335 | Andrew Jaffe on Models, Probability, and the Universe

Episode Summary

Executive Summary: Sean Carroll and Andrew Jaffe argue that science works by building and comparing probabilistic models rather than discovering certain foundations. Using cosmology, quantum mechanics, thermodynamics, and AI, they show how Bayesian reasoning unifies inference from messy data, why models are unavoidable, and why debates over interpretation remain open even when calculations agree.

Main Topics: Models as the basis of knowledge (Priority: 5/5): They argue that all understanding—scientific and everyday—requires models, which are simplified stories or mappings of the world used to navigate and predict experience. Probability and Bayesian inference (Priority: 5/5): The conversation centers on probability as the proper tool for updating beliefs about models in light of data, especially when comparing competing explanations. Historical examples in physics and cosmology (Priority: 5/5): Newton vs. Einstein, Mercury’s orbit, gravitational lensing, CMB analysis, and the Hubble constant illustrate how models gain or lose credibility through evidence. Frequentist vs. Bayesian statistics (Priority: 4/5): They contrast frequentist error bars and repeatable procedures with Bayesian posterior beliefs about what a measured quantity likely is, noting both can be useful but answer different questions. Quantum mechanics and interpretation (Priority: 4/5): They discuss the probabilistic nature of quantum mechanics and interpretations such as many worlds and QBism, emphasizing that they yield the same practical predictions. Entropy and statistical mechanics (Priority: 3/5): Statistical mechanics is used to show how macroscopic laws emerge from ignorance about microscopic states and why entropy is tied to probabilistic descriptions. AI, LLMs, and the Chinese room (Priority: 3/5): The discussion briefly connects model-based cognition to large language models, asking whether they possess a world model or merely an effective lookup structure.

Key Arguments: Scientific knowledge is inherently provisional: no finite set of observations yields absolute certainty, only probabilistic confidence in a model. A model is any structured way of representing relationships in the world; even ordinary perception and conversation rely on implicit models. Bayesian inference is the right framework for updating beliefs because it explicitly conditions probabilities on prior information and new data. Frequentist error bars are mathematically valid but answer a different question than the one scientists usually care about: what is the probability that the parameter takes a given value? Newtonian gravity was an excellent model, but Einstein’s model became more probable because it explained additional anomalies such as Mercury’s perihelion and light bending. In cosmology, especially CMB analysis, Bayesian methods are especially powerful because the data are indirect and the models contain many nuisance parameters that must be marginalized over. The Hubble tension shows how model dependence and calibration differences can produce competing high-confidence estimates from different methods. Quantum mechanics is fundamentally probabilistic in our experience, but its interpretation is still contested; Bayesian-friendly interpretations may avoid adding unnecessary ontology. Thermodynamics and entropy can be understood as laws about what can be known and extracted from systems, not just about hidden microscopic reality. Large language models may count as models of the world in an operational sense, even if it is unclear whether their internal structure resembles human understanding.

Data Points: Mercury perihelion discrepancy: about 1/60th of 1/1300 of a circle - Used as an early anomaly that Newtonian gravity could not explain well, motivating Einstein’s theory. Gravitational lensing test difference: factor of 2 - Newtonian and Einsteinian predictions for light bending differed by a factor of two, helping confirm general relativity. Supernova neutrino event year: 1987 - A nearby supernova triggered one of the first major Bayesian analyses in particle astrophysics/cosmology. CMB age after Big Bang: about 400,000 years - Describes when the cosmic microwave background light was released. Temperature of Hubble constant example: 67 km/s/Mpc ± 3 km/s/Mpc - Example of a Bayesian-style uncertainty statement for the universe’s expansion rate. CMB-based Hubble constant: about 67 km/s/Mpc ± 1 km/s/Mpc - Jaffe’s side of the modern Hubble tension from cosmic microwave background analysis. Distance-ladder Hubble constant: about 72 km/s/Mpc ± 1 km/s/Mpc - Local-universe measurements that disagree with the CMB-inferred value. Early competing Hubble values: 50 vs. 100 km/s/Mpc - Historical disagreement before the field converged toward intermediate values around 75. Cosmic convergence value: about 75 km/s/Mpc - The approximate midpoint that earlier measurements were drifting toward around the year 2000. Gas particle count scale: 10^23 particles - Illustrates why statistical mechanics can reduce microscopic complexity to macroscopic variables like temperature.

Pivotal Quotes: "We can't know anything at all about the world without a model for the world." — Andrew Jaffe: Core claim about the necessity of models in science and everyday reasoning. "The answer is you can get more certain of things." — Andrew Jaffe: Summarizing the Bayesian response to the problem of induction. "All probabilities are conditional." — Andrew Jaffe: His compact statement of the Bayesian view that probabilities depend on background information and model choice.

Implications: Listeners should expect scientific claims to be model-dependent and probabilistic, not certain. In practice, Bayesian thinking is especially valuable for cosmology, quantum theory, and complex data analysis, even as interpretation debates continue.