The course

A complete, free, open-source active-inference curriculum: 11 chapters, 39 sessions, 435 minutes of guided work end to end. Every session is written four times over, so a ten-year-old and a mathematician walk the same corridor.

Read this first. This syllabus is not typed onto this page. It is parsed at build time out of the two Elixir data modules the running Workbench serves its curriculum from — active_inference/apps/workbench_web/lib/workbench_web/book/sessions.ex and active_inference/apps/workbench_web/lib/workbench_web/book/chapters.ex, read at commit c2c9c7246251— and the build fails if the parse disagrees with the modules’ own declared totals, so a shortened course cannot ship quietly. What this page cannot claim: that a Workbench is running anywhere right now (this is a static export, not a probe), or that the minute figures measure anyone’s reading — they are the course’s own declared session lengths. The curriculum is built around the textbook Active Inference(Parr, Pezzulo & Friston, 2022, MIT Press). The book is not redistributed here; the course pairs its own sessions, labs, quizzes and narration with the reader’s own copy.

chapters, the preface included

sessions, each in four registers

435

declared minutes — 7 h 15 min end to end

TheORCHESTRATEActiveInferenceWorkbench @ c2c9c7246251 — chapters.ex · sessions.ex — the repository is fully public, and both links open the exact files, at the pinned commit, that these counts were parsed from. (read at 2026-08-31T07:18:53.201Z)

Four ways through every session

Every session exists in four registers. Pick one, switch at any moment — the corridor is the same corridor, only the voice changes. This is how one syllabus serves a child and a mathematician without lying to either.

kid: second person, grade-5 vocabulary, one concrete image
real: plain English, grade-8 vocabulary, one analogy
equation: the math in Unicode, tied to the labelled equation
derivation: formal voice — proof sketch or citation

The syllabus

Chapter 0 is the preface — it is part of the 11-chapter count above, not extra to it.

0. Preface

preface · 1 session · 6 min · this chapter is the preface

Why this book exists — and how to read it.

Orientation — how to use this suite — 6 min

1. Overview

theory · 3 sessions · 26 min

Perception, action, learning — one loop, one theory.

What Active Inference claims — 10 min
Perception and action — one loop — 10 min
Why one theory — and what this book covers — 6 min

2. The Low Road to Active Inference

theory · 4 sessions · 46 min

From Bayes' rule to variational free energy — the minimal machinery.

Inference as Bayes' rule — 12 min
Why free energy — bounding surprise — 12 min
Variational free energy, decomposed — 10 min
Action as sampling from a prior — 12 min

3. The High Road to Active Inference

theory · 4 sessions · 42 min

Expected Free Energy: the value of a plan, as a bill with two lines.

Expected Free Energy — scoring a plan — 12 min
Epistemic vs pragmatic value — 12 min
The softmax policy — 8 min
What makes an agent 'active' — 10 min

4. The Generative Models of Active Inference

theory · 5 sessions · 57 min

Every belief, every action, every thought — inside one generative model.

Why generative models — the engine of AIF — 10 min
A as emission — observation model — 8 min
Expected Free Energy — your first look — 12 min
A full discrete MDP — 12 min
Build your first agent — 15 min

5. Message Passing and Neurobiology

theory · 4 sessions · 48 min

The cortex as a factor graph — and the neuromodulators as precision knobs.

Factor graphs and message passing — 12 min
Predictive coding in hierarchy — 12 min
Precision, neuromodulation, and ACh/NA/DA/5-HT — 12 min
Anatomy of belief updating — 12 min

6. A Recipe for Designing Active Inference Models

practice · 3 sessions · 32 min

Ship your first agent — what's hidden, what's seen, what costs what.

What's hidden, what's seen, what costs what — 10 min
Fill in A, B, C, D — 12 min
Ship your agent — run, glass, iterate — 10 min

7. Active Inference in Discrete Time

practice · 5 sessions · 61 min

POMDPs in full colour — message passing, Dirichlet learning, hierarchy.

Discrete time — a refresher — 8 min
Eq. 4.13 message passing, unpacked — 14 min
Dirichlet learning of A and B — 12 min
Hierarchical discrete AIF — 12 min
A worked discrete example — 15 min

8. Active Inference in Continuous Time

practice · 4 sessions · 51 min

Motion of the mode is the mode of the motion.

Generalised coordinates — motion of the mode — 12 min
Eq. 4.19 — the quadratic free energy — 14 min
Action on sensors — u̇ = −∂F/∂u — 10 min
Open sandbox — play with the forge and tower — 15 min

9. Model-Based Data Analysis

practice · 3 sessions · 36 min

Fit an Active Inference model to real data — and know when to trust it.

Fitting AIF models to data — 12 min
Bayesian model comparison — 12 min
A worked case study — 12 min

10. Active Inference as a Unified Theory of Sentient Behavior

practice · 3 sessions · 30 min

Where the theory goes — and where it bends.

One machine — perception, action, learning — 12 min
Where the theory bends — 10 min
Where to go next — 8 min

Run the whole Workbench yourself

The deep surfaces of this course — the visual model builder at /builder, the glass inspection views, the tutor — are a Phoenix/LiveView application. A static site cannot embed a live BEAM process, and this one does not pretend to; those surfaces are run locally instead. One command, no account:

# from repo root
./scripts/start_suite.sh      # macOS / Linux / Git-Bash
./scripts/start_suite.ps1     # Windows PowerShell

quoted verbatim from RUN_LOCAL.md in TheORCHESTRATEActiveInferenceWorkbench @ c2c9c7246251 — not paraphrased, because a paraphrased command is an invented one

And to stop it:

./scripts/stop_suite.sh
./scripts/stop_suite.ps1

quoted verbatim from the same RUN_LOCAL.md @ c2c9c7246251

Said plainly rather than discovered later: the launcher boots more than the Workbench. It also starts the optional local services behind the narrator and the tutor’s help drawer, and none of them is required — the same RUN_LOCAL.md documents, service by service, how the app degrades when each is absent. First-run setup (dependencies, content sync) lives in that file too, and is deliberately not re-typed here, where it would rot.

The source is fully public at the pinned commit this whole page was parsed from: TheORCHESTRATEActiveInferenceWorkbench @ c2c9c7246251.

And some of the course needs no BEAM at all: the learning labs are standalone simulations that run in the browser, and those run right here. (Seven of them, by RUN_LOCAL.md’s own count at c2c9c7246251 — that figure is quoted from the source, not measured by this page.)

This course is one door of the estate. The hallway lists the rest, each with the state it is honestly in and the commit its numbers were read from.