343 | Tom Griffiths on The Laws of Thought

What It Covers

Tom Griffiths explores how mathematical frameworks like logic, probability theory, and neural networks provide laws governing rational thought. He examines the gap between ideal Bayesian reasoning and actual human cognition, explains resource rationality as a framework for understanding cognitive shortcuts, and discusses how large language models differ from human learning through their massive data requirements versus human inductive biases.

Key Questions Answered

• •Computational Level Analysis: David Marr's framework separates intelligence into three levels: computational (what problem is being solved and the ideal solution), algorithmic (actual cognitive processes that approximate solutions), and implementation (physical realization in brains or silicon). This multi-level approach explains why there cannot be a single unified theory of mind, as correct theories at each level must be compatible but need not map one-to-one between levels.
• •Resource Rationality Framework: Human cognitive biases often represent optimal solutions given finite computational resources rather than irrational errors. Setting subgoals, using sampling strategies instead of considering all possibilities, and employing heuristics become rational when viewed through constraints on time, energy, and processing capacity. This reframes apparent irrationality as efficient resource allocation rather than flawed thinking.
• •Language Learning Data Gap: Children learn language from approximately five years of exposure, while large language models require training data equivalent to 5,000 to 50,000 years of continuous speech. This 4,995-year gap represents the inductive bias and prior knowledge that evolution and early childhood experience provide humans. Understanding this gap reveals fundamental differences between human and artificial intelligence learning mechanisms.
• •Meta-Learning for Inductive Bias: Model-agnostic meta-learning trains neural networks by optimizing initial weights across multiple learning tasks, creating systems that learn faster from limited data. The algorithm uses nested loops: an inner loop learns individual tasks while an outer loop adjusts starting weights to improve performance across all tasks. This approach helps identify what biases human learners need to acquire language efficiently.
• •Probability Distribution Sensitivity in LLMs: Large language models exhibit unexpected biases based on output probability rather than correctness. When counting letters in strings, models perform better on answers like 30 versus 29 because 30 appears more frequently in training data. This reveals how training objectives shape behavior differently from human cognitive goals, creating jagged intelligence boundaries where systems excel in some areas but fail spectacularly in adjacent tasks.
• •Neural Networks as Spatial Computation: Neural networks transform points from one vector space to another, providing a mathematical framework for computation with concepts represented as spatial positions rather than logical symbols. This approach resolves Eleanor Rosch's finding that human categories lack clear logical definitions, instead showing gradient membership (armchairs are clearly furniture, rugs are marginal). Multiple layers enable multi-step transformations through increasingly complex spaces.