A Treatise on Probability
A Treatise on Probability is Keynes's first major work and, in its philosophical depth, perhaps his most rigorous. Written in fragments over a decade and published in 1921, it argues that probability is not a property of events (the frequency with which they occur in repeated trials) but a logical relation between evidence and conclusion (the rational degree of belief that given evidence warrants in a given proposition). The distinction seems technical. It is in fact the epistemological foundation of all of Keynes's subsequent economic work — and the most powerful available critique of the frequency-based architecture underlying contemporary AI.
In the AI Story
The frequency interpretation of probability — dominant in statistics, machine learning, and the computational infrastructure of every large language model — holds that probability is the ratio of occurrences in the limit of repeated trials. This interpretation works for repeatable events with stable distributions: dice, cards, insurance actuarial tables. Keynes argued that this interpretation fails for the decisions that matter most — decisions made under genuine novelty, where no comparable series of past events provides a frequency to derive probability from.
For such decisions, Keynes argued, probability is logical rather than frequentist: a measure of the warrant that evidence provides for belief. Two rational agents confronting the same evidence should arrive at the same probability assignment, but this probability is not a frequency ratio. It is a judgment about evidential weight.
The framework has direct implications for AI. Large language models are frequency machines — trained on vast corpora, they predict the probability of the next token conditional on prior tokens. Within the training distribution, their predictions are superb. Outside the distribution — in genuinely novel situations — the models continue to produce confident output without any mechanism to register the shift from calculable to uncalculable probability. This is the structural root of confident wrongness.
The Treatise also developed the concept of evidential weight — distinct from probability — that measures the amount of evidence supporting a judgment rather than the judgment's probability. This distinction is critical for AI evaluation: an output can have high probability (high confidence) with low evidential weight (little actual evidence), and the user has no way to detect the difference from the output alone.
Origin
Keynes began the work as a 1908 Cambridge fellowship dissertation and revised it through successive versions before publication by Macmillan in 1921. It remains his most sustained work of philosophy.
Key Ideas
Probability as logical relation. Not frequency but the warrant evidence provides for belief.
Radical uncertainty. Some probabilities cannot be calculated at all — the situations Keynes called genuinely uncertain.
Evidential weight. Distinct from probability itself, measuring the amount of evidence rather than the probability it supports.
Non-numerical probabilities. Some rational degrees of belief admit no numerical expression.
Foundation of economics. The framework underlies the entire Keynesian apparatus of expectation and uncertainty.
Debates & Critiques
The Keynesian interpretation competes with Bayesian, frequentist, and propensity interpretations. The AI application has given the debate renewed urgency: frequency machines that cannot distinguish calculable from uncalculable probability produce specific categories of error that purely frequentist frameworks cannot diagnose.
Appears in the Orange Pill Cycle
Further reading
- John Maynard Keynes, A Treatise on Probability (1921)
- Frank Ramsey, 'Truth and Probability' (1926)
- Frank Knight, Risk, Uncertainty, and Profit (1921)
- Rod O'Donnell, Keynes: Philosophy, Economics and Politics (1989)