NeuroGenesysFounded to understand.
Built to restore.
NeuroGenesys is a frontier technology research lab in San Francisco. We study the brain, work toward aligned intelligence, and measure the silicon both will run on. This is our story and the research that shapes it.
The brain is the blueprint.
We believe the mind is knowable — and buildable. NeuroGenesys was founded by Will Schulz of AW3 Technology and a collective of neuroscientists, ML researchers, and silicon architects.
Our mission is one continuous loop: decode the brain with high-bandwidth interfaces and connectomics, model what we learn in the Genesys family of frontier world models, and rebuild those principles in neuromorphic, photonic, and custom silicon. Each pass is designed to teach us something about minds and make the next pass faster — safely, openly, and with humans in the loop.
The Impact
Pre-registered experiments
Run-once, seeded experiments across the analog compute co-design campaign, rounds 1–13.
Publications in the open
Preprints, defensive publications, and negative-results studies — each reproducible from fixed seeds.
Measured precision robustness
Random-feature optical inference tolerates roughly 3000× the precision degradation that collapses a conventional matmul.
The research,
in its own words.
"It is the attention window policy — sliding versus full — not the attention fraction, that decides whether the hybrid keeps analog compute's advantage."
AIMC Window Co-Design
Defensive Publication • Jul 2026
"Train one small expert per task and freeze it; all of class-incremental difficulty is then which expert to trust — a cross-task selection problem."
Forgetting as Selection
Preprint • Jun 2026
"Discrimination buys ceiling height but not immunity — you cannot out-discriminate an exploitable verifier."
When More Sampling Hurts
Preprint • Jun 2026
"Across three domains, the complex/wave degree of freedom is load-bearing nowhere at equal budget. Phase was the wrong implementation of a right idea."
Complex Amplitudes Are Decorative
Negative Results • Jun 2026
"Against a fused INT8 ASIC the headline 48× collapses to 1.23×. The defensible claim shrinks — so that is the claim we publish."
Optical Kernel Inference
Falsification Study • Jun 2026
"The environment never changes the total coefficient weight; it only transfers influence between the present state and the inherited state."
Spectral Response of Recurrences
Research Note • Jul 2026
