Cluster: active-inference-fundamentals

Reading Order for Newcomers to Active Inference

By Michael Polzin. Published 2026-07-01.

Active inference is not one paper you read on a Sunday afternoon. It is a stack: a probability layer, a control-theoretic layer, a variational-inference layer, and a modeling layer for perception and action under a generative model. The order in which you meet those layers matters. This post is the order we recommend, with the prerequisite at each step named honestly.

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 the claim on faith. Test the build, inspect the gates, and help us find where it fails. This reading order is one way in.

Stage 0, before you start

You do not need measure theory. You do need comfort with three ideas: a probability distribution over hidden states, expectation as a weighted average, and the log of a ratio as a way to score surprise. If those words are unfamiliar, spend an evening with any solid undergraduate probability text first. Everything downstream sits on that base (Class E, standard curriculum).

Stage 1, orient with the family map

Start with the resource map from Themesis: Where to Start with Active Inference, A Resource Map for 2026. In our voice: this is an outside map that lists SWU among five pathways into the field, and it is the cleanest single anchor for seeing where each family (theirs, ours, others) sits (Class E, published resource map).

Also read our own vocabulary primer, so the words carry the same load across everything that follows: A glossary of terms we actually use.

Stage 2, the textbook

The center of gravity is Parr, Pezzulo, and Friston (2022), Active Inference: The Free Energy Principle in Mind, Brain, and Behavior, MIT Press. You do not read it front to back on the first pass. Read the setup chapters on generative models and the discrete-time POMDP formulation, then jump to the chapters on precision and expected free energy. We wrote a companion for exactly this path: Reading Parr, Pezzulo, Friston (2022): a guided tour. It marks where to slow down and where you can skim without losing the thread (Class E, textbook; Class C, our companion notes).

Stage 3, see it move

Open a browser tab and drive the Precision Lab. Three dials: sensory precision, transition precision, policy temperature. Change one at a time and watch the behavioral regime shift. Then try the Echo Lab for range-2 echolocation and the Loop Lab for a 2-state bifurcation map that makes the precision claim visible (Class B, live in-browser labs; the code is inspectable, the outputs are reproducible).

This is the point where the textbook math starts to feel like a control law rather than a proof.

Stage 4, math-hungry side quest

If you want more mathematical grounding before returning to the textbook, Themesis offers a short course on the statistical-mechanics vocabulary that shows up throughout active inference: T3, Top Ten Terms in Statistical Mechanics for AI. In our voice: we recommend this as prep for anyone who wants the math scaffolding before booking a seat at our workshop (Class E, external course).

Stage 5, hands on a different stack

If you code and want a second, complementary implementation path, Themesis runs a Python course: Building Active Inference in Python. In our voice: this is a different stack than our Elixir workbench, and reading both is how you tell which parts of a working system are the theory and which parts are the language (Class E, external course).

Stage 6, video companion

For deep-dive explanations, the Themesis YouTube channel walks through active-inference concepts on the whiteboard. In our voice: recommended for learners who retain concepts better by watching someone work them out loud (Class E, external video library).

Stage 7, the falsification benchmark

Once the vocabulary is stable, read The Science page and the Cell Lab. The benchmark is pre-registered and shows both wins and losses. That is the honest test: a single active-inference controller is not universally best, and the benchmark is designed to surface exactly that (Class B, pre-registered benchmark with committed cache and open loss reporting).

What to do next