Agentic Pattern Mining: Introducing Knowledge Diffusion at Meta AI Hackathon Reception

Meta hosted a kickoff reception for the Llama Hackathon weekend in Austin, and I was invited by their Innovation Policy team to present a demo of my novel work on pattern mining in large language models. The technique is called knowledge diffusion.

This technique is explores the emergent capabilities of LLMs to generalize new information and knowledge structures from their training data, using a novel agentic pattern. We design and deploy a multi-agent system tasked with creating a digital garden on user-provided seed topics. The collective attention of the agents spirals generatively through the conceptual space of the seed topic, ultimately building an organic digital garden of connected files that reflects the system’s exploration of the topic space.

Every iteration is unique. Bulk knowledge diffusion runs can be mined for novel and innovative material that can be applied elsewhere.

Read: More about this technique on my personal blog ↦ “Knowledge Diffusion: A New Paradigm in LLM Information Extraction”.

Read: More about the Hackathon ↦ Austin-American Statesman.