We are Bayesians. We come from a long lineage — Laplace, Gauss, Jaynes — and we are building its next instrument.
In 1807, on Berthollet's estate in the village of Arcueil outside Paris, a small society began meeting on weekends: Laplace and Berthollet presiding, and around them the young scientists they mentored — Gay-Lussac, Biot, Arago, Malus, Poisson. The Société d'Arcueil had no charter beyond doing science carefully and together, and it published what it learned as the Mémoires de la Société d'Arcueil. This organization is named for that society, and works the same way: curate what the community actually knows, grade the evidence, publish the memoirs.
Bayes left a theorem; Laplace made it a method. He rediscovered inverse probability independently in 1774 and spent four decades applying it — to the mass of Saturn, the stability of the solar system, the ratio of births, the credibility of witnesses — insisting all along that probability theory is "nothing but common sense reduced to calculation."
Gauss made it a tool. His 1809 derivation of least squares is, read plainly, a Bayesian argument: a normal law for the errors, a uniform prior over the unknowns, and the estimate taken at the posterior's peak. Every regression since is downstream of that page.
Then came a long winter, when inverse probability fell from fashion — and the flame passed through hands that refused to let it gutter: Jeffreys, who rebuilt it as a working scientist's theory of inference; Cox, who proved that any consistent calculus of plausibility must be probability; de Finetti, who rooted it in exchangeability and coherence; Savage, who gave it decision-theoretic teeth.
Jaynes made it a logic. Probability as the extension of deductive reasoning to incomplete information — and he opened Probability Theory: The Logic of Science not with axioms but with a design brief: build a robot that reasons plausibly and consistently from its prior information.
Then computation made it a practice: Metropolis and Hastings, the Gibbs sampler, Gelfand and Smith; Hamiltonian Monte Carlo, born in lattice field theory and brought home to statistics; NUTS, Stan, PyMC, BlackJAX. And with practice came something the textbooks never fully captured — craft. Which parameterization funnels and which doesn't. What a divergence is trying to tell you. When a prior is load-bearing. Two centuries after Laplace, the theory is written down, but the craft lives scattered across forum threads, case studies, and the memories of practitioners.
memoires — the craft, written down. A curated, evidence-graded, contradiction-aware catalog of Bayesian modeling knowledge, distilled from tens of thousands of practitioner discussions and expert case studies, reviewed adversarially, and organized the way Laplace's society organized its science: a small set of principles you can actually read, over a deep, searchable record of what works — and what doesn't, and under which conditions.
jaynes-robot — the reasoner that reads them. Jaynes asked for a robot that reasons from organized prior information; large language models finally make the design brief buildable. The Mémoires are its priors.
« La théorie des probabilités n'est, au fond, que le bon sens réduit au calcul. » — Laplace
Common sense, reduced to calculation.
Images: title page of the Mémoires de la Société d'Arcueil, Tome I (Paris, 1807) and portrait of Laplace by Paulin Guérin — public domain, via Wikimedia Commons (portrait).