David Matolcsi Argues Probabilities Are Not the Right Concept — What Replaces Them?

David Matolcsi opens his LessWrong sequence by questioning the very foundation of probability. He argues that subjective Bayesian credences—'probabilities are just in my head'—fail to explain why some probability assignments are clearly wrong, such as a 50% chance of Bigfoot in the next room. Frequentist definitions, which rely on sequences of similar events, break down for one-off questions like 'probability that the Russia-Ukraine war ends in 2026.' Matolcsi also examines two common answers: defining an objective prior (via simplicity or Occam's razor) and defining probabilities through betting odds. He finds both valuable but ultimately insufficient, as the problem of induction (e.g., why expect the sun to rise tomorrow after seeing it only until June 1?) remains unsolved. The post sets the stage for a proposed replacement of probabilities with a different concept, drawing on work by Wei Dai, Paul Christiano, Joe Carlsmith, and others.

This critique matters beyond philosophy. If probabilities are not the right tool, then Bayesian reasoning—the backbone of modern AI, forecasting, and decision theory—may need a new foundation. Matolcsi hints that future posts will sketch a unified framework for dealing with infinite ethics, anthropics, and the origin of priors. For professionals in AI safety, epistemology, and rationalist communities, this work could challenge core assumptions about how to handle uncertainty in complex systems. The post is deeply technical but accessible to those familiar with Bayesian debates, and it signals a potential shift toward reasoning primitives that are more robust than traditional probability theory.