The problem Metacast solves

Problem :

Every social media user has the problem of what to see and how to be seen. As a growing Web3 social media platform, Farcaster has many organic quality casts, but how to find and engage in those conversations, esp. if you don't want to scroll the home feed or trending feed all day to find out?

Metacast tries to find this problem in two folds:

1. Format: A farcaster channel integrated to the platform for fricionless user experience. A channel bot posts automatically, every week day, 5 top engaging casts from farcaster users, encouraging channel followers to engage with those casts. If new users engage more with those casts, they have a better chance to be seen; they also don't have to worry about where to find quality discussions on farcaster;

2. Method: curation is important to make sure the casts can serve its purpose. We designed 3 fold system for content curation: a) Top/Trending casts fetching from airstack api; b) a LLM powered curation process for filtering top casts; c) human curation from the admin console to make final curation decisions;

In season 1, we aim to make the curation process more permissionless by integrating an onchain subscription model, any one who's interested in the human curation process can purchase an subscription on base and join the curation process. After season 1, we are going to integrate the human curation rules in the LLM module.

We built our solution based on following assumptions:

1. most social media users nowadays don't have information scarcity problem, but an information overload prolem: curation is essential to solve this problem

2. AI helps curating large amount of contents, but AI has no taste and values, only human beings have taste and also needs to make a value judgment on how to make sure the algo is good for the collective public. Thus we need to explore more on how to build hybrid intelligence: human-AI collboration.

Challenges we ran into

We encountered 2 major problems:

1 Airstack API unstability, there were several bugs influencing whether we can successfully fetch casts from farcaster, as well as whether our channel bot can cast automatically. We solved this problem after bring these issues to airstack team and it was solved. However, in future, we will iterate on our product so it relies less on these APIs.

2 Working with Large language models: When working with LLM models, especially in natural language processing tasks, it's crucial to remember that these models operate on probabilistic principles rather than strict logical rules. This approach often yields more accurate and flexible results in language-related tasks. Our solution is to use probability-based judgment, because AI's underlying training logic is based on probabilities, not strict logic.

However, to solve above 2, this is just the first step. Following roadmap is designed to fully utilize AI for social media content recommendation:

1. Vector Database and Embeddings

• Implement text embeddings and vector storage
• Select and integrate a suitable vector database (e.g., Faiss, Milvus)
• Optimize vector indexing and search algorithms

2. Multimodal Content Processing

• Integrate image recognition models (e.g., ResNet, CLIP)
• Implement joint image-text embeddings
• Extend to other media types (video, audio)

3. Context-aware and Personalized Recommendations

• Design user profile system
• Implement conversation history and long-term interest modeling
• Develop dynamic interest decay mechanisms

4. Local Model Deployment

• Evaluate and select suitable small LLMs for local deployment (e.g., BERT, GPT-2)
• Implement model quantization and optimization for faster inference
• Design incremental learning mechanisms for continuous local model improvement

5. Multi-Agent System

• Design specialized agent roles (e.g., content filtering, personalization, explainability)
• ImpImplement inter-agent communication and collaboration mechanisms