# NeuroGenesys

> NeuroGenesys is a frontier AI research lab publishing open work on continual learning, test-time compute, post-silicon compute and brain-computer interfaces, and measuring selection reliability in AI agents. Operated by AW3 Technology, Inc., San Francisco. Founded by Will Schulz. Contact: neurogenesys@aw3.tech.

Research papers are published openly, including pre-registered negative results. Research Desk posts are sourced analyses with links to primary sources. RSS: https://neurogenesys.ai/news/feed.xml (Research Desk) and https://neurogenesys.ai/research/feed.xml (papers, full text). Full text of the papers, posts, findings ledger and tool descriptions in one file: https://neurogenesys.ai/llms-full.txt.

## Hubs

- [Research papers](https://neurogenesys.ai/research): All papers, grouped by research area
- [Research archive](https://neurogenesys.ai/research/archive): All papers, chronological
- [Research library](https://neurogenesys.ai/research/library): External AI papers by other researchers that our work cites, with where each is used; searchable by field, year and study
- [Genesys Frontier Models](https://neurogenesys.ai/programs/genesys): The continual-learning world-model program, claim by claim: evidence level, prior work, where each claim fails, and the pre-registered tests running now
- [Findings ledger](https://neurogenesys.ai/research/findings): Every continual-learning finding with its evidence level (toy, small LLM, literature), post-hoc flag, date and where it is recorded
- [Research Desk](https://neurogenesys.ai/news): Sourced analysis of continual learning, AI agents, neurotech and BCI, photonics, AI silicon and policy, plus Field Watch: daily summaries of new research from other labs, not NeuroGenesys results
- [Free tools](https://neurogenesys.ai/tools): In-browser calculators: LLM API Cost Calculator; Token Counter & Context Window Checker; Test-Time Compute Cost Estimator; AI Eval Sample Size Calculator; LLM Training Compute Calculator; VRAM Calculator: Can I Run This LLM?; LLM Inference Speed Estimator (Tokens per Second); Agent Tool Count Calculator; LLM Judge Agreement Calculator; EU AI Act Risk Checker; Neural Data Rate Calculator (EEG & BCI); BCI Information Transfer Rate Calculator; Neural Data Privacy Law Checker; BCI Company Comparison
- [Can I run it? (open-weight LLM memory guides)](https://neurogenesys.ai/tools/can-i-run): VRAM needed at 4-, 8- and 16-bit, which GPUs and Macs fit, and tokens per second, for 24 open-weight models; plus what fits each GPU and Mac memory size
- [LLM API pricing guides](https://neurogenesys.ai/tools/llm-pricing): Input and output prices per 1M tokens, common workload costs and nearest alternatives for 18 Claude, GPT and Gemini models
- [Research tools](https://neurogenesys.ai/research/tools): Calculators and harnesses built from the papers
- [API and MCP server](https://neurogenesys.ai/developers): 61 operations: JSON/Markdown API for all content and the free tools (free), the research calculators and hosted selection reliability for AI agents (metered); OpenAPI at https://neurogenesys.ai/api/v1/openapi.json, MCP over Streamable HTTP at https://neurogenesys.ai/mcp
- [Selection reliability for agent teams](https://neurogenesys.ai/agents): Tells an AI agent team whether a regression is forgetting or mis-selection, and gates skill, memory and model updates before they ship
- [How selection reliability compares](https://neurogenesys.ai/compare): How selection reliability differs from agent evals and observability platforms
- [AI Academy](https://neurogenesys.ai/academy): Free courses on the mathematics of AI and the research frontier
- [Will Schulz](https://neurogenesys.ai/people/will-schulz): Founder; author of the papers
- [About](https://neurogenesys.ai/about)
- [Safety, ethics and neural data](https://neurogenesys.ai/trust)

## Topic hubs

- [All topic hubs](https://neurogenesys.ai/topics)
- [Superintelligence: definitions, evidence and timelines](https://neurogenesys.ai/topics/superintelligence): Updated Nov 10, 2026. What superintelligence means, where AI already beats the best humans and where it doesn't, how anyone could test it, and the 2026 forecasts and evidence on how fast it could arrive.
- [Neural data privacy laws: state-by-state tracker](https://neurogenesys.ai/topics/neural-data-privacy-laws): Updated Nov 09, 2026. Which US states protect neural data, how each defines it, the consent model, enforcement, and federal and international status. Every row links to its source.
- [Test-time compute and verification: when more samples hurt](https://neurogenesys.ai/topics/test-time-compute): Updated Nov 09, 2026. Why Best-of-N accuracy peaks and then falls, what verifier exploitability ε is, how to choose a sampling budget, and why correlated judges are worth less than they look.
- [Optical and analog AI compute: what survives measurement](https://neurogenesys.ai/topics/optical-analog-compute): Updated Nov 09, 2026. Claimed versus measured results from our pre-registered tests of photonic and analog in-memory hardware for AI inference: what failed, what narrowed, and what survives.
- [BCI companies compared (2026)](https://neurogenesys.ai/topics/bci-companies): Updated Nov 09, 2026. Neuralink, Paradromics, Synchron, Precision Neuroscience and Blackrock Neurotech compared on electrodes, reported performance, trials and FDA status, plus the academic speech-BCI records.

