Not a neural network with biological names — a functional reproduction of the circuits that produce intelligence.
A biologically-faithful neural substrate that learns new objects from a single exposure, knows what it doesn't know, directs its own gaze toward what surprises it, and consolidates memories during sleep. It learns through local plasticity rather than backpropagation, runs continuously rather than in train-then-test phases, and carries a metabolic clock and a developmental course. It is a running system, not a simulation of one.
Three outputs, not one. A system that can decline to answer is a different kind of system. The internal organisation that produces that behaviour is the contribution, and it is not documented publicly.
Learns a new object from a single exposure through hippocampal episodic binding. Multi-view understanding built through active saccade exploration.
When uncertain, reports what it doesn't know rather than hallucinating. An internal gate decides whether it knows enough to answer at all.
Directs its own gaze via a superior colliculus with inhibition-of-return. Fixates on the most surprising part of a scene, avoids re-examining what it already knows.
Memories captured during the day consolidate during sleep through complementary learning systems replay, without catastrophic forgetting.
Binds objects to locations through hippocampal place-cell analogs. Answers "where did I see that?" from episodic memory, not a lookup table.
Every fixation has a reason — novelty, unfamiliarity, goal relevance. The system explains why it looked where it looked, in mechanistic terms.
The self-directed stack — the layer that forms its own goals rather than reacting — outperforms a reactive baseline by roughly fifty times. Every ablation is attributable: the gain traces to identifiable subsystems rather than to the ensemble, which is the harder result and the one that matters.
Learning without forgetting is closed end to end. Episodes captured while awake consolidate during sleep and survive subsequent learning — the failure mode that defines continual learning does not occur here.
Accuracy is treated as a population property across seeds rather than a single benchmark number — developmental stochasticity is a property of the system, not noise to be averaged away.
It is not a model you download. It is a runtime that keeps organisms alive.
Each organism runs its own loop on a warm pool with cold routing and eviction, and persists across restarts through digest-verified checkpoints. Environments are pluggable and arrive as declarations rather than code — bring a simulator, a dataset, or a live feed, and swap the world under a running organism without replacing the organism.
Reachable over a versioned HTTP API and a CLI, multi-tenant with owner-scoped enforcement, with an async job surface, spend ceilings on any outside call, and a supervision console that renders what each organism can do, what it is working toward, and what it has spent. One deployment is live.
A body is not a demo surface for this system. It is the condition it was built for.
Curiosity needs somewhere to look. One-shot learning needs a novel object placed in front of it. Abstention only matters when something is about to be acted on. Every capability above was designed against an agent that senses and moves, and the same substrate that runs in simulation runs on the device with no retraining step in between.
Current work targets humanoid and quadruped platforms with onboard depth cameras and edge compute: local sensing and reflexes on the device, consolidation off it, and a refreshed substrate returned after sleep.