Why We Cite Parr, Pezzulo, Friston (2022)

When we write about prediction, error, and felt safety in plain language, one book is doing the heavy lifting under the floorboards. This piece names it, tells you what the citation is doing, and tells you what it is not doing.

The book is Thomas Parr, Giovanni Pezzulo, and Karl J. Friston, Active Inference: The Free Energy Principle in Mind, Brain, and Behavior, MIT Press, 2022. It is the current textbook consolidation of the active inference research program, and it is the source we lean on whenever this cluster uses words like generative model, prediction error, or free energy. Class E

What the citation is doing

Three things, no more.

• It grounds our vocabulary. When a primer here says a learner is always predicting, that word is not casual. It points at a formal object in an active inference model: a generative model that produces expectations about sensory input, updated by prediction error. The 2022 textbook is where the object is defined. Class E
• It points readers who want the math to a real place. The book has the equations, the derivations, and the worked examples. If a reader wants to check whether the words on this site are grounded, that is the volume to open. Class E
• It fences off overclaim. Citing a specific textbook is a discipline. It forces us to say, when we go past what the textbook covers, that we are going past it. That is where the evidence class tags do their work.

What the citation is not doing

Four fences, said plainly.

• It is not an endorsement of this site by the authors. We cite their book. They have not read our blog. Read any implied approval as our failure to be clear, and tell us where the sentence reads that way.
• It is not a claim that the textbook has been mapped to our program equation by equation. Where we make a mathematical claim about our own build, the receipt is a separate document on the science front, with the specific mapping. Until that receipt is published for a given claim, the honest label is Class U, unverified. Class U
• It is not a claim that active inference is settled science across every domain it touches. The core framework is well developed. Specific neuroscientific mappings to brain regions, and specific clinical translations, are still in active debate, and we do not paper over that here. Class F A straightforward falsifier: if well designed learning experiments consistently show update rates rising under higher predicted threat with cognitive load held constant, the felt safety claim we build on is in trouble.
• It is not a substitute for reading the book. Our primer is the on ramp. The textbook is the destination.

Where the technical spine actually lives

IamHITL is the plain language on ramp. The math itself, the generative model structure, the KL divergence formulations, the Markov blanket picture, the free energy minimization derivations, lives on the sister site, Universal Natural Intelligence. UNI is a working hypothesis on an attainable path toward General Natural Intelligence: a natural, active inference approach whose evidence is growing, evidence-classed, and tested in the open. Do not take it on faith. Test the build, inspect the gates, help us find where it fails. Class C

Reading order: this post, then the primer on how prediction shapes experience, then the science page at UNI for the formal treatment. If you want the frame that keeps us out of the "AI" language, read why we say natural, not artificial.

The rule we hold ourselves to

If a sentence on this site uses active inference vocabulary and cannot be traced to Parr, Pezzulo, Friston (2022), or to a specific paper on the science page, that sentence is either weakened, cited to the right source, or cut. That is the whole rule. It is the same rule that keeps the rest of the family honest, applied to the one place readers are most likely to be dazzled: technical language. Full method and taxonomy live at the transparency page.

Evidence classes used in this post: Class C configuration and integration, Class E expert citation, Class F falsifier present, Class U unverified. Full method and taxonomy at the transparency page.