## Glossary

- [Glossary](https://neurogenesys.ai/glossary)
- [Verifier exploitability](https://neurogenesys.ai/glossary/verifier-exploitability): Verifier exploitability (ε) is the rate at which a verifier, such as a reward model, judge or checker, gives a high score to a wrong candidate. In Best-of-N sampling it decides whether more samples help or hurt; accuracy alone does not.
- [Co-packaged optics](https://neurogenesys.ai/glossary/co-packaged-optics): Co-packaged optics (CPO) puts the optical transceiver, the engine that converts electrical signals to light, on the same package as, or right next to, a switch or compute chip instead of in a pluggable module at the front panel. Shortening the electrical path cuts interconnect power in AI clusters.
- [Neural data](https://neurogenesys.ai/glossary/neural-data): Neural data is information generated by measuring the activity of a person's nervous system, for example from EEG headbands, implanted brain-computer interfaces or other neurotechnology. Several US states now treat it as sensitive personal data, but they disagree on whether it covers the peripheral nervous system and data inferred from other signals.
- [Organoid intelligence](https://neurogenesys.ai/glossary/organoid-intelligence): Organoid intelligence (OI) is a research field, named in 2023, that aims to use lab-grown 3D brain organoids (stem-cell-derived neural tissue) interfaced with electrodes and machine learning as a form of biological computing. It also studies learning and memory in a dish under an embedded ethics program.
- [Brain foundation model](https://neurogenesys.ai/glossary/brain-foundation-model): A brain foundation model is a large neural network pretrained, often self-supervised, on many hours of neural recordings across sessions, subjects or devices. It can then be adapted with little data to new decoding tasks or new people, instead of training a bespoke decoder for every patient and session.
- [Analog in-memory computing](https://neurogenesys.ai/glossary/analog-in-memory-computing): Analog in-memory computing (AIMC) performs matrix–vector multiplication inside the memory array that stores a neural network's weights. It uses device physics, Ohm's law for multiplication and Kirchhoff's current law for summation, instead of moving weights to a separate processor, which targets the energy cost of data movement in AI inference.

