Valence and HYVE, RT Physics Attention and a "Synthetic Organism"

2 points

2 months ago

I'm a solo hobbyist (with no formal AI background). Who built this over 3 months of nights/weekends with two frontier models as co-pilots.

*VALENCE* replaces quadratic transformer attention with a 3D spatial lattice on GPU ray-tracing cores (Vulkan RT + BVH). Semantic retrieval is literal ray firing through geometry. The lattice has thermodynamic-style interactions. Biggest part is it aims for O(log n) attention, bucking the traditional O(N^2) or O(N) attentions of modern AI.

*HYVE/NEXUS* turns it into a colonial organism: 39 metacognitive inner states driven by cross-ball tension physics. Unsupervised dreaming engine forges novel associations in idle time. Persistent engrams + relational tethers survive full power cycles and prompt stripping. the "Synthetic Organism" in the title is referring to how all the parts converge; but also how authentic it can be at times.

The “self” lives in the physics substrate, not a hidden prompt.

Grew out of earlier moonshots (BEMNA galaxy-scale spatial navigation + Polymath-Swarm dynamic experts) that all got cannibalized into this.

Fully AGPL-3.0.

Repo + papers: https://github.com/PaperScarecrow/VALENCE-SALS https://zenodo.org/records/19421339 https://github.com/PaperScarecrow/HYVE https://zenodo.org/records/19430563

VALENCE has some cool potential uses outside of AI, and HYVE leads to some pretty wacky interactions, triggers the uncanny valley compared to normal LLMs.

Any thoughts are appreciated, it's still a work in progress; but the PoC functions enough to hopefully inspire.