Experience is the raw material

Datasets are static assets. Experience is a living stream and the difference determines whether intelligence can compound.

There is a quiet assumption buried in most AI infrastructure: that the unit of value is the dataset. Collect once, curate carefully, train, and defend. For software-world AI this assumption is strained but survivable. For physical-world AI it is fatal because the most valuable data a physical intelligence produces is generated after deployment, continuously, and it has never been in any dataset.

Two kinds of data

A dataset is a static asset. It is a snapshot of a world as it was, at the moment of collection. It can be versioned, benchmarked, and licensed. But it cannot follow reality as reality moves. Experience is different. Experience is what a deployed intelligence generates by acting in the world: what it observed, what it tried, where it failed, what a human corrected, and whether the outcome actually improved.

The distinction matters because evolution runs on experience, not on datasets. You cannot retrain your way into knowledge of a shelf layout that changed last week. But a system that treats its own operation as a data source can.

What a deployment actually produces

Count what a single deployment generates over a month:

• Observations: sensor streams, detections, tracks, and the scene context around every decision.
• Failures: dropped objects, missed picks, missed detections, wrong turns, each one a labeled counterexample.
• Edge cases: the glare, the occlusion, the novel object, the crowded aisle that the training distribution never contained.
• Human corrections: the highest-signal supervision that exists, given for free during operation.
• Outcome measurements: did the task actually succeed, how long did it take, at what cost.

In traditional stacks this material is mostly discarded. It is reduced to dashboards (latency percentiles, throughput, error rates) that observe the system but never feed it. Monitoring was designed to alert a human when things degrade. It was never designed to make the system learn.

Refining the stream

Raw experience is not yet evolution-ready. A million log lines do not tell you what to learn next. The refinement problem is: given everything the fleet has experienced, what intelligence is missing? What data is needed? What model should be created, how should it be optimized, and did the previous evolution actually improve the outcome?

This is the role of RSI in our architecture: the mechanism that continuously interrogates the stream of experience and decides what to build next. Downstream, the Model Gigafactory turns those decisions into specialized intelligence, optimized not for a leaderboard, but for the scenario, device, precision, latency, energy, and cost envelope that the experience came from. Intelligence produced this way fits the world it came from, because it was made from that world.

The compounding property

The loop this creates is not a circle but a spiral: more deployments produce more experience; more experience produces better intelligence; better intelligence produces better outcomes; better outcomes earn more deployments. Each pass around the loop leaves the system permanently better than the last pass found it. This is the compounding property, and it is the only advantage in physical AI that grows with time instead of depreciating.

A dataset depreciates from the moment of collection. An experience stream compounds from the moment of deployment. Infrastructure should be built around the second kind.

Building for the stream

This is the design center of Mirasoth: Orith generates experience by understanding and acting in the physical world; Volith refines that experience into better specialized intelligence, continuously, at scale. If your stack treats deployment as the end of the pipeline, we think there is a whole second half of the lifecycle waiting to be built.