The Interconcept Space (enricobottazzi.github.io)

A new interactive project titled "Interconcept Space" invites users to explore the nuanced space between two concepts as interpreted by a large language model (LLM). Inspired by Stephen Wolfram’s explorations, this tool allows users to select two activation vectors sampled from layer 20 of the Qwen2.5-7B model, adjusting a slider to visualize and understand the intermediate combinations of these vectors. The mathematical underpinning uses a linear combination formula, z = (1 − α)A + αB, where the mixing coefficient α varies from 0 to 1, enabling a detailed traversal through conceptual gradients. This project is significant for the AI/ML community as it offers insights into LLM functionalities and deeper model interpretability. By utilizing a Natural Language Autoencoder (NLA) to describe each vector along the slider, users gain a clearer understanding of how LLMs perceive relationships between concepts. This exploration not only enhances our comprehension of LLM behavior but also opens avenues for refining AI systems in terms of semantic understanding and representation learning, potentially benefiting diverse applications from natural language processing to knowledge extraction.