If you teach, you already know the shape of the problem. A student hands in an essay that reads clean, cites nothing, and cannot be re-derived. The tool that made it is a black box, the classroom time spent teaching that tool is time not spent teaching thinking, and the receipts do not exist. There is a different subject we can teach instead, and it has gates you can inspect.
The choice we are actually making
Two curricula are on the table. One teaches LLM tooling: prompt patterns, jailbreak avoidance, "which model for which task." The other teaches Universal Natural Intelligence (UNI), a working hypothesis on an attainable path toward General Natural Intelligence, natural not artificial. Under UNI, the machinery is active inference: generative models, Bayesian updates, Markov blankets, and free-energy minimization (Class E, per Parr, Pezzulo, and Friston, 2022). The subject is legible. The parts have names. A student can point to a gate and ask, calmly, "why did this update the way it did?"
We are not neutral on this. We think the LLM-tooling curriculum trains guessing behavior. It rewards the student who intuits the model's mood. UNI-as-subject trains the opposite: state your prior, state your generative model, state what would falsify the update, and show the receipt (Class F).
What "teach UNI" concretely means in a classroom
Teaching UNI is not teaching a product. It is teaching a family of ideas the student can hold in their hands.
- Generative models as objects of study: what the model claims the world is like, and how it would be wrong.
- Bayesian inference as a small, honest arithmetic: prior in, evidence in, posterior out, and every step is visible (Class B, since the update code is inspectable and re-runnable).
- KL divergence as a way to measure "how surprised was the model, and by what." Not a mystery, a distance.
- Markov blankets as a bookkeeping tool for what the system is allowed to see and touch.
- Free-energy minimization as the through-line: the system prefers models under which the world is less surprising, over time.
A middle-schooler will not derive the variational bound. That is fine. A middle-schooler can absolutely draw a prior, update it against three observations, and notice that the posterior moved. That is the whole game in miniature. The subject scales up cleanly into undergraduate work, into Parr and Friston (Class E), and into current open questions we do not pretend to have closed (Class U).
The Themesis map, and why it matters here
The field is being mapped in public. Themesis published a resource map that lists several routes into active inference for 2026, and our workshop is one of the pathways on that page. Where to Start with Active Inference, A Resource Map for 2026 is the map: our honest one-line frame is that it names five reasonable places a learner could start, and it puts our workshop in that list. It is not a credential we hold. It is a signal that "learn active inference" is now a live curricular option, not a niche.
For teachers deciding where to invest classroom minutes, the second Themesis piece is the wider context. The AGI Landscape Just Changed is our field-level backdrop: our honest one-line frame is that the terms have shifted, and pretending the LLM-tooling curriculum is future-proof is now the risky bet.
The honesty fence, non-negotiable in the classroom
If we teach UNI as subject, we owe students the fence around what UNI is and is not. This is doctrine on every EducateWright artifact, and it belongs on the wall of the classroom.
- UNI is a working hypothesis. It is not "artificial intelligence." It is not "AGI." It is a natural, active-inference approach whose evidence is growing, evidence-classed, and tested in the open. Do not take the claim on faith. Test the build, inspect the gates, and help us find where it fails.
- Every claim carries an evidence class: A empirical-in-session, B code and inspection, C configuration and integration, E expert citation, F falsifier present, U unverified. A claim without a class does not go on the board.
- AI-authorship disclosure. Any classroom artifact that used an assistant to help draft, summarize, or format carries the disclosure. Not because assistants are shameful, but because "who wrote which sentence" is a fact, and facts are the whole subject. See ai-authorship-fence-in-the-classroom for the exact language and the routing rules.
- No claims of endorsement, no claims of "recognized as," no claims of "proven." Those are receipts we do not have. We link, we cite, we frame in our own voice.
The fence is a teaching artifact. It shows students what honest scoping sounds like. Once they have seen it a few times, they start doing it to their own work, which is the actual objective.
What students inspect, week by week
The curriculum has gates in it. That is the point. A gate is a checkpoint where a student can literally look at the state of the system, name what changed, and say why (Class B). Contrast that with a chatbot exercise, where the "why" is always inaccessible.
Sketch of the arc, and each piece links out to its own post in this cluster:
- The generative model as a drawing on the whiteboard, then as a small program. See curriculum-gates-that-teach-thinking for the specific gates and rubrics.
- The workbench itself: a set of small tools a teacher can run on their own machine, that show the student the update happening. See the-teachers-workbench-tour for the tour of what is on it today (Class C, since the integration between the bench and the classroom LMS is documented and reproducible).
- Falsifiers and re-runs: every notable claim in the semester ends with "here is what would show this is wrong" (Class F). This is a habit, not a lecture.
- Open questions: the parts of UNI that are Class U. We do not hide them. Students learn what "we do not know yet" looks like when it is written down honestly.
Where LLM tooling still fits, and where it does not
We are not banning LLMs from the school building. We are saying they are not the subject. A calculator has a legitimate place in a math classroom, and it is not the subject either. If a student uses an assistant to reformat a table, disclose it and move on. If a student is being taught prompt engineering as the pedagogical core, we think the school is spending its finite minutes on the wrong subject.
The deeper reason: LLM tooling trains a student to be a good user of one company's product. UNI-as-subject trains a student to reason about any inference system, including future ones and including the one between their own ears. The transfer is real. The receipts are inspectable. The falsifiers are on the page.
Where to go from here
- Book the Workshop if you want to see the workbench and the gates in a live classroom setting.
- Read the Transparency page for the evidence-class rubric and the exact honesty fence language.
- Walk the cluster: start with curriculum-gates-that-teach-thinking, then the-teachers-workbench-tour, then ai-authorship-fence-in-the-classroom.
- Bring a colleague. The subject is more fun to teach in twos, and the fence is easier to hold when a second teacher is holding the other end.