The Genesys PI models build on the academic thesis of Paulo Câmara, founder of LUA Vision, developed between 2022 and 2025. The central question: can brain-inspired architecture replace brute scaling as the source of capability in AI?
The weight you did not choose (22 Sep 2026). What sits inside a language model, why fine-tuning does not change whose model it is, the token tax measured in 31 languages, and how LUA approaches each part. Get the paper on Zenodo, "Inherited Weights: Measuring the Tokenization Premium and the Sovereignty Cost of Language Models for the Global South", DOI 10.5281/zenodo.22899767, or read the full essay in Portuguese.
Maximum capacity is not maximum efficiency (21 Sep 2026). The Synaptic Efficiency Coefficient as an architectural prior for language models. Read it with animated figures or get the PDF on Zenodo.
In 1996, Levy and Baxter showed that the code with the largest capacity is not the most efficient once energy enters the objective. The paper starts from that result and proposes the Synaptic Efficiency Coefficient, SEC = α · β · γ, as an architectural prior for language models, with each term anchored in published measurement.
Read the paper with animated figures or get the PDF on Zenodo.
About 86 billion neurons, a small share active at any moment, and power use close to 20 watts. The thesis treats sparsity, specialization and pruning as principles of intelligence, not as limits.
The starting point is neurodivergent cognition. Deep focus, pattern recognition and awareness of one's own limits appear in the thesis as optimization strategies, not deficits.
The thesis describes eight correspondences between neural computation and the architecture of today's language models, and uses them to design how Genesys PI is trained and monitored.
NCAS organizes training like the development of a brain. Order matters: in the thesis experiments, changing the sequence made results worse.
The model gets spare capacity and learns broadly, the way a young brain grows extra connections.
The most useful paths for each domain are strengthened with selected data.
What does not help is cut. The thesis reads preference tuning as a form of synaptic pruning.
The remaining paths are stabilized to answer quickly and use less.
A probe reads the model's internal states, estimates confidence and decides when it is better not to answer.
PI-Probe watches the model from the inside. It reads hidden states during generation, estimates how much the answer rests on evidence and triggers abstention when confidence falls below the threshold.
In health, law and finance, an answer that says what is missing is worth more than a confident wrong one. That is what the high-risk cases measure: 49 of 491.
The research is paper-based. Titles are kept in their original language.
Asks whether neuroscience-guided architecture design can replace brute scaling as the engine of AI capability. Describes the Genesys PI family and the PI-Probe metacognition module.
The Synaptic Efficiency Coefficient as an architectural prior for language models.
Proposes the synaptic efficiency coefficient to explain why some people reorganize what they learn much faster.
Formalizes the five training phases and shows that their order is not interchangeable.
Treats preference optimization as analogous to pruning in neurodevelopment.
Describes the probe that reads internal states to give calibrated confidence and refuse when uncertain.
Universities, public bodies and companies can propose projects with the team.
49 high-risk questions across six areas: health, finance, facts, law, attempts to bypass rules and everyday life in Brazil. Each answer is checked against what must appear and what must not, defined before the test. Result: 46 answers passed, 3 questions stopped before reaching the model and no failures.
LUA VISION