What this covers

Tom Griffiths, a psychologist and computer scientist at Princeton, discusses his book The Laws of Thought with Sean Carroll, tracing how mathematicians and cognitive scientists have attempted to formalize the principles governing human reasoning. The conversation moves across three centuries, from Gottfried Leibniz and George Boole's pioneering work in formalizing logic, through the mid-twentieth-century rise of cognitive science, to contemporary debates about artificial intelligence and human learning. Griffiths argues that three complementary mathematical frameworks—formal logic, [probability theory](jargon:probability theory), and spatial representation—best capture how thought actually works, organized through David Marr's framework of computational, algorithmic, and implementation levels.

The episode traces a historical arc showing how each mathematical advance unlocked new ways to think about thinking. Griffiths explains how Bayesian inference extends logic into domains of uncertainty, how Eleanor Rosch's work on fuzzy categories motivated representing concepts as points in space, and how the mathematics Leibniz developed centuries ago underlies modern backpropagation and deep learning. The discussion then pivots to human irrationality—apparent failures of logic and probability—which Griffiths reframes through resource rationality: the idea that observed heuristics and biases are rational strategies for agents with finite time and cognitive capacity. He also addresses a tension between human and machine learning, arguing that the orders-of-magnitude difference in data requirements between children and large language models reveals the role of inductive bias that humans inherit from evolution and embodied experience.

Sharpest takeaway

Griffiths argues that thought is best understood through three complementary mathematical threads—logic, probability theory, and spatial/neural-network representations—organized across Marr's three levels of analysis, such that the 'laws of thought' are most secure at the abstract computational level (logic and Bayesian inference) while human apparent irrationality often reflects rational use of limited cognitive resources.

  • Marr's three levels (computational, algorithmic, implementation) mean there is no single theory of mind; multiple compatible explanations can all be correct.
  • Bayesian inference and logic define ideal solutions at the computational level for inductive and deductive problems respectively.
  • Human heuristics and biases can be reframed as resource-rational strategies—the best one can do with finite cognitive resources.

The argument · threads3 threads · 45 claims
0.85

Intelligence is task-specific; optimization for one capacity trades off against others due to constraints.

3 pointscentrality 5/5
  • Large language models in their natural state are bad at arithmetic and number-based tasks (adding medium numbers, counting letters, generating uniform random numbers) not because the underlying computer can't compute, but because you're talking to a program optimized to sound human rather than to do arithmetic—exactly like humans, who are also intrinsically bad at arithmetic without tools.

    “It's almost like you tried really hard to make a program that sounded human, and in the course of doing that, it lost the ability to do arithmetic”

  • A human brain capable of arbitrarily good arithmetic might be worse at other things more important for survival, because the brain operates under constraints of energy, fuel, and time, so optimizing one capacity trades off against others.

    “If you had a brain, a human kind of brain that was able to do arbitrarily good arithmetic, that might make it worse at other things that were more important for survival.”

  • When we say 'thought' or 'thinking' we are not referring to a single ability; adding numbers and carrying on a conversation are very different capacities, and a system (organism or program) may be optimized for one at the expense of the other.

    “when you say thought or thinking, you're not really referring to a single thing. The ability to add numbers together and the ability to carry on a conversation, those are two very different abilities”

0.61

Apparent irrationality in humans reflects rational design constraints, not simple error or blank-slate learning.

2 pointscentrality 4/5
  • Humans are not blank slates; as thinkers from Kant to Chomsky argued, we are born with some predispositions, and 21st-century cognitive science is better at teasing out which ideas we are born with (innate inductive biases) versus which we pick up from experience in the world.

    “we're not blank slates, right? I mean, I guess various thinkers from Kant to Noam Chomsky have said that like, yeah, we're born with some ideas in our heads”

  • Certain ways human beings act irrationally—when compared against the perfect laws of logic—actually have good reasons; there are reasons we have particular inclinations and biases, so the apparent irrationality is not simply error.

    “certain ways in which human beings act irrationally, if you compare them to the perfect laws of logic that you might aspire to, actually have good reasons for them”

0.38

LLMs can execute formulaic tasks but struggle with creative conceptual leaps beyond their training data.

1 pointcentrality 3/5
  • An LLM could likely solve Einstein's general relativity equations in the right context but would struggle to have the first creative moment of suggesting that gravity is the curvature of spacetime, because such models are trained on things that have already happened; whether they can make such extrapolative inferences is an open question.

    “it would really struggle to have that first creative moment where it it suggested that gravity is the curvature of spacetime, because simply because they are trained on things that have already happened”