## Research papers

- [Merge confusable experts, split only separable ones](https://neurogenesys.ai/research/merging-splitting-moe-experts): Continual Learning, Synthesis, Nov 11, 2026. Re-read as a merge-and-split experiment, the lab's Split-CIFAR-100 task-count sweep (ledger 8.18) shows that under max-softmax routing, merging adjacent expert pairs (10 → 5 tasks) raised system accuracy from 0.342 to 0.422, while splitting every expert (10 → 20) cut it to 0.208.
- [Stable wrong beliefs evade every self-check: a literature synthesis on model self-knowledge and confident errors](https://neurogenesys.ai/research/model-self-knowledge-confident-errors): Evaluation, Synthesis, Nov 11, 2026. Semantic entropy flags confabulations at an average AUROC of 0.79 and exact-answer hidden-state probes reach 0.83–0.95, but on errors a model makes the same way every time detection falls to about 0.59–0.67.
- [Trust the premises, not the confidence: which belief to drop when an LLM contradicts itself, and why the solver is not the bottleneck](https://neurogenesys.ai/research/belief-repair): Continual Learning, Research note, Nov 10, 2026. When llama3.2 3B and llama3.1 8B contradicted themselves on MQuAKE 2-hop chains, repairing by graph entrenchment gave 85.8% and 95.2% accuracy, against 79.4% and 92.0% for model confidence; the pre-registered prediction that evidence priority would win failed.
- [Iterated self-closure settles fast, but only evidence that overrides self-belief makes the settled state true](https://neurogenesys.ai/research/iterated-self-closure): Continual Learning, Research note, Nov 10, 2026. On a toy knowledge-base transformer, iterated self-closure always converged: the unanchored loop changed 5.3% of 1,888 beliefs in round 1 and 0.06% in round 10.
- [A gradient kernel flags contradictions but does not predict forgetting, and steering rehearsal by it backfires](https://neurogenesys.ai/research/kernel-edit-ripple): Continual Learning, Research note, Nov 10, 2026. In a sequential toy update stream, low confidence predicted which facts would be forgotten at AUROC 0.737, and the best gradient-kernel variant reached only 0.663 (t = −8.1).
- [EWC stiffness freezes stale facts, and no per-fact Fisher ledger can release it](https://neurogenesys.ai/research/ewc-stiffness-stale-facts): Continual Learning, Research note, Nov 10, 2026. On a toy knowledge-base transformer with an exact Fisher, EWC at the dev-chosen λ raised unrelated retention from 44.3% to 92.0% while edit success on contradicting updates fell from 62.5% to 13.5%.
- [Route on shared statistics, never expert confidence: the router architecture for frozen experts](https://neurogenesys.ai/research/optimal-router-architecture): Continual Learning, Synthesis and results, Nov 12, 2026. On Split-CIFAR-100, routing by the experts' own confidence lost almost all accuracy to selection: error rose from 0.534 to 0.788 as experts went from 5 to 20, while within-task accuracy rose (ledger 8.18).
- [Successors should earn authority, not win contests: graded succession in continual learning](https://neurogenesys.ai/research/graded-succession-earned-authority): Continual Learning, Synthesis, Nov 11, 2026. The lab's record argues against the mechanisms "competition" invites: self-confidence filtering was worse than none, a stale majority left 44% of updated facts coherently wrong (E3), and selector error grew from 0.53 to 0.79 as experts went from 5 to 20 (ledger 8.18).
- [Coherence is selected, anchored, then pruned: a literature synthesis on logical coherence in LLM world models](https://neurogenesys.ai/research/logical-coherence-llm-world-models): Continual Learning, Synthesis, Nov 10, 2026. A 2026 benchmark found a frontier model 98.9% accurate on single questions but only 60.3% consistent across logically equivalent rewrites (an unreviewed preprint, as are many sources here).
- [EWC protects what its Fisher sees: a literature synthesis on elastic weight consolidation, scale and coherence](https://neurogenesys.ai/research/elastic-weight-consolidation-review): Continual Learning, Synthesis, Nov 10, 2026. EWC is the local quadratic of a KL anchor on old inputs, so in theory it protects exactly what its Fisher dataset exercises.
- [Real judges did not break Best-of-N at N ≤ 32, and measured ε did not predict the budget](https://neurogenesys.ai/research/real-judge-best-of-n): Evaluation, Research note, Nov 10, 2026. In a pre-registered test on 9,600 Llama 3.1 8B candidates, none of three real judges (a reward model, an LLM-as-judge and self log-probability) made Best-of-N accuracy fall between its best N and N = 32.
- [Giving experts a way to say "not mine": ten pre-registered selector and abstention experiments](https://neurogenesys.ai/research/selector-abstention): Continual Learning, Research note, Nov 9, 2026. On Split-MNIST, a task selector that never reads expert confidence, RanPAC over streaming-PCA features, reached 0.926 ± 0.004 class-incremental accuracy at selector ε 0.067, against 0.857 for nearest-class-mean and 0.786 for confidence selection.
- [A critic that shares the generator's prompt shares its errors: measured self-critique correlation](https://neurogenesys.ai/research/correlated-critique): Evaluation, Research note, Sep 9, 2026. A same-model critic given the generator's poisoned context note missed all 12 of 12 planted errors in both replicates, while fresh-context and different-model critics missed 0 of 12.
- [The novelty gate: five checks that caught seven fluent ideas before they shipped](https://neurogenesys.ai/research/novelty-gate): Methodology, Methods note, Sep 8, 2026. The novelty gate was calibrated on seven attractive ideas from one project, all of which felt like discoveries and none of which were, including an ε law that turned out to be Thurstone's 1927 identity.
- [The wavefield gate: Fourier modes are the wrong basis for visual computing](https://neurogenesys.ai/research/wavefield-gate): Photonics, Falsification study, Sep 8, 2026. The wavefield graphics pitch claimed a few thousand modes could replace millions of pixels, but Fourier modes needed 21.6% to 49.0% of coefficients to reach 35 dB PSNR, missing the pitch by 40 to 100 times.
- [The log-arithmetic gates: weight sharing buys compute, not storage](https://neurogenesys.ai/research/log-arithmetic-gates): Silicon & Analog Compute, Negative results, Aug 26, 2026. Eight novelty gates over two days found the log-neuron, super-neuron and photonic-multiplexing proposals not novel; the first three gates scored 0 of 16 threads novel, and no later gate recorded a novel thread.
- [Diamond is a material, not a system: no case for quantum memory in AI datacenters](https://neurogenesys.ai/research/diamond-photonics-gate): Photonics, Falsification study, Aug 4, 2026. Halving the 130 TB/s bandwidth of a 72-GPU rack would need about 7.2 trillion entangled pairs per second, while the best deployed-fibre entanglement swapping found runs at 1.5 pairs per second, a gap of about 12.7 orders of magnitude.
- [Can analog or photonic hardware replace LLM inference silicon? AIMC for compute, optics for decode interconnect](https://neurogenesys.ai/research/analog-photonic-codesign-synthesis): Silicon & Analog Compute, Synthesis, Jul 30, 2026. In the end-to-end decode build, analog mixture-of-experts at 10,108 µJ per token tied digital expert parallelism at 10,258 (±1.5%) while paying a 1.9-point accuracy tax.
- [Could the instrument have shown it? Power and saturation checks for pre-registered predictions](https://neurogenesys.ai/research/instrument-gate): Methodology, Methods note, Jul 30, 2026. On 29 July 2026, seven pre-registered predictions across four research tracks missed for one shared reason: the registered statistic could not have shown the effect, not that the hypothesis was wrong.
