The Vexafracta design is not an attempt to make a conventional neural network marginally more efficient. It begins from a more severe question:
What computation can be removed before the system stops being useful?
The answer may be: far more than modern software habits suggest.
At the bare-metal level, every bit has a cost. Every unnecessary state field occupies memory and consumes bandwidth. Every broad memory lookup weakens locality. Every dense calculation repeated across a quiet field spends energy describing nothing.
Vexafracta therefore treats efficiency not as a later optimization pass, but as a design constraint from the beginning. If a bit is not useful and meaningful, it should not be present.
This is the experimetal part of the project: experimental in its willingness to test unfamiliar computational structures, and metallic in its preference for things that are forged, discrete, load-bearing, and resistant to unnecessary ornament.
The existing Vexafracta model describes a 16×16 substrate of VN units surrounding a central 2×2 core. An action begins at that core. The resulting state propagates exward through successive VN neighborhoods until it reaches the boundary of the VEXEL. There the direction reverses. The transformed state then propagates inward through the same substrate, eventually returning to the core and producing the VEXEL’s new output state.
The outward and inward journeys are therefore not two unrelated processes. They are two directions through the same computational material.
This suggests another useful way to think about the VN unit itself. A VN should not be merely a disposable pixel inside a larger VEXEL. It may instead be understood as a minimal micro-VEXEL: not a geometrically identical miniature, but a computational object governed by the same discipline at a smaller scale.
A VN receives a constrained local condition. It interprets that condition according to its local geometry, direction, present state, and retained history. It applies a compact rule and emits a correspondingly compact state to the next local neighborhood.
At the larger scale, the VEXEL performs an analogous operation. It accepts an input at its core, transforms that input through local interactions across its substrate, reverses direction at its boundary, integrates the returning result, and produces a new state at the core.
Neither scale requires global knowledge. Each object needs only enough local information to obey its rule.
That distinction matters. The goal is not to construct a detailed numerical portrait of the entire system at every instant. The goal is to allow useful global behavior to emerge from repeated, inexpensive local operations.
The 256-entry sigmoid lookup table is a useful example. It is not merely a shortcut for avoiding repeated mathematical evaluation. It is a declaration of intent. A structured local input becomes an 8-bit address. The lookup table filters that condition and returns a two-bit activation state:
00— inactive01— approaching inactive10— approaching active11— active
Four states. Two bits.
The lookup is therefore both a nonlinearity and a compression boundary. Whatever complexity produced the local input is reduced immediately to the amount of information the VN actually needs to carry forward.
The same economy applies to inhibition.
Inhibition does not necessarily require signed arithmetic or the explicit storage of a negative activation value. A local condition can simply fail to reinforce the propagating state. Direction, phase, retained history, threshold relationships, and other local rules may eventually provide richer forms of inhibitory behavior without requiring the architecture to represent inhibition as a conventional negative number.
This is not an argument against complexity.
It is an argument about where complexity belongs.
Vexafracta does not attempt to place complexity inside every individual operation. Instead, complexity should emerge from the composition of simple, local, legible acts repeated across a structured field.
That is also why the VN-as-micro-VEXEL idea is interesting. The VN and the VEXEL need not be physically identical or execute identical machinery. What may be recursive is the computational ethic:
Receive locally. Transform locally. Retain only what matters. Emit only what matters.
At one scale, that operation occurs across neighboring VN units. At another, it occurs across neighboring VEXELs. The larger structure should not suddenly require an entirely different computational philosophy merely because the scale has changed.
If the VEXEL is to become a useful computational object, its VN substrate must first demonstrate that real work can emerge from very little machinery: compact state, local geometry, minimal history, inexpensive filtering, directional propagation, and disciplined exchange of information.
The experiment is to discover whether those constraints produce only austerity—or whether austerity itself creates room for a different kind of expressive, resonant computational field.
As below, so above.
Not as mysticism, but as engineering discipline: the larger structure should not require a different kind of truth than the smaller one.