Block A
Byte Unit
The first neuron layer. Reads a raw byte (0–255), splits it into individual bits, and learns a compact binary fingerprint. Frozen champion: binary + C19 activation, H=16.
FrozenA gradient-free substrate that learns to wire itself — no backprop, no tokenizer pretrain, no GPU cluster required.
VRAXION is a research prototype for building AI systems that grow their own internal wiring from scratch — without gradient descent, without pre-trained weights, without a tokenizer.
It processes raw bytes, learns compact binary representations, and assembles higher-level understanding one building block at a time.
The goal: a substrate that is fully transparent, reproducible, and runnable on a laptop.
Each block is a standalone layer that can be understood independently. Click any card to dive in.
The first neuron layer. Reads a raw byte (0–255), splits it into individual bits, and learns a compact binary fingerprint. Frozen champion: binary + C19 activation, H=16.
FrozenTakes pairs of Byte Unit outputs and merges them into a single combined representation. Learns which byte-pairs belong together without supervision.
+ Native 7-bit variant → ActiveEmerges from the Merger: byte-groups that consistently fire together become de-facto tokens. No BPE, no vocabulary file — the tokens grow from data.
PlannedMaps token-level representations to a dense embedding space. Quantized to int8 — the LUT bakes directly from byte-unit weights, no separate training step.
BetaThe reasoning substrate. A growing network that wires its own graph at runtime — connections appear where the signal is useful, and fade where it is not.
PlannedReproducible, Apache-licensed, runs on CPU. File a finding or an honest critique — both welcome.