- [What language buys: re-sampling, transmission, and the limits of a discriminability metric](https://neurogenesys.ai/research/what-language-buys): Language, Negative results, Jul 30, 2026. In a small recurrent network, extra thinking steps raised accuracy from 0.978 at 16 steps to 1.000 at 256 when the noisy input was re-drawn each step, but stayed flat at 0.935 when the input was drawn once.
- [When the exploit is constructed: ranking collapses, counting survives](https://neurogenesys.ai/research/constructed-exploits): Test-Time Compute, Research note, Jul 30, 2026. At an exploit rate matched to three decimal places, picking the top-scoring candidate out of about 8,000 gave 0.862 accuracy with sampled exploits and 0.000 with exploits constructed against the verifier.
- [The capacity audit: most class-incremental routing error is placement slack, not capacity](https://neurogenesys.ai/research/capacity-audit): Continual Learning, Research note, Jul 30, 2026. In a vision-transformer audit on a 100-class benchmark, sphere packing stopped limiting class-incremental routing error at about 16 embedding dimensions.
- [The last hardware flank: analog prefill survives only as an energy-bound niche](https://neurogenesys.ai/research/analog-prefill-flank): Silicon & Analog Compute, Falsification study, Jul 30, 2026. Analog prefill's energy win over a digital baseline dilutes with prompt length: 19.1× at 512 tokens, 9.3× at 2048, 5.9× at 4096, 3.6× at 8192 and 1.7× at 32k, because causal attention cannot go on a crossbar.
- [The wrong phase: optical interconnect's case is in decode, not prefill](https://neurogenesys.ai/research/optical-interconnect-decode): Photonics, Synthesis, Jul 30, 2026. The talk claim that networking consumes up to 50% of Llama MoE prefill runtime blends three sources; TokenWeave measures 9–23% of end-to-end inference latency across three large models on an eight-GPU node.
- [What survived: a claims register with falsifiers for a pre-registered AI research program](https://neurogenesys.ai/research/claims-register): Methodology, Synthesis, Jul 29, 2026. The claims register lists verifier exploitability as the lead surviving thread: Best-of-N is bounded by ε, N*(ε) shipped as a tool, and ε swings 6.2× with the candidate pool.
- [Exploitability is a property of a verifier and its candidate pool, not the verifier alone](https://neurogenesys.ai/research/exploitability-verifier-pool): Test-Time Compute, Research note, Jul 29, 2026. The same verifier's measured exploitability ranged from 0.153 to 0.949, a factor of 6.2, depending only on which upstream prefilter supplied its candidates at a funnel width of ten.
- [Recurrent depth as test-time compute: parameter-efficient, extrapolating, and bounded by contractivity](https://neurogenesys.ai/research/recurrent-depth-test-time-compute): Test-Time Compute, Synthesis, Jul 27, 2026. A weight-tied recurrent-depth network matched a matched-compute feedforward network in distribution at 1/8 the parameters, on synthetic graph tasks such as reachability.
- [Thurstone bounds Best-of-N: a 1927 identity, calibrated for modern verifiers](https://neurogenesys.ai/research/thurstone-bounds-best-of-n): Test-Time Compute, Preprint draft, Jul 27, 2026. In the Best-of-N reference simulation (p = 0.10, d′ = 2, ε = 0.01), coverage approached 1 while accuracy peaked at 0.70 at N = 64 and fell to 0.36 by N = 8192.
- [Which sequence architecture keeps analog in-memory compute's advantage? A cost-model answer with a design rule](https://neurogenesys.ai/research/aimc-window-codesign): Silicon & Analog Compute, Defensive publication, Jul 27, 2026. In the cost model, analog in-memory compute's decode advantage on a transformer dilutes from 3.86× at 2k context to 1.05× at 131k, while a pure SSM holds 11.8× at every length.
- [The displacement theorem, and the aspect no-go](https://neurogenesys.ai/research/displacement-theorem): Mathematics & Logic, Machine-checked results, Jul 26, 2026. The displacement theorem shows that a distinction-free object level can be specified as such only from a metalanguage that itself contains a distinction; the paper states plainly that the proof is the pigeonhole principle.
- [Correction: the ε selection law is Thurstone's 1927 comparative judgment, not a discovery](https://neurogenesys.ai/research/epsilon-thurstone-correction): Foundations, Correction, Jul 26, 2026. The project's ε law, ε = Φ(−gap/(σ√2)), is Thurstone's 1927 Law of Comparative Judgment, Case V, an exact algebraic identity rather than a discovery.
- [Escaping the saturated game: plasticity, harder benchmarks, and retrieval in continual learning](https://neurogenesys.ai/research/three-frontiers-continual-learning): Continual Learning, Synthesis, Jul 26, 2026. Across 30 pre-registered continual-learning experiments (29 run), loss of plasticity behaved as an Adam, long-stream, small-network weight-norm pathology, with Adam losing plasticity 8× faster than SGD.
- [A closed-form spectral response for balanced perturbations of positive recurrences](https://neurogenesys.ai/research/balanced-spectral-response): Mathematics, Expository note, Jul 19, 2026. The closed-form growth response for balanced perturbations of a positive two-term recurrence is a worked example of the 1986 Arnold–Papanicolaou–Wihstutz weak-noise expansion, not a new theorem.
- [Spectral response of positive recurrences under memory-filtered balanced coefficient perturbations](https://neurogenesys.ai/research/spectral-response): Mathematics, Research note, Jul 19, 2026. For a positive two-term recurrence perturbed by balanced AR(1) memory noise, growth changes only at second order in noise strength ε, with no first-order effect.
- [Analog chips can do nonlinear ML — if you co-design around the device's native nonlinearity](https://neurogenesys.ai/research/analog-nonlinearity-feasibility): Silicon & Analog Compute, Research note, Jun 29, 2026. Device-shaped analog nonlinearities matched or beat ReLU after hardware-aware retraining on an MNIST MLP: tanh 0.932 and a ReRAM-like clamp 0.947 against ReLU's 0.938.
- [Catastrophic forgetting in class-incremental learning is a verifier-exploitability problem](https://neurogenesys.ai/research/forgetting-as-selection): Continual Learning, Preprint draft, Jun 27, 2026. With one frozen expert per Split-MNIST task, each near-perfect at 0.987 oracle-task accuracy, class-incremental difficulty reduces to selecting which expert to trust.
- [When more sampling hurts: verifier exploitability, not accuracy, bounds test-time compute](https://neurogenesys.ai/research/verifier-exploitability): Test-Time Compute, Superseded draft, Jun 27, 2026. The superseded draft's ε law is Thurstone's 1927 Case V identity, not a novel or universal law, per the project's own correction; its measured constants, tools and p/(1−p) threshold stand.
- [Where optics-native inference actually pays: a cost-model study](https://neurogenesys.ai/research/optics-native-inference): Photonics, Cost-model study, Jun 27, 2026. An optics-native state-space model's Phase-1 decode speedup of 15–56× shrank to single-digit× once a fused GPU removed about 5× of launch overhead.
- [Reverse-engineering a trained network for all-optical inference via random-feature kernels](https://neurogenesys.ai/research/optical-kernel-inference): Photonics, Falsification study, Jun 27, 2026. Requoted against a fused INT8 digital ASIC instead of an H100, the all-optical kernel's headline 48× speedup collapses to 1.23×, winning at only one of five complexities.
- [Complex amplitudes are decorative: a pre-registered falsification of the wave/quantum thesis for learning and cognition](https://neurogenesys.ai/research/negative-results-wave-cognition): Foundations, Negative results, Jun 27, 2026. At equal real-parameter budget, complex or wave amplitudes were load-bearing in none of three domains tested: continual learning, quantum-style cognition effects, and compositional binding.

