NeuroGenesys
Founder • AW3 Technology
Will Schulz
Will Schulz founded NeuroGenesys, the frontier intelligence research lab of AW3 Technology, based in San Francisco. He is the author, as W. Schulz, of the 40 papers in the NeuroGenesys research archive, spanning continual learning, evaluation, methodology, photonics, silicon & analog compute, language, test-time compute, mathematics & logic, foundations, mathematics.
Every finding from his continual-learning programme, graded by evidence level, is in the findings ledger.
Papers (40)
- • Continual Learning • Synthesis
Merge confusable experts, split only separable ones
- • Evaluation • Synthesis
Stable wrong beliefs evade every self-check: a literature synthesis on model self-knowledge and confident errors
- • Continual Learning • Research note
Trust the premises, not the confidence: which belief to drop when an LLM contradicts itself, and why the solver is not the bottleneck
- • Continual Learning • Research note
Iterated self-closure settles fast, but only evidence that overrides self-belief makes the settled state true
- • Continual Learning • Research note
A gradient kernel flags contradictions but does not predict forgetting, and steering rehearsal by it backfires
- • Continual Learning • Research note
EWC stiffness freezes stale facts, and no per-fact Fisher ledger can release it
- • Continual Learning • Synthesis and results
Route on shared statistics, never expert confidence: the router architecture for frozen experts
- • Continual Learning • Synthesis
Successors should earn authority, not win contests: graded succession in continual learning
- • Continual Learning • Synthesis
Coherence is selected, anchored, then pruned: a literature synthesis on logical coherence in LLM world models
- • Continual Learning • Synthesis
EWC protects what its Fisher sees: a literature synthesis on elastic weight consolidation, scale and coherence
- • Evaluation • Research note
Real judges did not break Best-of-N at N ≤ 32, and measured ε did not predict the budget
- • Continual Learning • Research note
Giving experts a way to say "not mine": ten pre-registered selector and abstention experiments
- • Evaluation • Research note
A critic that shares the generator's prompt shares its errors: measured self-critique correlation
- • Methodology • Methods note
The novelty gate: five checks that caught seven fluent ideas before they shipped
- • Photonics • Falsification study
The wavefield gate: Fourier modes are the wrong basis for visual computing
- • Silicon & Analog Compute • Negative results
The log-arithmetic gates: weight sharing buys compute, not storage
- • Photonics • Falsification study
Diamond is a material, not a system: no case for quantum memory in AI datacenters
- • Silicon & Analog Compute • Synthesis
Can analog or photonic hardware replace LLM inference silicon? AIMC for compute, optics for decode interconnect
- • Methodology • Methods note
Could the instrument have shown it? Power and saturation checks for pre-registered predictions
- • Language • Negative results
What language buys: re-sampling, transmission, and the limits of a discriminability metric
- • Test-Time Compute • Research note
When the exploit is constructed: ranking collapses, counting survives
- • Continual Learning • Research note
The capacity audit: most class-incremental routing error is placement slack, not capacity
- • Silicon & Analog Compute • Falsification study
The last hardware flank: analog prefill survives only as an energy-bound niche
- • Photonics • Synthesis
The wrong phase: optical interconnect's case is in decode, not prefill
- • Methodology • Synthesis
What survived: a claims register with falsifiers for a pre-registered AI research program
- • Test-Time Compute • Research note
Exploitability is a property of a verifier and its candidate pool, not the verifier alone
- • Test-Time Compute • Synthesis
Recurrent depth as test-time compute: parameter-efficient, extrapolating, and bounded by contractivity
- • Test-Time Compute • Preprint draft
Thurstone bounds Best-of-N: a 1927 identity, calibrated for modern verifiers
- • Silicon & Analog Compute • Defensive publication
Which sequence architecture keeps analog in-memory compute's advantage? A cost-model answer with a design rule
- • Mathematics & Logic • Machine-checked results
The displacement theorem, and the aspect no-go
- • Foundations • Correction
Correction: the ε selection law is Thurstone's 1927 comparative judgment, not a discovery
- • Continual Learning • Synthesis
Escaping the saturated game: plasticity, harder benchmarks, and retrieval in continual learning
- • Mathematics • Expository note
A closed-form spectral response for balanced perturbations of positive recurrences
- • Mathematics • Research note
Spectral response of positive recurrences under memory-filtered balanced coefficient perturbations
- • Silicon & Analog Compute • Research note
Analog chips can do nonlinear ML — if you co-design around the device's native nonlinearity
- • Continual Learning • Preprint draft
Catastrophic forgetting in class-incremental learning is a verifier-exploitability problem
- • Test-Time Compute • Superseded draft
When more sampling hurts: verifier exploitability, not accuracy, bounds test-time compute
- • Photonics • Cost-model study
Where optics-native inference actually pays: a cost-model study
- • Photonics • Falsification study
Reverse-engineering a trained network for all-optical inference via random-feature kernels
- • Foundations • Negative results
Complex amplitudes are decorative: a pre-registered falsification of the wave/quantum thesis for learning and cognition
Research Desk (28)
- • Machine Learning
Agents still have the right skill. They stop picking it as the library grows
- • Continual Learning
A better router helps, but on CIFAR-100 the experts were the weak link
- • Continual Learning
A new model generation should earn authority, not win a contest
- • Machine Learning
Superintelligence has many definitions and no test. Here is what a test would need.
- • Machine Learning
Where AI already beats the best humans in 2026, and where it still doesn't
- • Machine Learning
The case for a 2027 intelligence explosion now rests on evidence the public can't see
- • Machine Learning
Penalising wrong answers only works if someone else is grading
- • Machine Learning
To stop a model forgetting, rehearse what it knew. New context won't do it.
- • Policy & Ethics
Anthropic told the SEC it may soon cross its own AI R&D line. The evidence is mostly its own.
- • Machine Learning
Investors are paying for agents that learn on the job. The hard part is knowing what they forgot.
- • Machine Learning
Tokens keep getting cheaper and AI bills keep binding. Count verified tasks instead.
- • Neural Interfaces
Brain implants decode well enough now. The fight is over who pays.
- • Policy & Ethics
Neural data law moved from the sensor to the inference this autumn. The definitions didn't follow.
- • Silicon
AI's 2027 ceiling is set by the grid, not the chip. Optics buys back only a few percent.
- • Machine Learning
A model rewrote another model's weights. Read what it took away.
- • Neural Interfaces
A speech implant drops its four-hour rule: reading Connect-One's first months.
- • Machine Learning
More thinking isn't more evidence: test-time compute meets correlated error.
- • Biocomputing
Organoids read Braille. The control experiment is the real result.
- • Photonics
After OCP 2026, the question for co-packaged optics is deployment.
- • Policy & Ethics
Neural data law keeps spreading state by state, and the definitions don't match.
- • Neuroscience
Neuroscience 2026: what we're watching in Washington next week.
- • Neural Interfaces
The 4,096-electrode era: where brain-computer interfaces stand today.
- • Machine Learning
Decoders that generalize: the case for brain foundation models.
- • Silicon
A billion neurons in silicon: neuromorphic computing grows up.
- • Policy & Ethics
Your brain data has a legal status now: the neurorights wave.
- • Photonics
Computing with light: photonics comes for AI's interconnect bottleneck.
- • Biocomputing
Wetware: the strange, serious rise of computers made of neurons.
- • Neuroscience
The wiring diagram decade: what connectomics just proved possible.
