Two questions come up before every workshop. What is actually covered, and what am I supposed to already know. This post answers both, in the same language we use inside the room.
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. The workshop is where that hypothesis becomes something a small team can operate.
What the workshop covers
The curriculum walks the discrete-time POMDP formulation of active inference, following Parr, Pezzulo and Friston (2022) chapter for chapter Class E, and mapping each construct onto a lab you can steer in the browser Class C. The five topic blocks:
- Generative models and Markov blankets. How to name the joint distribution over hidden states and observations, and how to draw the boundary between internal, sensory, active, and external states. We build one from scratch inside the Precision Lab configuration.
- Perception as inference. Sensory precision, prior precision, and the Kullback, Leibler divergence between approximate and true posterior. Behavioural regimes emerge from where the dials sit, not from hand-coded rules Class C.
- Action as inference. Expected free energy, the pragmatic term over prior preferences, and the epistemic term over uncertainty reduction. Why active-inference agents look curious without a bolt-on exploration bonus Class E.
- Planning depth and policy temperature. How horizon and temperature shape the policy space, run against the Loop Lab and Echo Lab so participants can watch each dial move behaviour.
- Falsification discipline. The pre-registered Cell Lab benchmark, how the five falsification criteria were written before the runs, and how the honest losses are logged next to the wins Class C.
What the workshop does not cover
We are direct about the fences. The workshop is not a clinical training, not a diagnostic training, and not therapy instruction. Behavioural labels in the labs are candidate computational phenotypes, hypotheses, not diagnoses. It is also not a general machine-learning primer: we assume you already know what a probability distribution is and what conditional independence means. It is not a survey of every active-inference variant in the literature. It is one working stack, taught deeply, with the trade-offs named.
What you should already know
A participant will get the most out of two days if they can already do the following: read basic probability notation, follow a Bayes rule derivation, and hold a matrix operation in their head. Programming background helps but is not required for the conceptual sessions. If you want to touch the Elixir workbench live, bring a laptop with a working development environment.
What you will be able to do after
On the last afternoon, participants build one working POMDP in the workbench, wire a generative model into it, and run it against a small disturbance battery drawn from the Cell Lab family Class C. Concretely, after the workshop you should be able to name the generative model behind an existing system, articulate which precision dial is doing the work in a given regime, and write a falsification criterion that could show your controller is worse than a rule-based baseline. Those three moves are the minimum viable practice. They are also, in our experience, the moves most teams have never been asked to make explicit.