Active Inference, taught with the math actually worked through

I kept hitting the same wall in Karl Friston's work: explainers that gesture at the free energy principle without ever running the equations, and papers that run them without explaining why. So I built the course I wanted — and I'm opening it as a free 12-week pilot cohort (25 seats), starting next week.

This is the version I wish existed when I first hit the active-inference literature: university-level, every equation executable in a clonable Elixir/Jido workbench, every shortcut named out loud.

The math, taught honestly

Most courses blur the parts that are easy to get subtly wrong. This one names them:

• Mean-field VMP throughout. The state-belief update uses (ln B)·s, and the variational free energy uses (ln B)·s as well — the same form across both the update and the functional. No silent marginal/Bethe blend, which is where a lot of implementations quietly diverge.
• Policy posterior σ(ln E − γG − F), with the precision γ placed on the expected free energy G where it belongs.
• Expected free energy = ambiguity + risk, in nats — not a vague "exploration bonus."
• The bound is never inverted. F[q] ≥ −ln p(o|m) stays an upper bound on surprise, always.

What the full 12-week arc covers

1. Perception as inference — the variational free energy F
2. Action as expected free energy G
3. Policy posteriors and precision
4. Markov blankets, made numerical (an actual conditional-independence residual, not just a diagram)
5. Dirichlet learning, wired live — E[ln A] via the digamma function, not ln E[A]
6. A capstone cue task and its five ablations — signed as risk-driven safe cue-seeking, each ablation breaking the agent in a predicted direction

You run it, you don't just read it

Every result in the course reproduces in a clonable Elixir/Jido workbench on the BEAM. Clone it, run mix test, watch the numerical trust gate pass to ~1e-9, watch the ablations fail exactly where the theory says they should.

A one-minute sample from Week 8 (perception as inference via mean-field VMP): https://youtube.com/watch?v=-Jcox5oGAYg

Reserve a seat

The pilot is free in exchange for deep feedback that helps me finalize the materials. 25 seats. To claim one, email Michael.Polzin@SolutionWright.com.