CTO Pulse Podcasts

What's Missing Between LLMs and AGI - Vishal Misra & Martin Casado

In this episode, Martin Casado interviews Vishal Misra, a Columbia professor who has built a mathematical model of how large language models (LLMs) operate, treating them as massive sparse matrices that update token probabilities via Bayesian-like inference. Misra explains his early work using GPT‑3 for in‑context learning to translate natural language queries into a custom domain‑specific language for cricket statistics, and how this led to a series of papers showing that LLMs perform predictable, matrix‑based updates rather than true reasoning. He argues that while LLMs excel at pattern matching, achieving AGI will require continual post‑training learning and a shift from correlation to causal modeling. The conversation also touches on the debate over whether LLM behavior is genuinely Bayesian and the practical tools Misra created to probe token probabilities.