Learn · the guided tier

The UNI Workshop

Build it with us. Then try to break it. A live, mentored 12-week program where you build a working active-inference agent, inspect the gates, and help find where it fails.

Two ways to learn UNI, stacked.

The Workshop is not a separate world from the free CBT. It is the next step after it: the CBT teaches active inference you can read and run by hand, in your browser; the Workshop is where a small group builds it for real, with mentors, over twelve weeks.

101 · free, self-paced

The CBT

Fifteen interactive modules you run in the browser, free, no account. You build free energy by hand, from a numberless hunch to a mind that acts. This is the 101 foundation, and the prep we ask you to walk first.

Open Module 1 ›  ·  Browse all 15 modules ›

201 · live, mentored

The UNI Workshop

The same active inference, taken as a guided build cohort. You ship a running, non-LLM agent and sit a certification exam. Twelve weeks, with people, in the open. Taught live by Michael Polzin.

Apply for the next cohort ›  ·  Scholarships & grants

Twelve weeks. You ship, not just watch.

No prior coding background is required to start, but it is not effortless. The work is technically scaffolded, built gently into the math, and hands-on with the open Active-Inference Learning Workbench from week one.

By week 5

Your first running generative model

Not a slide. A model that runs on your own machine, that you built and can inspect.

By week 12

A non-LLM agent on a CPU

A working, constrained-domain active-inference agent with inspectable state, preferences, and uncertainty, running on an ordinary CPU.

Throughout

The workbench, in your hands

You clone and follow the open Workbench, or watch the lab run live. Every step reproduces.

Said plainly

“No LLM” means no large language model at the agent’s runtime. Any use of other tools for teaching, documentation, or debugging is disclosed separately and is not part of the deployed agent.

“Working” means a constrained-domain agent with inspectable state and predictable behavior, not open-domain, general-purpose chat.

The discrete-time POMDP formulation, chapter for chapter.

The curriculum follows Parr, Pezzulo & Friston (2022), Active Inference (MIT Press), and maps each construct onto a lab you can steer in the browser. Five topic blocks, plus the discipline that keeps the claims honest.

Generative models and Markov blankets

Name the joint distribution over hidden states and observations, and draw the boundary between internal, sensory, active, and external states.

Perception as inference

Sensory precision, prior precision, and the KL divergence between approximate and true posterior. Behavioural regimes emerge from where the dials sit, not from hand-coded rules.

Action as inference

Expected free energy: the pragmatic term over preferences and the epistemic term over uncertainty. Why these agents look curious without a bolt-on exploration bonus.

Planning depth and policy precision

How horizon and policy temperature shape the policy space, run against the labs so you watch each dial move behaviour.

Falsification discipline

The pre-registered benchmark, how the falsification criteria were written before the runs, and how the honest losses are logged next to the wins.

The correctness discipline

Bound directions, model versus process, the limits of the analogies. You learn to state things exactly: hypothesis, prediction, observation, provenance.

A real challenge needs public rules.

These are the gates the build must pass: what is tested, what would count as failure, and what would disconfirm the claim. Where claims are public, the gate definitions are published: the test, the tolerance, the reproduced example, the expected failure under ablation.

Build gate

Nine structural checks

Probabilities conserve, the free-energy bound holds, actions follow the posterior, the Markov-blanket boundary is respected. If a check fails, the build fails, publicly.

Trust gate

Deterministic agreement

Specified reference calculations reproduce to roughly 1e-9. This tolerance applies only to those named tests, never as a global claim about intelligence or real-world validity.

Capstone gate

Ablations must break it

Remove the right piece and the agent must fail in the predicted way. A model that cannot be broken on cue was never doing the work it claimed.

On the method

The methods are aligned with the worked examples in Parr, Pezzulo & Friston (2022). Where a mapping is public, the exact examples, equations, and tolerances are shown. UNI’s internal method stays private; the gates that any correct build must satisfy do not.

Learning with another person, in the open.

Part of the same workshop is trauma-informed learning design and relational reflection practice: consent, repair, co-regulation, and making uncertainty explicit when a learning interaction misattunes.

To be clear

This is educational, not therapy. UNI does not diagnose, treat, cure, or claim to reduce trauma. The relational strand teaches trauma-informed ways to learn with another person: slowing down, checking predictions, lowering blame, practicing repair.

What the Workshop is, and is not.

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. It is not an announcement that we have arrived.

A preprint and reproducibility materials are archived on Zenodo. Expert peer review is invited and pending; until it completes, treat the work as a preprint and build-audit artifact, not an authoritative scientific result.

We do not claim the Workshop produces achieved general intelligence, consciousness, biological equivalence, or a finished system. On some standard benchmarks a larger general model scores higher, and we say so.

It is not a clinical or machine-learning survey course, and not a promise of a production controller by the last week. It is one working stack, taught deeply, with the trade-offs named.

What is wrong ›  ·  The Claim Ledger ›  ·  The preprint on Zenodo ›

What it is, and what it costs.

Length
12 weeks, hands-on with the workbench, building gently into the math.
Who it is for
Designed for non-programmers to start: motivated, not effortless. Technically scaffolded, not technically empty.
By week 5
Your first running generative model.
By week 12
A working non-LLM active-inference agent on a CPU (constrained-domain, inspectable).
Tuition
$75,000 USD for the full 12 weeks, plus a 4-hour examination to sit for certification.
Scholarships
Partial to full ride through the EducateWright grant, so cost is not the gate.
Outcome
You can build and ship, and you can read the gates to judge the claim for yourself.

Apply through EducateWright.

Enrollment, scholarships, and payment are handled by EducateWright, the education arm that fronts training and the grant. One honest application: curiosity and the willingness to be corrected matter more than credentials. Scholarships run from partial to full ride. This page collects nothing; the button takes you to EducateWright’s own application.

Not ready to apply? Walk the free prep first: the CBT, the interactive labs, and Prep Work Before the Workshop.

See the work for yourself.