You do not need to love math to teach with active inference. You do need one honest intuition about what a belief is, and how far one belief sits from another. That is what KL divergence measures.
The one-sentence version
KL divergence is a number that says: if I hold belief A about what will happen, and reality (or a better belief B) turns out different, how surprised should I be, on average, and in which direction? It is not a symmetric distance like inches between two points. It is a directional gap: how costly it is to keep believing A when B is the truer picture (Class E, Friston and colleagues use this framing throughout the active inference literature).
That is the whole idea for classroom purposes. The rest is bookkeeping.
Why teachers should care
Every lesson is a small belief update. A student walks in with belief A about how magnets work, or how a sentence should end, or what a fraction is. Something happens in the room: a demo, a question, a peer's answer. Now they hold belief B. The gap between A and B is the learning. If the gap is zero, nothing changed. If the gap is huge and unpleasant, you have shock, not learning. If the gap is right-sized and the student can trace how they moved, you have a lesson they will keep.
KL divergence is the mathematics of that gap. You do not have to compute it to teach it. You have to notice it.
What a fuller treatment requires (the honest part)
If you want the actual formula, it is a sum (or integral) over possible outcomes, weighted by how likely belief A thought each outcome was, of the log-ratio between the two beliefs. That sentence hides three things a real course has to unpack:
- What a probability distribution is, and why beliefs are shaped like one (Class E).
- Why the logarithm is there (it makes surprises add up cleanly when events are independent).
- Why the direction matters, so KL(A to B) is not the same number as KL(B to A).
None of that is out of reach. It is just not a single blog post. If you want the honest, careful walk-through, the strongest prep we know of is Themesis T3, Top Ten Terms in Statistical Mechanics for AI. We recommend it as prep for the UNI Workshop for math-hungry learners who want the equations under the intuition, taught by someone who takes the math seriously (Class E).
We do not paraphrase that course here. We point to it, we send our math-curious teachers there, and we come back to the classroom with cleaner language.
What we teach instead, in the room
For teachers who do not want to become statisticians, the classroom move is smaller and more useful. Ask three questions across a lesson (Class C, this is how the Workbench prompts are structured):
- What did you expect to happen? (belief A)
- What actually happened? (evidence)
- What do you now think? (belief B)
Then, and this is the KL part, ask: how big was the jump from A to B, and can you say why? A student who can name the size and the direction of their own update is doing, in words, what the equation does in symbols. That is enough to teach on. It is not enough to publish on, and we do not pretend otherwise.
Where to go next
- Start by grounding the vocabulary in what active inference actually says, so KL divergence sits inside a story and not in the air.
- Then walk the recommended reading order for teachers if you want the fuller arc before Workshop.
- For a concrete classroom scene using the same three questions, see Bayesian updating in a classroom conversation.
- If the math itch is real, take Themesis T3 before the Workshop. You will get more out of both.