Series: The Learn Arc — 50 posts through the Active Inference workbench.
Previous: Part 49 — Session §10.3: Where next
Hero line. Fifty posts. Ten chapters. One framework. The Learn Arc closes here — with a reader's map, a short what-to-keep list, and a pointer to what is worth building next.
What the Arc covered
- Posts 1–11 — The orientation arc. Why a BEAM-native workbench; the ten chapters in one page each; the single-loop view that runs under every chapter.
- Posts 12–22 — Chapters 1–3 up close. Inference as Bayes; why free energy; the high road; expected free energy; epistemic vs pragmatic value; softmax policy; what makes an agent active.
- Posts 23–27 — Chapter 4: A, B, C, D. The four matrices, the POMDP world, the first shippable agent.
- Posts 28–31 — Chapter 5: The cortex. Factor graphs, predictive coding, neuromodulation, the brain map.
- Posts 32–34 — Chapter 6: Shipping an agent. States/observations/actions, filling A-B-C-D, run and inspect.
- Posts 35–39 — Chapter 7: The muscle chapter. Discrete refresher, Eq 4.13 in depth, Dirichlet learning, hierarchy, the capstone worked example.
- Posts 40–43 — Chapter 8: Continuous time. Generalized coordinates, Eq 4.19, action on sensors, the continuous sandbox.
- Posts 44–46 — Chapter 9: Fit to data. Parameter inference, model comparison, a case study.
- Posts 47–49 — Chapter 10: Synthesis. One equation with three gradients, the honest limits, the roadmap.
What to keep — five things
-
One free energy, three gradients. Perception (
∂F/∂μ), action (∂F/∂a), learning (∂F/∂θ). Every Active Inference agent is this sentence. - A, B, C, D is the design contract. Shapes before semantics; semantics before code; code before inference. In that order.
- Eq 4.13 is message passing. Softmax is not folklore; it is the normaliser of a two-node factor graph's posterior.
- Hierarchy is a taller graph, not a new algorithm. A top-level posterior becomes a lower-level prior via the same message.
- Fitting is inference one level up. Parameters are latents; model comparison is free-energy comparison. Self-similar all the way up.
What to skim
- The biology chapters are beautiful but optional for engineering readers. Keep them bookmarked for intuition.
- Continuous-time math rewards a second pass after you have shipped at least one discrete agent.
- The limitations session is worth re-reading after you try to fit your own data. It will read very differently the second time.
Where the workbench sits
The ORCHESTRATE Active Inference Learning Workbench is one of several teaching/engineering surfaces for Active Inference. It is distinctive in three ways:
- Pure BEAM / Jido. No Python agent runtime; every agent is an Elixir process. Fault tolerance is a feature, not a surprise.
- Sessions and labs coexist. The book's structure is preserved in 39 sessions; labs are reusable across sessions so the same "Bayes chips" demo can anchor four different lessons.
- One screen per concept. The LiveView UI shows belief, matrices, EFE, and surprise simultaneously — the debugger you wish papers came with.
What to build next
If you got this far, the three best first projects are: (a) port the worked example (Session §7.5) to a task from your own domain; (b) run Session §9.1's fitting pipeline on a small behavioral dataset; (c) implement one item from Session §10.3's roadmap and write it up.
Whichever you pick, the framework is now yours. Fifty posts is a long path — thank you for walking it. The workbench is open-source; the issues tab is open; the next arc starts whenever you do.
Powered by The ORCHESTRATE Active Inference Learning Workbench — Phoenix/LiveView on pure Jido.
Top comments (0)