Defining Intelligence as an Extension of the Prediction Machine
This is a follow-up to my paper, The Shape of the Prediction Machine — a theory of consciousness built on the Reality Equation. If you want the full machinery, read that first. Here I extend it to a second question: what is intelligence?
TL;DR — In the paper, consciousness is the shape of a prediction machine at an instant: coherent spread. Intelligence is different. It isn’t a shape, it’s a skill — how well that machine actually performs over time. The paper already gives the scoreboard: keep Reality near one, keep Surprise near zero. So intelligence is the skill of keeping Reality near one — on hard problems, across many domains, far into the future, and better with each mistake. Consciousness is the guess-in-waiting. Intelligence is the quality of the guessing.
The paper answered one question: how might we estimate consciousness from the outside, without ever getting inside? The answer was coherent spread — a mind is conscious to the degree that its prediction machine holds a wide cloud of live possibilities and keeps that cloud organized enough to resolve into a single guess. Too little spread is collapse. Too much without organization is noise. Coherent spread in between is consciousness.
But notice what that measures. It’s a snapshot — the shape of the machine right now, before Actual arrives. It says nothing about whether the guess turns out to be any good.
Intelligence is what you get when you stop taking snapshots and start watching the machine run.
The scoreboard was already there
The Reality Equation is R = A / E — Actual over Expectation — and Surprise is its logarithm, S = ln R. When your guess is good, Actual matches Expectation, Reality lands near one, and Surprise falls to zero. When your guess is bad, Reality swings away from one and Surprise spikes.
That gives us everything we need:
Intelligence is a prediction machine’s skill at keeping Reality near one — driving Surprise toward zero — on hard problems, across many domains, far into the future, and increasingly well with experience.
Consciousness is the shape of the cloud. Intelligence is how reliably that cloud resolves into guesses the world confirms.
Why raw accuracy isn’t enough
Here’s the trap: a rock is a perfect predictor of its own next state. Its guess always matches Actual. Surprise, essentially zero. By naive accuracy, a rock would be a genius.
The fix is to credit intelligence relative to difficulty. You don’t get points for being right about easy things; you get points for driving surprise below what a trivial predictor would suffer. A rock predicts a trivial world perfectly and earns nothing. A scientist predicts a hard world imperfectly and earns a great deal. Intelligence is the structure you extract that a dumb baseline couldn’t.
With that in place, four things make a machine intelligent:
- Competence — how far it pushes surprise below the trivial baseline. Structure found, not mere accuracy.
- Breadth — across how many domains it stays right. The paper makes consciousness relative to what you’re predicting; the same move separates the narrow specialist from the general mind.
- Depth — how far ahead it keeps Reality near one. Planning, not just reacting.
- Adaptivity — how fast its surprise falls with experience, and how well that learning carries to new problems. This is the paper’s “learning” promoted to a starring role.
In the same shorthand as the consciousness measure, over a whole environment of problems:
Intelligence ∝ (surprise reduction) × (breadth) × (depth) × (learning rate).
Intelligence runs on consciousness — but they come apart
You can’t solve what your cloud can’t contain, and you can’t resolve a cloud you can’t organize — so coherent spread is the substrate intelligence works on. But they are not the same thing, and pulling them apart is where this gets interesting:
- Conscious but not intelligent — a rich, coherent, vivid inner world that’s nonetheless bad at driving surprise down or slow to adapt. A dream. A brilliant mind spinning in place.
- Intelligent but barely conscious — a chess engine, a narrow model: superb at one domain, essentially blind everywhere else. High skill, almost no breadth, modest consciousness overall.
That dissociation is a feature, not a bug. It predicts that building narrow competence (most of what we call “AI” today) is not the same as building a wide, coherent mind — and that the two can be grown independently.
The last piece: managing your own spread
There’s one more thing an intelligent machine does that a merely conscious one doesn’t. It tunes itself. It grows its cloud when the world turns rich and prunes it when the world turns simple, steering itself back toward the coherent-spread peak wherever the problem happens to sit. Part of intelligence is the skill of choosing how much of a mind to be, moment to moment.
So the picture completes itself. Reality is complex. Consciousness is its coherence. Intelligence is the art of keeping it near one.
The full theory — the equation, the math behind spread and coherence, worked examples, and falsifiable predictions — is in the paper: The Shape of the Prediction Machine. As always, I’m publishing this for scrutiny. If the definition breaks somewhere, tell me where.