It's not a bug, it's a feature.
Overflow is also computation.

Exploring a hypothetical solution by using bit overflow logic as an advantage.

This project is an open-ended experiment in rethinking how neural computation could work beyond traditional 2D matrix constraints.

Instead of linear tensors and rectangular attention windows, I'm exploring circular structures — where memory wraps around, overflow is meaningful, and the "shape" of computation becomes dynamic.

Imagine:

• A 2D Matrix of for example 32 x 32, which are 32 node columns (nodes) and 32 node rows (lets call them layers)
• The 2D Matrix curving into a cylinder
• the cylinder curving into a ring shape (Toroid)
• A "fixed-width" toroidal memory pipe forming
• Each node only storing a tiny state — maybe 16 or 128 bits, depending on what it will perform
• Where the 2D Matrix would overflow, the toroid flips signals instead of breaking them
• Neighboring logic becomes the basis for attention, not full-matrix dot products. Therefore each node takes into account its neighboring nodes ([n-1]) and (layer[n+1])
• On each iteration, where the matrix would reach its bottom 32th row, it would continue with row 33
• As there is no layer 33, it continues with Layer 1 (or index[0])

I'm learning and building as I go. This isn't production code — it's a blueprint for a different way to think about AI.

PRs welcome. Thought experiments encouraged. Commercial use? Nope. See license.