## Research Desk

- [Agents still have the right skill. They stop picking it as the library grows](https://neurogenesys.ai/news/skill-routing-breaks-as-libraries-grow): Machine Learning, Nov 12, 2026. Two papers from October and November 2026 measure what happens to LLM agents as their skill libraries grow into the hundreds and thousands. The skills keep working; the agents stop choosing them. That is the failure our forgetting-as-selection results predict, and both papers fix it without touching the skills.
- [A better router helps, but on CIFAR-100 the experts were the weak link](https://neurogenesys.ai/news/router-shared-space-results): Continual Learning, Nov 12, 2026. We pre-registered six predictions about how to route inputs among frozen experts and tested them on 30 cached expert libraries. A router in a shared feature space beat the experts' own confidence, by a little without pretraining and by a lot with it. But once routing was good, a plain readout of the shared features beat the experts themselves by 13 points.
- [A new model generation should earn authority, not win a contest](https://neurogenesys.ai/news/successors-should-earn-authority): Continual Learning, Nov 11, 2026. When new models or new knowledge replace old ones, the usual picture is competition: old and new fight, the winner takes over. Our synthesis argues for graded succession instead. The successor gains authority step by step as independent evidence shows it is better, the old generation stays recoverable, and a protected constitution changes only by explicit amendment.
- [Superintelligence has many definitions and no test. Here is what a test would need.](https://neurogenesys.ai/news/what-would-count-as-superintelligence): Machine Learning, Nov 10, 2026. Labs now say superintelligence is in sight, but their definitions compare AI with different humans, and none comes with a way to check. Benchmarks saturate, ceilings are set by label errors, and the best long-task metric stops working at 16 hours.
- [Where AI already beats the best humans in 2026, and where it still doesn't](https://neurogenesys.ai/news/where-ai-already-beats-experts-2026): Machine Learning, Nov 10, 2026. AI now matches or beats top humans on olympiad maths, contest programming and most reasoning benchmarks, and has disproved a famous conjecture. It still fails on long reliable work, learning after deployment and the physical world, and many of the strongest claims are graded by the companies themselves.
- [The case for a 2027 intelligence explosion now rests on evidence the public can't see](https://neurogenesys.ai/news/how-fast-could-superintelligence-arrive): Machine Learning, Nov 10, 2026. Forecasters moved their dates for automated AI research back to 2027 this week, on the strength of METR's finding that Anthropic is improving about twice as fast as a year ago. Most of the evidence behind that finding is redacted, and the gains may be limited to tasks that grade themselves.
- [Penalising wrong answers only works if someone else is grading](https://neurogenesys.ai/news/confident-wrong-experts-selection): Machine Learning, Nov 09, 2026. Experts that never saw a task still vote on it with near-certainty. We ran ten pre-registered fixes. Penalties inside the expert failed; a selector that ignores expert confidence reached 0.926 on Split-MNIST.
- [To stop a model forgetting, rehearse what it knew. New context won't do it.](https://neurogenesys.ai/news/rehearse-old-knowledge-not-new-context): Machine Learning, Nov 09, 2026. We tested whether bridging knowledge added alongside new facts protects old ones. It doesn't, on a toy model or on Qwen 1.5B. Re-expressed old knowledge does, and outdated facts need pruning.
- [Anthropic told the SEC it may soon cross its own AI R&D line. The evidence is mostly its own.](https://neurogenesys.ai/news/anthropic-ipo-ai-rd-threshold): Policy & Ethics, Nov 09, 2026. Anthropic's amended IPO filing says it “may imminently meet” its policy's definition of AI-driven acceleration. The numbers that would settle it are internal, redacted and lagging.
- [Investors are paying for agents that learn on the job. The hard part is knowing what they forgot.](https://neurogenesys.ai/news/agents-that-learn-on-the-job-2026): Machine Learning, Nov 09, 2026. Trajectory just raised $45M at a $300M valuation to make deployed models keep learning. But on CL-Bench, the best system gained only 25.4% over a stateless baseline, and plain long context beat dedicated memory.
- [Tokens keep getting cheaper and AI bills keep binding. Count verified tasks instead.](https://neurogenesys.ai/news/ai-inference-cost-per-verified-task): Machine Learning, Nov 09, 2026. Apple cut some engineers' daily AI budget from $300 to $200 and Revolut moved coding to self-hosted GLM at about $0.30 per million tokens. Agentic workloads multiply tokens faster than prices fall. The unit that matters is cost per verified task.
- [Brain implants decode well enough now. The fight is over who pays.](https://neurogenesys.ai/news/bci-market-who-pays-2026): Neural Interfaces, Nov 09, 2026. This autumn brought a $250 million US round, a RMB 700 million Chinese one, and the first US billing codes for brain-computer interfaces. Regulatory class, durability evidence and payment now decide who reaches patients first.
- [Neural data law moved from the sensor to the inference this autumn. The definitions didn't follow.](https://neurogenesys.ai/news/neural-data-law-autumn-2026): Policy & Ethics, Nov 09, 2026. California now bars employers from using AI to collect neural data or infer emotions, and its governor says the law lacks key definitions. Six states still can't agree on what neural data is.
