Why Large Language Models think differently to us

This episode explores the world of embeddings, mathematical representations that allow Large Language Models (LLMs) like ChatGPT to “think” in thousands of dimensions. While humans are limited to conceptualizing in three dimensions, LLMs operate in 2048 or more, using embeddings to encode meaning and capture semantic relationships between words. 

The discussion contrasts this form of statistical pattern recognition with the richer, experience-driven reasoning of the human brain. It also introduces a new technique called ‘vec2vec,’ which enables translation between embeddings from different models. While powerful, this raises potential security concerns about reverse-engineering sensitive data from vector databases. The episode sheds light on the impressive capabilities of LLMs, while also questioning what it means for a machine to “understand.”