343 | Tom Griffiths on The Laws of Thought

Episode Summary

Executive Summary: Sean Carroll and cognitive scientist Tom Griffiths explore the idea of “laws of thought”: mathematical principles governing ideal reasoning. The conversation traces logic from Aristotle to Leibniz and Boole, then shifts to probability/Bayes for uncertain inference, and finally to neural networks and resource rationality as models of bounded human cognition and AI.

Main Topics: Laws of thought as a scientific project (Priority: 5/5): Griffiths argues that cognition can be studied at an abstract computational level, where logic and probability theory describe how ideal reasoners should think under constraints. History of logic from Aristotle to Boole (Priority: 5/5): The discussion follows the development of formal logic from syllogisms and universal language schemes to Boole’s algebra of logic, which helped mathematize reasoning. Probability theory and Bayesian inference (Priority: 5/5): Bayes and Laplace extend mathematical reasoning to uncertainty, treating belief updates as probabilistic inference rather than all-or-nothing truth. Resource rationality and bounded cognition (Priority: 4/5): Human irrationality is reframed as adaptation to finite time, memory, and energy; heuristics and sub-goals can be rational under resource constraints. Neural networks, spaces, and modern AI (Priority: 5/5): A third thread links thought to geometric representations and neural networks, showing how modern AI implements cognition differently from logic-based systems. Inductive bias and language learning (Priority: 4/5): Griffiths contrasts human language acquisition with LLM training, arguing that humans rely on strong prior structure/inductive bias to learn efficiently from limited data. Multiple levels of explanation (Priority: 5/5): Using Marr’s levels, the episode emphasizes that thought can be explained at computational, algorithmic, and implementation levels without reducing everything to one theory.

Key Arguments: Logic captures deductive reasoning, but uncertainty requires probability theory; ideal reasoning therefore includes both logical and Bayesian principles. Human cognition is not simply irrational; many biases and heuristics may be efficient solutions to problems under strict resource constraints. Large language models can appear intelligent while still exhibiting odd failures because they optimize different objectives and have different inductive biases from humans. Neural networks provide an algorithmic way to approximate complex cognitive computations, especially when scaled and trained appropriately. Language learning is a key example of the human need for strong inductive bias: children learn from far less data than current LLMs require. A complete theory of mind requires multiple compatible explanations at different levels, not a single all-purpose law.

Data Points: LLM language training data equivalence: 5,000 to 50,000 years of continuous speech - Griffiths contrasts the data needs of large language models with human children learning language in about five years. Human language learning exposure: About 5 years - Used to highlight how little data children need compared with LLMs. Book release timing: Within a week of each other - Sean Carroll notes Griffiths has two related books coming out almost simultaneously. Boole’s book structure: Two halves - His later treatise on the laws of thought is described as half logic and half probability theory. Neural network learning limit in early criticism: One layer of adjustable weights - Minsky and Papert’s critique focused on simple perceptron-like networks.

Pivotal Quotes: "The laws of thought are a tool for then generating what are the sort of appropriate ways of using your cognitive resources at the object level, the actions you take in the world." — Tom Griffiths: Explaining how abstract reasoning principles connect to practical behavior under resource constraints. "We are good at solving the kinds of problems that we face with the resources that we have." — Tom Griffiths: Summarizing the resource-rational view of human cognition. "There’s not necessarily a one-to-one mapping between those levels." — Tom Griffiths: Clarifying why multiple valid explanations can coexist across computational, algorithmic, and implementation levels.

Implications: Listeners should think of reasoning as layered: ideal logic/probability, resource-bounded heuristics, and physical implementation all matter. For AI, better inductive biases and human-aligned representations may improve efficiency, interpretability, and generalization.