- [AI's 2027 ceiling is set by the grid, not the chip. Optics buys back only a few percent.](https://neurogenesys.ai/news/ai-power-wall-2027): Silicon, Nov 09, 2026. PJM is 6.8 GW short, turbines take six years, and FERC just pushed the backstop to 2027. Interconnect is a small slice of facility power, so photonics can't fix this.
- [What a model lets into long-term memory matters more than how it learns](https://neurogenesys.ai/news/gate-what-models-consolidate): Machine Learning, Nov 09, 2026. In our experiments, a per-example trust gate cut a poisoning attack from 94% to 4%, where a sleep phase added only 3.6 points. The leverage is in what gets consolidated.
- [Can a guardian chip stop a learning model from forgetting? Only partly.](https://neurogenesys.ai/news/guardian-core-continual-learning): Machine Learning, Nov 09, 2026. A protected core model on its own silicon can watch a continually learning model for drift. The hardware is ready; the judgment about what to keep is not.
- [To keep what a model knows, control what it is allowed to move](https://neurogenesys.ai/news/fortifying-knowledge-continual-learning): Machine Learning, Nov 09, 2026. Meta cut forgetting from 89% to 11% by updating 500 memory slots. Across continual learning, the methods that hold up limit what moves and keep old data in play.
- [A model rewrote another model's weights. Read what it took away.](https://neurogenesys.ai/news/llms-editing-open-weights): Machine Learning, Nov 09, 2026. An MIT agent edited 0.016% of a small Qwen model and its math scores jumped. Across weight editing in 2026, subtraction works and addition still lags.
- [A speech implant drops its four-hour rule: reading Connect-One's first months.](https://neurogenesys.ai/news/speech-bci-connect-one-goes-national): Neural Interfaces, Nov 05, 2026. The FDA has opened Paradromics' study to candidates nationwide. One participant is now talking through the device, and the hard evidence is still to come.
- [More thinking isn't more evidence: test-time compute meets correlated error.](https://neurogenesys.ai/news/test-time-compute-is-not-evidence): Machine Learning, Nov 05, 2026. Sampling a reasoning model sixteen times does not give you sixteen opinions. New preprints put numbers on an old statistical warning.
- [Organoids read Braille. The control experiment is the real result.](https://neurogenesys.ai/news/organoids-braille-control-experiment): Biocomputing, Nov 05, 2026. Three human brain organoids classified letters at 83%. A classifier with no organoids scored 82%, and the gap only opens once noise arrives.
- [After OCP 2026, the question for co-packaged optics is deployment.](https://neurogenesys.ai/news/co-packaged-optics-after-ocp-2026): Photonics, Nov 05, 2026. October's summit showed optics inside real racks, with open specifications forming behind them. The published numbers are still scarce.
- [Neural data law keeps spreading state by state, and the definitions don't match.](https://neurogenesys.ai/news/neural-data-law-delaware-vermont): Policy & Ethics, Nov 05, 2026. Delaware has added neural data to its opt-in categories, and Vermont has both recognized neurological rights and made neural data sensitive. The federal bill remains a study.
- [Neuroscience 2026: what we're watching in Washington next week.](https://neurogenesys.ai/news/neuroscience-2026-what-to-watch): Neuroscience, Nov 05, 2026. About 20,000 researchers meet on November 14. Our shortlist covers memory prosthetics, neuroethics and the tools that make new maps possible.
- [The 4,096-electrode era: where brain-computer interfaces stand today.](https://neurogenesys.ai/news/state-of-brain-computer-interfaces): Neural Interfaces, Aug 04, 2026. Channel counts are scaling like the early days of Moore's law — and the bottleneck is shifting from hardware to decoding.
- [Decoders that generalize: the case for brain foundation models.](https://neurogenesys.ai/news/brain-foundation-models): Machine Learning, Jul 28, 2026. Speech neuroprostheses crossed the usefulness threshold by borrowing the modern ML playbook — and the next leap is models that transfer across sessions, subjects, and tasks.
- [A billion neurons in silicon: neuromorphic computing grows up.](https://neurogenesys.ai/news/neuromorphic-computing-at-scale): Silicon, Jul 21, 2026. The brain runs general intelligence on twenty watts. Event-driven silicon is how the industry starts closing a five-order-of-magnitude gap.
- [Your brain data has a legal status now: the neurorights wave.](https://neurogenesys.ai/news/neural-data-privacy-laws): Policy & Ethics, Jul 07, 2026. Six US states, a global UNESCO framework, and a fast-maturing consensus: neural data is becoming a protected category — and the industry should welcome it.
- [Computing with light: photonics comes for AI's interconnect bottleneck.](https://neurogenesys.ai/news/photonics-ai-interconnect): Photonics, Jun 30, 2026. As clusters scale past a hundred thousand accelerators, copper hits its physical limits — and co-packaged optics moves from specialty tech to mainstream cluster design.
- [Wetware: the strange, serious rise of computers made of neurons.](https://neurogenesys.ai/news/biological-computing-wetware): Biocomputing, Jun 18, 2026. Lab-grown neurons learned Pong in five minutes. Now biological computing has commercial hardware, a research agenda — and an ethics literature it actually needs.
- [The wiring diagram decade: what connectomics just proved possible.](https://neurogenesys.ai/news/connectomics-milestones): Neuroscience, Jun 09, 2026. A whole fly brain, a cubic millimeter of mouse cortex with function attached — the connectome went from fantasy to roadmap in eighteen months.

