Summary
Zoubin Ghahramani, professor at Cambridge and co-lead of frontier AI at Google DeepMind, joins the Google DeepMind podcast to make the case that uncertainty quantification — not just scale — is a critical missing ingredient in today’s AI systems. Drawing on thirty years of Bayesian machine learning research, he argues that true intelligence requires representing, updating, and acting on uncertainty, and that current large language models conspicuously lack this capability.
The conversation distinguishes two types of uncertainty: aleatoric (irreducible randomness, like a coin flip or an unpredictable pedestrian) and epistemic (knowledge gaps that can be reduced by gathering more information). Ghahramani illustrates the practical stakes with self-driving cars encountering rare long-tail scenarios — a horse in a hailstorm — where a system aware of its own uncertainty would slow down, while one that is not would proceed with false confidence. He explains how Bayes’ rule provides a formal mathematical framework for updating beliefs as evidence accumulates, and contrasts this with LLMs, which mimic uncertainty language through next-token prediction but capitulate when pushed back on, revealing they are not maintaining coherent probabilistic beliefs.
The episode situates Ghahramani on one side of a live debate within AI research: against those who believe pure scale will solve remaining capability gaps, and alongside those who see architectural innovation — particularly around probabilistic reasoning — as necessary for reliable intelligence. For engineers and researchers building applications where confidently wrong answers carry real-world costs, the discussion offers a grounded technical critique of current LLM foundations.
📺 Source: Google DeepMind · Published August 26, 2026
🏷️ Format: Podcast
