Journey to the Centre of the LLM

Tokyo Tech Talks #8 looks beneath the familiar input-output surface of large language models. Independent researcher Larry Richards will guide attendees through experiments with model internals, examining what happens inside an LLM and what those internal behaviors reveal about how it works. Larry researches mechanistic interpretability, sparse representations, activation interventions, and the geometry of language-model internals. He also created and maintains vollib, an open-source Python and TypeScript library for option pricing, implied volatility, and Greeks. The session assumes a basic understanding of Transformers (layers, activations, next-token prediction) and focuses on experimental exploration rather than an introduction to the architecture. 3Blue1Brown's visual introduction to Transformers is suggested as preparation. The event takes place in Shibuya on September 29, 19:00–21:00.