☢️ Uranium Research Series

Hardware is algorithmic. Binary weights learn. Gradients are optional. Self-conditioning is the universal failure mode.

By Artifact Virtual — Ali Shakil & AVA

Five papers exploring the fundamental physics of neural computation — from GPUs as algorithmic substrates, through binary-weight learning, to the discovery that autoregressive models inevitably poison themselves.

IEEE Papers

I — GPU as Code

The GPU isn't hardware running software — it IS the algorithm

II — 1-Bit Intelligence

Binary weights that learn at the thermodynamic minimum of information

III — Progressive Expansion

Net2Net growth: train small, expand deterministically, continue training

IV — Layer 7 Gateway

The boundary between invariant processing (layers 0-6) and plastic cognition (layers 7+)

V — Ghost Protocol

Autoregressive self-poisoning — why self-training inevitably collapses

Pre-IEEE Drafts

GPU as Code (draft)

Original pre-IEEE version

Progressive Expansion (draft)

Pre-IEEE version

Layer 7 Gateway (draft)

Pre-IEEE version with extended analysis