## Free tools

- [LLM API Cost Calculator](https://neurogenesys.ai/tools/llm-cost-calculator): In-browser, free. Updated Nov 09, 2026. Compare what a workload costs per call, per day and per month across Claude, GPT and Gemini models, with prompt caching and long-context pricing.
- [Token Counter & Context Window Checker](https://neurogenesys.ai/tools/token-counter): In-browser, free. Updated Nov 09, 2026. Paste text or load a file to count its tokens, see what sending it costs as input, and check which models it fits in. Nothing leaves your browser.
- [Test-Time Compute Cost Estimator](https://neurogenesys.ai/tools/test-time-compute-cost): In-browser, free. Updated Nov 09, 2026. What Best-of-N and self-consistency cost per query, and the accuracy they can buy: oracle coverage and majority vote against the number of samples.
- [AI Eval Sample Size Calculator](https://neurogenesys.ai/tools/eval-sample-size): In-browser, free. Updated Nov 09, 2026. How many eval items you need: the margin of error on one model's accuracy, and the sample size to detect a difference between two models, paired or not.
- [LLM Training Compute Calculator](https://neurogenesys.ai/tools/training-compute-calculator): In-browser, free. Updated Nov 09, 2026. Training FLOPs from parameters and tokens (6ND), GPU-hours at a realistic utilization, wall-clock time and cost — or the compute-optimal model for a budget.
- [VRAM Calculator: Can I Run This LLM?](https://neurogenesys.ai/tools/vram-calculator): In-browser, free. Updated Nov 09, 2026. GPU memory for an LLM at any quantization and context length — weights, KV cache and overhead — and which GPUs and Apple Silicon Macs can hold it.
- [LLM Inference Speed Estimator (Tokens per Second)](https://neurogenesys.ai/tools/inference-speed-estimator): In-browser, free. Updated Nov 09, 2026. Tokens per second and time to first token for a model on a GPU or Mac, from memory bandwidth and compute, and how batching changes throughput.
- [Agent Tool Count Calculator](https://neurogenesys.ai/tools/agent-tool-count-calculator): In-browser, free. Updated Nov 09, 2026. How often an agent picks the wrong tool as its library grows, from published benchmarks: every tool in the prompt, embedding or BM25 shortlists, with the measured range marked.
- [LLM Judge Agreement Calculator](https://neurogenesys.ai/tools/llm-judge-agreement): In-browser, free. Updated Nov 09, 2026. Check an LLM-as-judge against human labels: Cohen's kappa with a 95% interval, the confusion matrix, sensitivity and specificity, and how many human labels the judge is worth.
- [EU AI Act Risk Checker](https://neurogenesys.ai/tools/eu-ai-act-checker): In-browser, free. Updated Nov 09, 2026. Answer a few questions to see your EU AI Act risk tier, the obligations with their articles, and the deadlines after the 2026 Digital Omnibus.
- [Neural Data Rate Calculator (EEG & BCI)](https://neurogenesys.ai/tools/neural-data-rate-calculator): In-browser, free. Updated Nov 09, 2026. Raw data rate and storage from channels, sample rate and bit depth, for EEG headsets through intracortical arrays — and whether it fits a wireless link.
- [BCI Information Transfer Rate Calculator](https://neurogenesys.ai/tools/bci-itr-calculator): In-browser, free. Updated Nov 09, 2026. Wolpaw ITR for any brain-computer interface: bits per selection and per minute from targets, accuracy and timing, typing-speed equivalents, a chance-level check and published benchmarks.
- [Neural Data Privacy Law Checker](https://neurogenesys.ai/tools/neural-data-law-checker): In-browser, free. Updated Nov 09, 2026. Pick the states your users are in and the kind of data you collect to see which neural data laws apply, from when, and what they require.
- [BCI Company Comparison](https://neurogenesys.ai/tools/bci-company-comparison): In-browser, free. Updated Nov 09, 2026. Filter brain-computer interface companies by approach, sort by electrode count and compare any of them side by side on device, results and regulatory status.

## Research tools

- [verifier-budget](https://neurogenesys.ai/research/tools/verifier-budget): In-browser · Python CLI. The modelled Best-of-N sampling budget N*(ε) and the selection rule (argmax vs. consensus) from a verifier's measured exploitability, with an ε estimator for a labelled validation set. A model: real judges did not collapse by N = 32 in the lab's pre-registered test (findings 8.13).
- [verifier-panel](https://neurogenesys.ai/research/tools/verifier-panel): In-browser · Python CLI. How much an LLM-judge panel is really worth: error correlation, effective panel size, the majority-vote ceiling — from two numbers or straight from a critics × items miss matrix — and a tail check before trusting a Φ-based ε.
- [selection-reliability](https://neurogenesys.ai/research/tools/selection-reliability): In-browser · Python CLI. Did the agent forget a skill or pick the wrong one? Selection error ε with confidence intervals, a forgetting vs. mis-selection verdict from counts or a decision log, and the largest library a router can face before accuracy drops below target.
- [exploitability-audit](https://neurogenesys.ai/research/tools/exploitability-audit): In-browser · Python CLI. Is your verifier or reward model exploitable? Measured ε, the consensus threshold ε* = p/(1−p) and the empirical Best-of-N peak N* from a log of scored candidates, broken out by pool and by problems the verifier saw in training.
- [selector-tune](https://neurogenesys.ai/research/tools/selector-tune): In-browser · Python CLI. Fix the selector, not the skills: per-candidate score standardisation, a deferral margin that meets a target ε and the smallest verifier shortlist, fitted on half a decision log and judged on the other half.
- [emergence-audit](https://neurogenesys.ai/research/tools/emergence-audit): Python harness. Four owned controls — stateless generator, non-solution generator, prior swap at fixed fit, family distance — run against your own training script for any "X emerges from a network" claim.
- [threshold-audit](https://neurogenesys.ai/research/tools/threshold-audit): In-browser · Python. Before you register a threshold: its standard error, the instrument's ceiling and the noise band of range statistics. After the run: was the miss malformed, underpowered or substantive?
- [photonics-budget](https://neurogenesys.ai/research/tools/photonics-budget): In-browser · Python. Four calculators from the log-arithmetic gates: the optical recirculation pass budget, the nonlinear material figure of merit, the logarithmic-number-system energy check, and why fixed pre-ADC aggregation collapses to one matrix.

