the second hundred · metaphor 178
Do you experience the world as it is — or do you mostly experience your expectation of it, corrected only where reality shoves back hard enough to notice? And when the two disagree, which one do you bend: your belief, or the world itself?
It rarely feels this way from the inside. Perception feels like a window: the world pours in, and you see it. But you can only ever meet the world through a fog of noisy, ambiguous signals, and to make anything of them your brain has to bring a guess — a prediction of what's out there — and compare. What you consciously experience is, to a startling degree, that prediction, trimmed by whatever surprise the senses could not explain away.
And there are two ways to make a surprise go away. You can change your mind to fit the world — ordinary perception, updating the guess. Or you can change the world to fit your mind — reach out and act until what you sense is what you predicted. One theory says the brain is doing nothing but this, everywhere, all the time: driving down the gap between prediction and evidence by whichever route is cheaper. Perceiving. Acting. The same single quantity, minimized both ways.
The mathematical idea
Give the agent a single scalar to hate: free energy, essentially the weighted prediction error, F = ½·π_s·(s − μ)² + ½·π_p·(μ − η)². The first term punishes the gap between what it senses, s, and what it believes, μ. The second punishes the gap between that belief and its standing expectation η — the prior it brought to the encounter. The weights π are precisions: how much the agent trusts each channel.
Now let it roll downhill. Perception lowers F by adjusting the belief: μ slides to a precision-weighted compromise between the sensation and the prior. Trust the senses and belief snaps to the world; hold your prior with conviction and belief barely leaves the expectation — you perceive, mostly, what you already predicted. Action lowers the very same F by adjusting the world: reach out and change things until the sensation s matches the prediction. Two descents on one landscape — update the map, or move the territory.
The panel draws both as literal bowls. The left bowl is F as a function of the belief; its floor is where perception rests. The right bowl is F as a function of the world; its floor is where action rests. Do only one and the ball reaches that bowl's floor — but a floor above zero, because the other gap remains. Do both and the whole surface drains to zero: the agent ends up believing its expectation and having made the world confirm it.
What to notice while it runs
Press Disturb the world: the true state jumps away from the expectation, the sensation follows, and free energy spikes. In Perceive only, the world stays put and the belief crawls toward it — fully if you trust the senses, barely if your conviction is high. Turn conviction up and watch the unsettling thing: the world has plainly moved, yet the agent's belief hardly does. It is perceiving its prediction and discounting the evidence — the machinery of denial, or of expertise, depending on who's right.
Switch to Act only and disturb again: now the belief is held and the world is dragged back until the senses read what was expected. The surprise is erased not by learning but by rearranging reality — the prediction made true by force. In Both, the fastest descent, the two cooperate: a little updating, a little acting, and free energy drains to near zero. Watch the pink trace fall each time. It falls whichever route you allow — that is the whole claim, made visible: the same surprise, minimized two ways.
The mapping to a human worry
The private worry here is a vertigo: how much of my life am I actually meeting, and how much am I just confirming? The theory's answer is not comforting. You walk around inside a forecast of the world and mostly experience the forecast; the raw world reaches you only as correction, and only when its surprise outweighs your conviction. Two people in the same room, holding different priors with different strengths, quite literally do not perceive the same evening.
And the action term names something we usually call by other words. The self-fulfilling prophecy, the person who is cold and so acts in ways that make others cold to them, the expectation that quietly arranges its own confirmation — all of it is action lowering free energy, the world edited until it stops contradicting the prediction. This is not a flaw bolted onto an otherwise honest perceiver; on this account it is the same drive as seeing. To hold a belief strongly is to be disposed to act the world into agreeing with it. The question the metaphor leaves you with is which of your predictions deserve that power.
Read as life lessons
Whether you bend belief or world turns on precision — how much you trust senses versus prior. It is not fixed; it's set, and mis-set precision is much of what we mean by bias, stubbornness, or naïveté.
A belief held with enough confidence stops updating and starts editing the world. Strong priors don't just resist evidence; they recruit your hands to manufacture the evidence they expect.
Low free energy — feeling unsurprised — can mean you've understood the world or merely bullied it into agreement. The number is the same. Peace of mind alone can't tell you which you did.
In the wild
Predictive coding models the cortex as trafficking in prediction errors: higher regions send guesses down, lower ones send back only the mismatch — a picture that fits much of how sensory brains are wired.
Aberrant precision is a leading story for hallucination and delusion: priors weighted too strongly overwrite the senses, so the mind perceives, and defends, what it already expects.
Active-inference agents choose actions that minimize expected surprise, unifying perception, planning and control under one objective instead of separate modules.
The mapping, exactly
| Mathematics | Life |
|---|---|
| belief μ | Your working picture of what's true right now — the guess you actually live inside. |
| expectation η (prior) | What you brought to the moment before it began — habit, hope, assumption, setpoint. |
| free energy F | Felt surprise: the ache of a world that isn't behaving as predicted. |
| precision π | How much weight you give a channel — trusting your eyes, or trusting your conviction. |
| perception (move μ) | Changing your mind to fit the world — updating the guess toward the evidence. |
| action (move the world) | Changing the world to fit your mind — arranging reality until it confirms the prediction. |
The honest model
One hidden world value, one belief, one fixed expectation η = 0. Each instant the agent senses s = world + noise and holds the free energy F = ½·π_s·(s−μ)² + ½·π_p·(μ−η)² — the Laplace/Gaussian form of variational free energy, shown up to an additive constant. It descends by gradient: perception integrates μ̇ = π_s(s−μ) − π_p(μ−η), action integrates ẇ = −π_s(s−μ), and the mode switch decides which are allowed to run.
Both wells, both balls, the pink trace and every chip are read straight from that live integration — no scripted curves. The perception ball settles at the precision-weighted mean μ* = (π_s·s + π_p·η)/(π_s+π_p); the action ball settles where the sensed world meets the belief; and only when both routes run does the joint state collapse to μ = world = η, free energy at its floor. Disturb, retune, and it recomputes from the equations.
Where the metaphor tears
Free energy can be told about perception, action, learning, even evolution — which makes it powerful and slippery at once. A principle that forbids nothing predicts nothing in particular; much of the human content lives in the parts it leaves open — where priors come from, how precisions get set — and those are exactly the parts that matter to a life.
Taken literally, the surest way to never be surprised is a dark, silent room forever. That we instead seek novelty, risk and the unknown means the story needs extra machinery — expected free energy, curiosity, priors that want exploration. The bare principle undershoots the restlessness of actually being alive.
Here the world is a single scalar with a single tidy expectation. Real minds juggle hierarchies of predictions across time, other agents with their own models, and stakes that no quadratic error captures. The panel shows the logic of the move, not the vertiginous depth of doing it about a whole life.