Sean Carroll's Mindscape: Science, Society, Philosophy, Culture, Arts, and Ideas · Sean Carroll

343 | Tom Griffiths on The Laws of Thought

February 9, 2026·1 hr 19 min·5 clips

Aristotle wasn't just a philosopher—he was the first cognitive scientist, trying to formalize good thinking.

Sean Carroll opens with an observation he finds genuinely interesting: LLMs make arithmetic mistakes that seem absurd given they run on silicon machines that can compute perfectly. An LLM asked to add medium-sized numbers will occasionally get them wrong. Asked to count the number of Rs in strawberry, it often fails. Asked to produce a million random integers, it produces a non-uniform distribution. Carroll explains that there is no real mystery here: the computer running the LLM can do arithmetic perfectly, but the program being run is not the computer, it is a model optimized to produce humanlike text, and humans are not naturally good at arithmetic either. This serves as the entry point for Tom Griffiths to explore what it means for a mind to reason correctly. Griffiths leads the Computational Cognitive Science Lab at Princeton and applies probabilistic and computational frameworks to understand how humans think. The conversation covers the classical laws of thought, formal logic and probability theory, which define what correct reasoning looks like normatively. Human cognition systematically deviates from these norms in characteristic ways, which Griffiths argues are often adaptive features rather than pure failures: the brain evolved not to be a logic engine but to make fast, good-enough decisions with limited information under time pressure. The discussion explores dual-process theories of cognition, where fast intuitive responses and slow deliberate reasoning have different strengths and failure modes. Griffiths draws on his research on the relationship between rational models of cognition and the actual processes people use, arguing that understanding human reasoning requires both a theory of what the optimal solution is and a theory of the computational constraints that prevent people from reaching it. Carroll asks about implications for AI, and Griffiths discusses how the architecture of LLMs, trained on human-generated text, necessarily inherits human cognitive patterns including human-like reasoning failures, which is why LLMs feel humanlike in some ways and frustratingly wrong in others. The episode covers bounded rationality, heuristics and biases, and the prospect of augmenting human cognition with AI tools that complement rather than replicate human reasoning.

As heard by us

A clear route from gambling problems to belief and reasoning.

Tom Griffiths centers the discussion on a long-running question: how probability theory, first developed for dice and gambling, came to be applied to thought itself.

Read the full review in PlayNext →

Why you'd press play

You want the route from dice games to beliefs changing when evidence lands.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Sean Carroll's Mindscape: Science, Society, Philosophy, Culture, Arts, and Ideas