## AI Academy

- [Probability & Inference](https://neurogenesys.ai/academy/probability-and-inference): Foundations track, 4 lessons. Reasoning under uncertainty: the rules of probability, Bayes' rule, why averages concentrate, and how maximum likelihood turns a model into a loss function.
- [The Geometry of Learning](https://neurogenesys.ai/academy/linear-algebra-for-learning): Foundations track, 3 lessons. Linear algebra as the language of neural networks: vectors and similarity, matrices as maps, and the eigenvectors and singular vectors that reveal what a layer does.
- [Information Theory](https://neurogenesys.ai/academy/information-theory): Foundations track, 3 lessons. Shannon's measure of surprise, the cross-entropy loss every language model minimizes, and the deep equivalence between predicting data and compressing it.
- [Optimization & Generalization](https://neurogenesys.ai/academy/optimization-and-generalization): Foundations track, 4 lessons. How networks learn and why they work on data they have never seen: gradient descent, backpropagation, stochastic optimizers, and the surprises of overparameterized models.
- [Transformers & Language Models](https://neurogenesys.ai/academy/transformers-and-language-models): Frontier track, 4 lessons. What a language model is, how attention and the transformer block work, and what scaling laws do — and don't — tell us about training at frontier scale.
- [Reasoning Models & Reinforcement Learning](https://neurogenesys.ai/academy/reasoning-and-rl): Frontier track, 3 lessons. From human-preference training to reinforcement learning on verifiable rewards: how reasoning models are trained, and what the evidence says that training actually buys.
- [Test-Time Compute & Verification](https://neurogenesys.ai/academy/test-time-compute-and-verification): Frontier track, 4 lessons. Spending compute at inference — sampling, Best-of-N, majority vote — and why the verifier, not the generator, so often sets the ceiling. Built on NeuroGenesys's own research and tools.
- [Agents & Tool Use](https://neurogenesys.ai/academy/agents-and-tool-use): Frontier track, 3 lessons. Models that act: function calling, agent loops with memory and planning, and how to measure long-horizon autonomy without fooling yourself.
- [Interpretability & Alignment](https://neurogenesys.ai/academy/interpretability-and-alignment): Frontier track, 3 lessons. Looking inside models — features, circuits, superposition — and the techniques and evaluations used to make capable systems behave as intended.
- [Efficient AI: Inference, Sparsity & Hardware](https://neurogenesys.ai/academy/efficient-ai): Frontier track, 3 lessons. Where the cost of AI actually goes — memory bandwidth, KV caches, batching — and the methods and hardware that cut it, including the lab's own analog and photonic results.

## Optional

- [Research programs](https://neurogenesys.ai/programs): The Genesys, Cortex, SynapseOne and Lumen research programs: continual-learning world models, neural decoders, analog in-memory compute and photonics, studied through papers, cost models and literature reviews.
- [The science](https://neurogenesys.ai/science): The thesis that ties NeuroGenesys research together: what brains suggest about continual learning, and which of those ideas are tested in models and cost models today.
- [Measured results and milestones](https://neurogenesys.ai/milestones): The numbers NeuroGenesys research stands behind: modelled analog compute results, exact factorizations and honest negative results.
- [Submit research](https://neurogenesys.ai/research/submit): Submit a paper, negative result or replication to the research archive. Every submission is reviewed before it is published.
- [Reference videos](https://neurogenesys.ai/academy/videos): Curated lectures and explainers to watch alongside the Academy courses and the research: continual learning, verifiers, transformers and the math of AI.
- [Access](https://neurogenesys.ai/access): Three ways to work with NeuroGenesys: free open research, metered API and MCP credits for builders, and design partnerships and research collaborations.
- [Contact](https://neurogenesys.ai/contact): Contact the lab about replicating a published result, selection reliability for your agents or a research collaboration.
- [Newsletter](https://neurogenesys.ai/newsletter): Subscribe to or manage the newsletter of new papers and Research Desk posts.
- [Sign in](https://neurogenesys.ai/signin): Sign in with 1Profile to save papers, posts and tools, follow submissions, and manage API keys and credits.
- [Your account](https://neurogenesys.ai/account): Signed-in dashboard: saved items, your submissions, newsletter, API keys and credits, and hosted selection-reliability projects.
- [Privacy policy](https://neurogenesys.ai/privacy): How AW3 Technology handles data for NeuroGenesys: no training on Genesys API traffic, no neural data collected today, and what is collected and why.
- [Terms of service](https://neurogenesys.ai/terms): Terms of Service for NeuroGenesys, operated by AW3 Technology, Inc.
- [Cookie policy](https://neurogenesys.ai/cookies): Cookies for technical necessity only, with no third-party advertising cookies, cross-site trackers or marketing pixels.