From synapse to silicon.
One continuous research stack.
Superintelligence & AGI
The Genesys program: charting the path to continual-learning world models that reason with test-time compute and improve without forgetting — grounded in our published results on selection, forgetting and verification, negative ones included, so capability never outruns understanding.
Neural Interfaces
The field has entered the 4,096-electrode era, and decoding — not hardware — is becoming the bottleneck. Our Cortex program tracks how foundation-model decoders generalize across sessions, subjects, and tasks, with mental privacy as a first-class constraint.
Post-Silicon Compute
Intelligence needs a new substrate. Our published cost-model campaigns map where analog in-memory compute keeps an 11.8× modeled decode advantage, where all-optical inference honestly wins — and where both collapse. SynapseOne and Lumen are the programs built on those measured boundaries.
Frontier science,
three lines of code.
A preview of the developer surface the Genesys program is building toward: one SDK for models and neural decoders alike — stream spike trains, query world models, recalibrate pipelines. Join the early-access waitlist to shape it.
import { Genesys, NeuralDecoder, SpikeTrain } from '@neurogenesys/sdk';
import { CortexArray, SafetyTier } from '@neurogenesys/types';
// ── [01] SESSION INITIALIZATION ──
const decoder = new NeuralDecoder({ channels: 4096, model: 'genesys-1' });
const spikes = new SpikeTrain({ rate: 30000, format: 'event-driven' });
console.log("Calibrating cortical array...");
await decoder.calibrate();
Built in the open.
Published for everyone.
Science that can't be reproduced isn't science. Our experiments run on fixed seeds, our claims trace to pre-registered runs where we have them, and our results — including the negative ones and our own corrections — are published for anyone to check.
NumPy
Run-once, seeded experiments
PyTorch
Real-checkpoint validation
Python
Cost-model campaigns
pytest
Automated verification suites
TypeScript
AW3 portfolio engineering
Next.js
Product & research surfaces
Frontier capability.
Responsibility committed in advance.
Neural Data Privacy
Mental privacy is a right, not a feature. Any neural data we ever touch will be collected only under informed consent, de-identified at acquisition, and never used outside the study it was gathered for.
Secured Frontier Weights
Our commitment: frontier model weights belong in hardware-isolated enclaves with strict access controls. What gets built at the frontier stays under lock until it clears the safety bar.
Contained Evaluation
Our standard: frontier models pass dangerous-capability evaluations in contained sandboxes before any external exposure. Capability ships only after oversight catches up.
How NeuroGenesys operates.
From neuron to knowledge to silicon.
Decode the Brain
The field's 4,096-electrode arrays and synapse-resolution connectomes are capturing how biological intelligence actually computes — recordings of the only general intelligence known to exist. Our field notes track every milestone, with sources.
Model the Mind
Neural data trains the field's brain foundation models, and we study which of their principles transfer into Genesys — our research path to continual-learning world models, built on published results about forgetting, selection, and verification.
Rebuild It in Silicon
What the science teaches gets etched into hardware — eventually. First we measure: our cost-model campaigns chart exactly where analog compute and optical inference win before SynapseOne or Lumen commit to silicon.
Three programs.
One frontier.
Thought to text at conversation speed.
The field's best speech neuroprostheses now decode attempted speech at 62–78 words per minute, and arrays have reached 4,096 electrodes. Our Cortex program tracks the layer that made it possible — foundation-model decoders — and what it takes for them to generalize across sessions, subjects, and arrays.
// ── THE DECODING SHIFT ──
// state of the field, sourced in our notes:
// 4,096-electrode cortical arrays (2024)
// 62–78 wpm attempted-speech decode (2023)
// pretrain-then-adapt decoders (POYO, 2023)
// open problem: cross-session generalizationThe frontier,
quantified.
Analog Decode Advantage, Cost Model
Frozen-Expert Within-Task Accuracy, Split-MNIST
Optical Precision Robustness vs Matmul, Simulated
Pre-Registered Experiments
Research Output
Open Research Publications
Progress you can verify.
Results from our research programs — published openly, pre-registered where possible, and reproducible.
Class-incremental accuracy factorizes cleanly
On Split-MNIST, accuracy equals selection accuracy times within-task accuracy to within a hundredth (±0.05 on harder synthetic data) — confirming Kim et al.'s 2022 theorem. With frozen per-task experts, catastrophic forgetting becomes a selection problem.
Real judges didn’t break Best-of-N
In a pre-registered test on 9,600 Llama 3.1 8B candidates, no reward model, LLM judge or log-probability made accuracy fall by N = 32, and measured ε did not predict the best budget — so our simulated “more sampling hurts” claim stays a simulation.
The honest optical speedup, requoted against an ASIC
A claimed 48× all-optical advantage collapses to near parity against a fused INT8 ASIC — a negative result we published in full, alongside the two advantages that survived.
"We didn't set out to build a company. We set out to answer the oldest question in science — and discovered that the answer would need new models, new medicine, and new silicon."

Will Schulz
Founder • AW3 Technology
Access the frontier.
Fund the science.
Researcher
Academic Access
- All publications, open access
- Free research tools & calculators
- Replication & falsification support
- No affiliation required
API & MCP
Metered Credits
- 100 free credits with your first key
- Research calculators & reliability analyses
- Hosted selection-reliability projects
- Prepaid credit packs, no subscription
Partners
Design Partners & Collaborations
- Eight-week design-partner program
- Hosted monitor, gate & audits at no cost
- Founder-led support & a written verdict
- Research collaborations, case by case
Clear answers.
No science fiction.
One question in several settings: how do intelligent systems learn without forgetting and choose the right answer? Our 40 published papers cover continual learning, test-time compute and evaluation, and cost-model studies of analog and photonic compute, alongside sourced field notes on brain-computer interfaces. The applied side is selection reliability — measuring how often an AI agent's router, retriever or judge picks wrong.
Yes. Our publications are fully open — preprints, defensive publications, and negative-results studies, with methods and seeds written up, and pre-registrations where they apply. If you want to collaborate, replicate a result, or challenge one, reach out through the contact page. Falsification attempts are welcome; that's the methodology.
No — and we won't imply otherwise. Our interfaces work today is research: tracking the field's 4,096-electrode era, following how foundation-model decoders generalize across sessions and subjects, and publishing sourced field notes. Any future clinical work would run under proper regulatory and ethical oversight, and we will say so publicly when it does.
Neural data is the most sensitive data there is, and we treat mental privacy as a right. We do not collect neural recordings today. Any we ever touch will be collected under informed consent, de-identified at the point of acquisition, never sold or shared with commercial third parties, and never used outside the study it was gathered for.
Capability should never ship ahead of oversight. Our published work on verifier exploitability and real-judge Best-of-N is about that gap — how far a system's own judges can be trusted to pick the right answer. Our standard for any frontier system: dangerous-capability evaluations in contained sandboxes, isolated weights, and interpretability audits of what models learn — not just what they output.