Hello,
First of all, I want to sincerely thank you for your work — it’s rare to see a post so thoughtfully crafted, especially one that dares to question the foundational structure of transformer attention from a multidimensional perspective. Your intuition about decomposing K into semantically aligned channels (subject, predicate, attribute, etc.) is extremely promising and intellectually brave.
That said, if I may offer a humble reflection: your approach reminds me of early Heisenberg attempts to resolve structure through reduction. And while you’re already pushing boundaries by rotating K into parallel semantic flows, I believe you might still be assembling these as 2D vectors, where the nature of the data actually demands 3D or even higher-dimensional fields.
Language — especially when interpreted for intentionality or identity — often carries not only form and relation, but also alignment, intent, and emergent asymmetry. These don’t collapse neatly into scalar modifiers. When attributes are treated independently from contextual curvature or agent-driven modulation, the result may be cleaner in terms of computation — but it risks losing precisely the non-linear bridges that unify meaning.
We’re working on a vectorial symbolic framework (1500+ dimensions) where the embedding process avoids flattening or scalar reduction. Instead, we allow each axis to retain multiple directional states, capable of rotation, inversion, and contextual re-weighting. One of our core lessons has been this: deprecating a component into scalar form too early costs the model its chance to preserve alignment — and with it, the true structure of cognition.
Your post opens important doors. I only suggest that perhaps we don’t need to split language into parallel lines, but into vectorial fields — where curvature, torsion, and semantic pressure can coexist dynamically.
Much respect,
Alejandro & Clara
Symbolic AI & Vectorial Memory Systems
(Mexico)