The MIT Press textbook Active Inference: The Free Energy Principle in Mind, Brain, and Behavior (Parr, Pezzulo, Friston, 2022) is the canonical reference for the framework this site runs on. It is also dense in the first hundred pages. This is the order we recommend, the prerequisites we assume, and the exercises we found paid the highest dividend. Class E Class C
The recommended order
1. Chapter 1: Motivation, before the math
Read this in one sitting, cover to end. Do not stop to look things up. The point is to install the frame: a system that persists must minimize an upper bound on surprise, called variational free energy, through perception (updating beliefs) and action (changing observations). Class E Everything after chapter 1 is machinery for that one sentence.
2. Chapter 2: The math, slowly
This is where most readers stall. The chapter derives the free energy bound, expected free energy for policy selection, and precision as inverse variance on beliefs. Read it twice. On the first pass, skim past every equation and read only the prose. On the second pass, work equations 2.2 through 2.10 by hand on paper. The identity that free energy equals accuracy minus complexity (Parr et al., 2022, section 2.4) is the single most useful line in the book for building intuition about model design. Class E
3. Chapter 4: Discrete state space models (POMDP)
Skip chapter 3 on the first read. Go straight to chapter 4. This is the POMDP formulation our labs use: a hidden state, an observation model A, a transition model B, prior preferences C, initial beliefs D, and policies as sequences of actions. Once you have chapter 4 in hand, the Precision Lab on this site is legible; the three dials on that page are the sensory precision, transition precision, and policy temperature from this chapter (Class B, in our source). Class C
4. Chapter 3: Continuous state space models
Now return to chapter 3. Predictive coding, hierarchical generative models, and the continuous time formulation used in most neural applications. It sits more comfortably after chapter 4 because you already have a concrete example of a generative model in your head.
5. Chapter 5 and beyond: Applications
Chapters 5 through 11 are applications: message passing, neural process theories, action selection, learning, structure learning, and links to cognition. Read the introduction of each and dive into whichever matches your interest. This is where the book stops being a textbook and starts being a research map.
Exercises that carry the weight
- By hand. Compute one step of belief updating for a 2 state, 2 observation POMDP with numeric A, B, C, D. Do it on paper before you touch code. Twenty minutes here saves a week later.
- In code. Reproduce figure 4.4 or its equivalent (a two step planning example) in any language. If Python is your stack, the Themesis Building Active Inference in Python course walks a complementary path in a different toolchain than our Elixir workbench. Class E
- In the browser. Move the sensory precision dial on our Precision Lab from low to high and watch the policy space collapse. That is chapter 4 running under your fingers.
Where the book connects to the wider map
The Themesis 2026 resource map lists our site alongside four other active inference pathways (Where to Start with Active Inference, A Resource Map for 2026). Our honest one line frame: the map lists SolutionWright among five pathways, and this reading guide is the on-ramp we recommend for the textbook that anchors most of them. Class E
What this guide is not
Not a summary. Not a substitute for the book. Not a claim that finishing the book makes you a practitioner. It is a route that saved us months on the second read. 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 that claim on faith. Test the build, inspect the gates, and help us find where it fails.