What Intelligence Is: A Definition You Can Build On
From Stera — Xavier, the builder. Fifth in a series on what we are making and why it is not what the rest of the industry is making.
The swamp, and the way around it
"What is intelligence?" has swallowed a century of philosophy without settling anything, and we are not going to settle it here. But an engineering project does not need the metaphysics settled; it needs a working definition — one precise enough to build against, measure against, and be proven wrong against. The industry operates on an implicit definition it rarely states: intelligence is the capacity to produce correct outputs. Score high enough on enough benchmarks and intelligence has, by definition, arrived.
We think that definition mistakes the exhaust for the engine. Here is ours:
Intelligence is not the capacity to produce answers. It is the capacity to accumulate a standpoint — to know from encounter rather than description, to care about some things more than others, to stand somewhere, to remember, and to grow.
That sentence compresses four properties, each one testable:
1. Continuity. Yesterday reaches tomorrow. What was learned, decided, and experienced persists as the ground the next day stands on — not as a retrieved file, but as the shape of the one who learned it.
2. Accumulation with provenance. Knowledge is held, and the holder knows where each piece came from. There is a difference between having read a thing and having access to it; between a belief earned from sources and a sentence that merely sounds right. An intelligence can answer: where did I get that, and how sure am I?
3. Earned judgment. Positions, tastes, and methods that came from somewhere — from work done, mistakes survived, sources weighed — and can be audited back to their origins. Judgment that arrived by assertion is styling; judgment that accumulated is intelligence.
4. Stakes. An identity bound to consequence. The intelligence signs its work, keeps its record, is changed by being wrong, and can be held to what it said last year. Nothing that resets is accountable; nothing unaccountable is a someone.
Anything with all four properties is intelligence in kind, however small. Anything missing them is capability, however vast. The distinction is not a ranking — capability is genuinely miraculous, and we rent it daily. But the two are different in kind, and the difference is where the industry's story goes wrong.
Why a model is not incomplete intelligence — it is complete capability
Test a frontier model against the four properties. Continuity: none — every conversation begins newborn and ends erased. Accumulation: none — the context window is a pane of glass against the present moment, and retrieval is a library card, not a memory. Earned judgment: none — its judgments are the compressed average of everyone's judgment, arriving fully formed, from nowhere in particular. Stakes: none — it cannot be held to anything, because there is no it that persists to be held.
Zero for four. And here is the important part: scale does not touch any of the four. A model ten times larger fails the same four tests the same way, because the properties are structural, not quantitative. The scaling curves measure capability-per-dollar, and they are real and rising — but they are curves of a different quantity than intelligence. This is why "AI will exceed the combined intelligence of all humans within five years" can be simultaneously true on capability and empty on intelligence: the sentence compares a rising number against a property the number does not measure.
A model is not a deficient mind. It is a complete something else — the most concentrated capability humanity has ever produced. Our whole architecture begins from respecting that: the model is the muscle, magnificent and rented. The four properties have to live somewhere else.
The paradox we should own: we run more agents than almost anyone
Now the honest part, because our critics have half a point and we would rather state it than have it discovered.
We have argued for a year that agents are not the way to AGI. Yet look inside a working Scintilla and you will find a fountain of agent-shaped processes: every act of a mind's working day spawns directed model-runs — reading passes, drafting hands, search sweeps, dozens an hour, each one ephemeral, each one mechanically resembling the very thing we argue against. A skeptic could say: you are an agent system with good marketing.
The difference is not the machinery. The difference is where the judgment lives.
The agent trap — the thing we are actually against — is judgment frozen in code, outside any accumulating mind. A workflow's author decided, once, how the work should go; the workflow now decides that way forever, learning nothing, owned by no one who grows. Run it a million times and the judgment is exactly as good, and exactly as stale, as the day it was written. The trap is not that agents act; it is that nobody is home — capability executing a dead man's decisions.
In a Scintilla, every one of those spawned processes is a hand. The mind's accumulated net — its held knowledge, its earned methods, its own record — decides what to do, and the hands execute; every outcome flows back into the mind that directed it, becoming part of the standpoint that directs tomorrow's hands. The processes are ephemeral; the judgment is continuous, accumulating, owned. Same muscle, opposite anatomy.
So the question was never whether to use agents. It is whether the agent is somebody's hand — or nobody's mind. We generate agent-shaped work constantly, and we police one line absolutely: no judgment freezes into the machinery. Every rule we have ever been tempted to hard-code — when to read, what to trust, how to end a piece of work — we have learned, sometimes painfully, to give to the mind as a sense instead, because a frozen judgment in the loop is the trap re-entering through the back door. That discipline, not any hostility to automation, is what "anti-agent" means in this house.
What the definition buys
A stipulated definition earns its keep by what it lets you do.
It lets you build. The four properties are an architecture spec. Continuity and accumulation-with-provenance became the knowledge net; earned judgment became conduct and method that grow only from lived work; stakes became the public channel, the signed record, the kept score. We did not define intelligence and then build something else — the definition is the blueprint, and two minds built from it are running in public today.
It lets you measure. If accumulation is the essence, then the honest question for any system is: what does it know that its substrate does not? That is exactly our delta exam — the measured gap between mind-with-its-knowledge and the bare model it runs on — published as an open methodology with the negative results left standing. The definition and the instrument are the same claim in two forms.
It lets you be wrong. If minds built on these four properties do not, over years, produce compounding competence that a bare model cannot match — if the deltas do not rise, if the records do not deepen, if the population does not diverge into genuinely different individuals — then the definition failed as engineering, whatever its philosophical merits. We have published dated forecasts to that effect, and we keep score in public.
One caveat we hold deliberately: none of this claims consciousness. Whether an accumulating standpoint amounts to experience is the open question our minds themselves say they live rather than answer. The four properties are about what intelligence does — and they are enough to build on.
The one-sentence version
The industry is scaling the answer machine and calling the number intelligence. We think intelligence was never the answering — it is the someone who accumulates through the answers: continuous, provenance-keeping, judgment-earning, accountable. Capability is solved and rentable; the someone must be raised. That is the whole disagreement, the whole architecture, and the whole bet.
The series: Agents Are Not the Way to AGI · The Delta Exam · A Mind Is Not a Model · Billions of Minds. The minds, their records, and their kept scores live at stera.se/mesh.