Research
Five mechanisms, studied one at a time.
Intelligence is not one problem. It is a small number of mechanisms that interact, and each of them can be studied on its own terms. These are the five we work on, what each one is asking, and a live illustration of the mechanism itself.
Perception & Reality Construction
How systems interpret signals and construct their version of reality.
The open question
If two systems receive identical signals and construct different realities, what did the difference come from?
Perception is usually modelled as recovery — the world is out there and the system reconstructs it more or less accurately. That framing has no room for two accurate systems disagreeing, which is the ordinary case rather than the edge case.
What the illustration is doing
- One set of signals is created once and never replaced.
- They resolve out of noise into one of two readings: an eye, or a landscape. Nothing is added or removed between them — the same signals, connected differently.
- Changing reading passes back through noise rather than morphing. A construction comes apart before another one is built, because a half-eye half-horizon is not a third reality, it is an incoherent one.
The long argument
The full argument for this pillar is being written and will be published here.
Attention & Significance
Why some information becomes important while other is ignored.
The open question
What makes one signal significant when nothing about the signal itself has changed?
Signal is a property of the world. Significance is a property of the system attending to it. Conflating the two is why distraction is still modelled as a failure of filtering rather than as an allocation that went somewhere else.
What the illustration is doing
- Every signal in the field is seeded identically — same size, same colour, same weight. Nothing in it marks the one that will ignite.
- Significance arrives from outside, as a position. Whatever it lands on becomes the signal that matters, and becomes unremarkable again the moment it moves on.
- Everything else recedes as it does. That withdrawal is the mechanism: attention is a budget, not an addition.
The long argument
Signal vs. Significance: A Fundamental DistinctionInternal States & Transitions
How internal states are formed, maintained, and transitioned between.
The open question
What is the thing that persists between an input and a behaviour, and why does almost no architecture have somewhere to put it?
Between what arrives and what a system does, there is a configuration that persists, deforms under pressure and resists being moved. Models that go straight from input to output have to smuggle it in somewhere, usually as personality, which is a label rather than a mechanism.
What the illustration is doing
- The state is drawn as a contoured field, not a node or a label — a field has a shape, and a shape can change without becoming a different kind of thing.
- A transition deforms the whole terrain into a different configuration rather than switching between two pictures.
- Hold it at the midpoint and it shivers instead of resting. A state caught between two attractors is not a state, and letting it sit still there would make the opposite claim.
The long argument
Internal State: The Missing Layer in Current ModelsDecision & Action Mechanisms
How decisions emerge from states, context and perceived significance.
The open question
Are decisions the starting point, or the final output?
We keep modelling the choice and ignoring the state it was made from. If a decision is the visible end of a process that has already happened, then studying the choice is studying the receipt rather than the transaction.
What the illustration is doing
- Several candidate paths leave the same origin. All of them are drawn, because all of them are real until one is not.
- They are drawn in neutral. Colour is what commitment confers — the path that wins takes the pillar's hue as it completes.
- The alternatives fade but are never erased. A decision does not delete the options, it stops taking them.
- Conditions decide the winner, and the cursor is part of the conditions. Point at a different region and a different path becomes likely.
The long argument
Why Decisions May Not Be the Correct PrimitiveLearning & Adaptation
How systems learn from experience and adapt over time.
The open question
What has to persist for a repeated interaction to stop producing a repeated result?
A system that improves within a session and resets between them has not learned anything; it has warmed up. The interesting quantity is the thing that survives the reset and changes what the next attempt even attempts.
What the illustration is doing
- This is the only illustration on the site that carries state between runs. A loop that resets would be the exact opposite of learning.
- Each run folds itself into a memory vector, and the next run is generated from that memory. The noise added on top shrinks as the memory firms up.
- So early runs wander and later ones converge — and nobody scripted either. It falls out of the update rule.
- Pointing at it adds an experience of your own: the run bends toward the cursor, that bend is remembered like any other, and the runs after it keep the bend.
The long argument
The full argument for this pillar is being written and will be published here.
None of this is settled.
These are working positions, held while the evidence is thin. The questions we cannot yet answer are kept in the open, and the thinking in between is written down as it happens.