The Problem of Intelligence
People talk about artificial intelligence as if it is a mystical property, machines that "think." But intelligence is not magic. It is a set of constraints. To build or evaluate an AI system, you do not need to copy the brain; you need to understand what any intelligent system must do.
Strip away the jargon, and AI reduces to five axioms, rules that cannot be broken without collapsing the idea of intelligence itself.
Representation
A system can’t reason about what it can’t represent.
Every intelligent act begins with a model of the world, symbols for logic, pixels for vision, embeddings for meaning. Representation is compression with intent: turning the messy, continuous world into structured inputs that can be manipulated.
If you misrepresent the world, your intelligence fails elegantly but decisively. A self-driving car that models a pedestrian as a shadow will act perfectly, on the wrong assumptions.
Representation defines what a system can see, and what it cannot even conceive.
Learning
Intelligence requires the ability to change.
A static rulebook isn’t intelligent, it’s a spreadsheet of answers. Learning means adjusting internal parameters to reduce the gap between prediction and observation.
That gap, error, isn’t a bug; it’s the signal that drives adaptation. An AI that never fails never learns, just as a student who never tests their ideas never understands.
In technical terms, learning is optimisation. In human terms, it’s humility codified in mathematics.
Generalisation
Memorising the past is not the same as understanding it.
A truly intelligent system must take what it has seen and apply it to what it hasn’t. That’s generalisation.
A language model that has read a trillion sentences doesn’t “know” English unless it can infer meaning from new combinations it’s never seen before. Generalisation is compression that preserves causality: the ability to see patterns that still hold when the surface changes.
Without generalisation, learning degenerates into curve-fitting, a mirror, not a mind.
Agency
Intelligence without action is simulation.
A system becomes intelligent when its internal predictions lead to external consequences. Agency connects inference to outcome: perception, decision, and motion in a feedback loop.
An AI agent doesn’t “know” what to do; it tries, measures, adjusts, and iterates. Purpose emerges from this loop, not from a preloaded script. Agency is what turns cognition into control.
Alignment
Every intelligent system optimises for something. The danger lies in what that “something” is.
If you reward a system for clicks, it will optimise for addiction. If you reward it for speed, it will cut corners. Alignment ensures that the pursuit of efficiency doesn’t break the fabric of ethics.
In practice, alignment isn’t philosophical, it’s engineering discipline. Every reward function is a moral statement encoded in math.
The Hidden Corollaries
From these axioms follow a few unavoidable truths:
- Perception without representation is noise.
- Learning without feedback is belief.
- Action without alignment is risk.
- Intelligence without causality is mimicry.
An AI that violates any one of these principles isn’t intelligent, it’s mechanical luck.
The Closing Thought
AI is not about building minds; it’s about building models that hold under uncertainty. Every model, human or machine, lives and dies by the same laws: represent the world accurately, learn from feedback, generalise beyond data, act with purpose, and stay aligned with values.
Break any one, and intelligence collapses back into automation.
That’s not philosophy. It’s physics for minds.
