Novel Neural Architectures

HELIX models stack HarmonicStateLayer → ElasticAttentionGate → ResonantExpertRouting blocks — a recurrent/sparse-attention/MoE hybrid trained entirely on Apple Silicon, three generations deep (56M, 127M, 209M). The honest answer so far: no. Matched dense transformers have beaten HELIX on held-out perplexity every time the comparison has been run — by ~8–9% at 56M and 15.0% at 127M — and that gap has never been re-measured at the current 209M scale. HELIX-200M's base pretraining is complete (244,141 steps, held-out val PPL 27.50); post-training is active, not finished.

The SynapNet/HibbieFormer line trains with explicit plasticity phases — hebbian, synaptic scaling, and pruning on a 50/30/20 schedule with adaptive neuromodulation. The completed TinyStories run logged 21,476 steps across 215 plasticity cycles.