terms in the sum
The number you built in Module 1 was your model’s. This module asks what the true surprise is — and discovers it hides behind a wall no mind can walk through. You’ll build the sum by hand, watch Module 1’s golden numbers fall out of it, then feel it explode past any hope of finishing.
Same five doors as Module 1 — the plain line runs straight through; take whichever extras fit: ▶ Picture it, ✎ Try it (commit before the reveal), ✓ Check it yourself, ▸ Go deeper, ⬡ See it in UNI. Runs locally in your browser — no account, no network — and adapts to light and dark.
The one fence that never moves. “Surprise,” “evidence,” “belief” are functional quantities — numbers a model computes about its own predictions. For you we use felt words, because you’re a person reading a sky; for UNI we use functional words — it reads a hint through a model and its dynamics move a number. Neither is awareness, feeling, or sentience.
No scrolling back to Module 1. A sailor comes up on deck at dawn. The swell is longer than yesterday, the air has gone heavy, the gulls have quit the water. Something leans. Rebuild the idea cold — in a world you’ve never used for it before.
From memory, in a brand-new domain (no rain, coins or dice): name the sailor’s numberless lean, commit a probability you build yourself, then the squall hits and you read the surprise. No re-reading — reconstruct it.
In one sentence, with no peeking: the sailor wasn’t handed a number — so where does a number even come from?
If that came back cold, you’re ready. Now the crack in that number.
Back to your own doorway — grey-green morning, damp air, your lean of 0.60 toward rain. In Module 1 you scored your surprise against your 0.60. But your neighbour, with a different life of skies, stood in the same doorway and leaned 0.40. Same morning, same rain, two different surprises. So neither is “the” surprise — each is a model’s surprise.
That pushes a harder question: what is the surprise of the morning in itself — the surprise of the sky you saw, not filtered through one person’s lean? And the moment you ask it, a thing you’ve been ignoring steps forward: the sky didn’t appear from nowhere. Something caused it — a real weather system you never get to see. You only ever get the sky, the hint. Call that hidden thing the cause.
Shrink the world to two causes, and borrow Module 1’s exact numbers so you can see the seam:
| Hidden cause η | Prior p(η) — how likely the cause | p(sky | η) — this wet-grey sky, given the cause |
|---|---|---|
| front ON (a real front) | 0.7 | 0.9 |
| front OFF (no front) | 0.3 | 0.2 |
How probable is today’s sky itself — not assuming any one cause, but all told? There’s one honest way. The sky can arrive by either road: the front is on and it produced the sky, or the front is off and it produced the sky anyway. Add the two roads — but commit your arithmetic first.
Compute the two products by hand and commit them; the tool adds them to the evidence, and Module 1’s surprisal and posterior drop out of that one sum. It was this sum all along.
So the true surprise of an observation is −ln of the evidence, and the evidence is a sum over the hidden causes. The one earned symbol of this rung:
p(y) = Ση p(y | η) p(η)
the probability of the observation y is the sum, over every hidden cause η, of (chance that cause produces y) × (chance of that cause).
Numerically in our world the evidence is 0.69, so the true surprise of the sky is −ln(0.69) = 0.3710636814 nats — Module 1’s sealed surprisal, now revealed to have been this very sum all along.
Maybe you’re thinking: forget the true surprise — I’ll just update my belief about the front and move on. That’s Bayes, the honest update. But look what it needs:
p(front on | sky) = p(sky | on) p(on) / p(y) = 0.63 / 0.69 = 0.913
That 0.913 is — again — Module 1’s posterior, exactly. And notice the denominator: it’s p(y) = 0.69, the very evidence sum from §2 (the tool above already showed it falling out). Even to update your belief, you must divide by that same sum over all causes. There is no back door.
Our sum had 2 terms. Suppose the morning’s cause isn’t just front / no-front, but one of many distinguishable weather situations — each needing its own term. Turn the dial and watch what happens to the sum you’d have to finish. (This isn’t a graph of a formula — it’s the count of terms, and its cost.)
The evidence sum has exactly ONE term per hidden cause. Add causes and the term count — and the work to finish the sum — grows with it. Find where “finishable” ends.
You now know three things precisely:
So a mind cannot compute the thing it most needs. And yet minds — you, an animal — plainly do make sense of the world, in real time, blind to the cause. (The organism is being built to do the same — that’s the project, not yet a finished claim.) Not by finishing the sum. By a trick: instead of the true belief it can’t reach, a mind carries a belief it can move — a guess it’s allowed to slide around — and a score it can actually compute that tells it which way to slide. Build that movable belief and the score, and you’ve built your way around the wall. That movable belief is Module 3; the score, two modules on, is the one this book is named for.