969: The Laws of Thought: The Math of Minds and Machines, with Prof. Tom Griffiths

Episode Summary

Executive Summary: Tom Griffiths argues that “laws of thought” are mathematical principles for understanding internal cognition, much like physics describes the external world. The conversation links psychology, computer science, and AI: human minds are powerful but resource-limited and use priors, probability, and heuristics; modern AI is powerful but different, often driven by training-data statistics and scaling. The episode explores curiosity, inductive bias, benchmarks, and how to design AI that is both capable and interpretable.

Main Topics: Laws of Thought and the mathematical theory of mind: Griffiths frames cognition as something that can be modeled with logic, probability, and computation, extending the idea of laws from physics to thought. Psychology vs. computer science views of intelligence: He contrasts psychologists’ emphasis on human irrationality with computer scientists’ admiration for what humans accomplish under severe constraints. Bayesian reasoning, priors, and inductive inference: The discussion explains how probability theory and Bayes’ rule support human-like inference from incomplete data, using the orangutan detective example. AI limits, training data, and autoregression: Griffiths describes how LLM behavior is shaped by probabilistic training objectives, making them sensitive to token frequencies and training distributions. Inductive bias, scaling, and human-vs-machine learning: The episode examines why human children need far less data than LLMs, and how meta-learning may help AI incorporate useful priors. Curiosity as rational information seeking: Curiosity is modeled as a strategy for gathering information that improves future action, producing an inverted U-shaped interest curve. Evaluation, benchmarks, and alignment: He argues for scientific experiments over broad benchmarks when studying AI, to diagnose internal mechanisms rather than just score performance.

Key Arguments: Human cognition is best understood as rational adaptation under tight constraints: limited time, memory, compute, and communication bandwidth. Psychologists and computer scientists emphasize different truths about humans: we are error-prone, yet remarkably capable for our constraints. Probability theory is more useful than pure logic for modeling inductive problems where uncertainty is unavoidable. LLMs are fundamentally statistical systems; even when a task has a single correct answer, outputs can be biased by token frequency and training distributions. Human language learning relies on strong priors/inductive biases that current neural networks lack, which is why children learn from vastly less data. The history of AI tends to move from one dominant approach to another until each hits a limit, after which a new paradigm emerges. Meta-learning is a promising way to encode soft priors by learning good initialization rather than hard-coded rules. Curiosity can be formalized as value-of-information maximization, which explains why moderately familiar things are often most interesting. Benchmarks are useful for engineering, but scientific experiments are better for understanding what AI systems are actually doing internally. AI should not be assumed to mirror human cognition; biological inspiration is useful at the level of principles, but not always at the level of mechanisms.

Data Points: Podcast episode: 969 - Super Data Science Podcast episode number Institution: Princeton University - Tom Griffiths’ affiliation Book title: The Laws of Thought - Griffiths’ new book discussed in the interview Prior book title: Algorithms to Live By - Griffiths’ earlier bestseller Sample size: more than 10,000 - Number of decision-problem responses collected online in one lab study Language-learning comparison: 5 years of speech vs. 5,000–50,000 years of language input - Contrast between human children and LLMs Publication milestone: 2012 - Year AlexNet won ImageNet and helped launch modern deep learning scaling Counting example: 30 more likely than 29 - LLMs are more likely to answer 30 because that token is more common in training data Curiosity curve: inverted U-shaped - Moderately familiar stimuli are modeled as most curiosity-provoking Relative timing: few decades - Approximate human lifespan used to emphasize cognitive constraints

Pivotal Quotes: "if you go and talk to a psychologist, you're going to be told people are terrible decision makers. We're irrational. ... If you go and talk to a computer scientist, they'll tell you: humans are impressive." — Tom Griffiths: Explaining the tension between psychological and computer-science perspectives on human intelligence "probability theory is the tool that we can use for solving these inductive problems, because it gives us a way of talking about how data should influence our beliefs, even when we don't have enough information to be certain." — Tom Griffiths: Describing why Bayes and probability are central to thought and cognition "if you ask a modern model to do something like this ... they're influenced by the expectations that they have about the distribution of language." — Tom Griffiths: Explaining why LLMs can produce counterintuitive errors on deterministic tasks

Implications: Listeners should expect AI progress to keep diverging from human cognition unless designers deliberately add human-like priors and structure. For industry, the big opportunity is not just scaling, but better inductive bias, evaluation, and alignment-aware design.