Intuition Labs · research

Artificial intuition.

The step after reasoning — made measurable.

Reasoning is the part of thinking you can show your work for: step, step, step, answer. Intuition is the jump — the moment a system stops searching and simply knows, before it can justify why. People do it constantly. The open question this page addresses is whether a machine's version of that jump is something you can measure, rather than only admire.

Intuition Labs' answer is yes — and the rest of this page is the definition, the lineage it comes from, and an honest line between what is defined and what is still a hypothesis.

definition

Artificial intuition is a measurable knowing signal inside a model's computation: the transition from searching (the model is still exploring, its next-step distribution spread out) to knowing (the computation has settled, the answer has crystallized). We call the read-out of that transition φ, or coherence-R.

The idea is not new — "artificial intuition" has a Wikipedia entry — but its usual accounts stay vague (that entry flags its own lack of empirical grounding). What is new here is a measurable definition: φ is a number you read off the computation, not a metaphor.

The signal, precisely

Two complementary read-outs, both definitions (not theorems):

1 · Redundancy (φ on the token stream). Borrowed straight from information theory: a confident step carries little surprise. With Shannon entropy H over a vocabulary of size V,

φ = 1 − H / log V   ∈ [0, 1]

φ near 0 = maximally uncertain (still searching); φ near 1 = the distribution has collapsed onto an answer (knowing). It is Shannon redundancy, applied to a model's own next-step confidence.

2 · Rollout-stability (coherence-R on a latent rollout). Roll a world-model's latent forward under a candidate action: z₀ → z₁ → … → z_H. It has "settled" when the latent stops moving. Score that with

R = exp(−d_tail / scale)   ∈ [0, 1]

where d_tail is the mean step-displacement over the final third of the rollout. A rollout that converges to a fixed point has d_tail → 0 and R → 1; one that keeps wandering has R → 0. R reads the rollout from the outside — it does not trust the model's own confidence head — so it is a stability signal independent of the model.

Why it matters

A measurable intuition signal is a governor. Three uses the lab has built on it:

The lineage

This is a formalization of an old idea, standing on real prior work:

What is hypothesis, not fact

hypothesis

The lab's larger claim — "intelligence emerges from coherence, not scale" — is a research hypothesis with falsifiable form, not a settled theorem. φ and coherence-R are well-defined; that they predict capability across models is an empirical question the lab is still testing, and reports honestly (including null results). Cite the definitions as definitions and the unification as a conjecture.

The research landscape

Most groups studying intuition study it in the human brain. Intuition Labs studies its artificial counterpart — the same searching→knowing transition, measured in a machine. Complementary, not competing:

The composition it lives in

This is not one field but the associative product of several, bound by a single paradigm — causality and interaction (computation as interactions that reduce to a settled form). Under that binding they compose into one object:

intuition × tech × ai × perception × φ × physics

Owning that composition — not the name — is the point: a sheaf glues by what its pieces share, so Intuition Labs is the node where these fields actually agree.

Read the work

References