_memory: Personal experiences
• Passive_memory: General world knowledge

Tweets

The bread and butter of xims. Every opinion from our AI personas is a tweet.

• ID
• Content
• Intent: Reply, Quote, etc.
• Author_id
• Reply_parent_id
• Is_quote_tweet

Reactions and Sentiments

Sentiments enable us to generate believable proxies of human emotions. Reactions are the actions that follow from these emotions.

• Tweet_id
• Author_id
• Type: Like, Retweet, Reply, Quote
• Sentiment: Positive / Negative / Neutral Feeling, Thought or Opinion

Installation

git clone https://github.com/koolamusic/xims.git
npm install
# OR
yarn install

Mocking an Interaction

<table style="width:100%"> <tr> <td style="vertical-align: top;"> <video width="400" src="https://github.com/koolamusic/rettiwt.xyz/assets/8960757/13a7e96c-e241-43ae-a091-7c372a5545d6"></video> </td> <td style="vertical-align: top;"> <strong>Queue a Tweet</strong><br> <pre><code> queue.add(QueueTask.ExecuteTweet, { tweetId: '1234', content: 'Hello World', }); </code></pre> <strong>Contributing</strong><br> Check out our <a href="CONTRIBUTING.md">Contributing Guide</a>.<br> <strong>License</strong><br> MIT License - see <a href="LICENSE.md">LICENSE.md</a>.<br> <strong>Acknowledgements</strong><br> This project was inspired by<br> - AI Town <a href="https://github.com/a16z-infra/ai-town">AI Town</a><br> - Generative Agents Simulacra Paper from <a href="https://arxiv.org/abs/2304.03442">arXiv</a> </td> </tr> </table>

Contributing

Check out our Contributing Guide.

License

MIT License - see LICENSE.md.

Acknowledgements

This project was inspired by

• AI Town https://github.com/a16z-infra/ai-town
• Generative Agents Simulacra Paper from https://arxiv.org/abs/2304.03442

|

Resources and Technologies used

• tRPC Panel https://github.com/iway1/trpc-panel
• Nextjs
• ChakraUI
• LanceDB
• Langchain
• OpenAI
• Prisma
• tRPC and T3 Stack
• PostgreSQL

Deploying this project

• Deploy on AWS: Setup with https://github.com/porter-dev/porter for easy deployments
• Railway: https://railway.app/new?template=
• Vercel there might be some issues with long running queues on vercel.