NeuroGenesysThe frontier,
measured.
Research claims are cheap; measurements are not. These are the numbers our programs stand behind — measured advantages, exact factorizations, honest negative results — published and reproducible.
Numbers that matter.
Measured Analog Compute Advantage
Analog in-memory compute holds a context-independent 11.8× decode advantage for static-weight sequence models — measured in our cost-model campaign, with the exact boundary published.
Frozen-Expert Accuracy, Split-MNIST
Freeze one small expert per task and each is near-perfect. All of class-incremental difficulty is selection — which is why forgetting reduces to a verifier-exploitability problem.
Optical Precision Robustness
Random-feature optical inference tolerates roughly 3000× the precision degradation that collapses a conventional matmul — one of two genuine optical advantages our falsification study confirmed.
Pre-Registered Experiments
Run-once, seeded, kill-criterion-first: the falsification campaign behind our analog compute design rule — every headline claim traceable to a numbered experiment.
One lab.
Many frontiers.
Open Research Publications
Where The Work Goes
- Models: forgetting & verification2 papers
- Silicon: analog & optical compute2 papers
- Foundations: cognition & mathematics2 papers
*Every program shares the same loop: neuroscience informs the models, and what the models need shapes the silicon.
2026 Milestones
- Wave-cognition falsification publishedJun
- Verifier exploitability preprintJun
- Spectral response theorem confirmedJul
- Analog compute design rule publishedJul
Bandwidth measured
in words
A decade ago, brain-computer interfaces moved cursors. The field's best speech neuroprostheses now decode attempted speech at conversational rates — progress we track, with sources, in our field notes.
78 wpm
Best Published Decode (2023)
4,096
Electrode Field Record (2024)
Air-gapped
frontier evals.
Our published standard: any frontier capability increase runs its dangerous-capability evaluations inside physically isolated sandboxes — no network path out until oversight signs off.
Built by every
discipline
- NeuroscienceField notes
- Machine LearningVerification
- Silicon EngineeringCost models
- Ethics & PolicyNeurorights
Neuromorphic
edge intelligence
The SynapseOne program: always-on, event-driven inference for implants, wearables, and robots — built on our measured analog compute design rules.
One SDK.
Every frontier.
The developer surface the Genesys program is building toward: models, decoders, and silicon targets behind one SDK — from notebook to implant. In design now.
→ package reserved — early access soon
