Continual learning
Models that keep improving in deployment — no retrain cycle, no catastrophic forgetting.
AI Research Lab
Living models.
AI that runs continuously, remembers, learns in deployment, and acts. We invent it — and put it to work inside the real operations of Japanese industry.
A categorically different object.
A living model lives inside a real environment. It holds persistent memory, learns while it runs, consumes a real-time stream, and takes action on its own. It works 24/7 and gets better the longer it runs.
That is not a chatbot, not retrieval over your documents, not a one-off fine-tune. It is a different kind of thing — and building it is a research problem worth a lab.
The invention
Four capabilities, composed into one.
Streaming & live
Always-on models that consume a continuous data stream and act in real time, not discrete request-and-response.
Recurrent memory
Persistent recurrent state (RLM) for unbounded memory and long-horizon tasks.
Agentic
Autonomous, multi-step action that drives real software — not question-answering.
Composed, these are living models.
The business
We commercialize the lab from Layer 3.5.
Between raw foundation models and a company’s real operations sits a missing layer. We call it Layer 3.5 — the seat where a living model plugs into a live business process and takes on work static AI cannot. Partners bring the domain and the environment; we bring the AI.
How we work with partners Foundation models Layer 3
Living models Where we operate Layer 3.5
Your real operations Ground floor
Proving grounds
One technology. Three live environments.
The living-model stack is horizontal — it runs wherever there is a live stream, real drift, long-horizon state, and consequential action. Japan’s deep-tech industry is exactly that.
Factory automation & robotics
Live sensor, vision and control streams; constant drift; long-horizon line state.
Finance & FinTech
Live transaction and market streams; patterns that shift by the hour; per-entity memory.
Energy, battery & grid
Continuous telemetry; real-time optimization; online degradation learning.
Sovereign AI
Sovereign AI for Japanese industry.
Japan’s industry runs on Silicon Valley’s AI stack. A country’s core operations deserve a domestic, adaptive AI layer it controls. Spicedust is that layer — a living model that runs inside real industrial operations and answers to the businesses that depend on it.
About the companyBring us a live environment.
A continuous stream, real drift, decisions that matter. We’ll bring a living model that works inside it. First engagements are a scoped, paid pilot on one workflow.
Become a design partner