| When you hear | What this is |
|---|---|
| “The model degraded, we have to retrain” | Re-adapts in place. Skip the retrain pipeline. |
| “Comms-denied, sensitive, disconnected” | Learns on the asset, no cloud. Units can share learned behavior peer-to-peer with no server, and a received behavior can’t destabilize the recipient |
| “That data can’t leave the box, the site, or the country” | Adaptation happens where the data already is. Nothing is transmitted, so there is nothing to transfer, disclose, or put in a data-processing agreement |
| “Remote condition monitoring” | Learns each unit’s own normal and responds locally, without adding false-alarm burden |
| “Every unit drifts differently” | Vehicles, ag equipment, wells, fleets, HVAC: per-unit adaptation instead of one averaged model |
| “The requirement changed after we shipped” | Bind a new signal or teach a new objective in the field, with no firmware redeploy |
| “It’s battery-powered, it wakes up on events” | Learns from irregular, duty-cycled streams where the interval between samples carries the signal |
| Edge silicon and MCU vendors | A learning workload for their chips, not just inference |
Opportunity
Where bounded live learning pays
Anywhere a frozen model goes stale and another cloud retrain is out of reach, not worth the cost, or the wrong answer because the asset stays offline.
When to think of us
Now
Condition monitoring and predictive maintenance
Battery-wireless and edge sensors where cloud retraining isn’t practical and each asset’s normal drifts. The unit self-commissions from minutes of the machine’s own stream instead of a multi-day baselining campaign, and keeps tracking that machine for the rest of its life.
Defense edge
Comms-denied is the default and the adversary keeps changing tactics. There is no cloud and no reward channel in the field, so a bounded learner that manufactures its own training signal on-device is the only thing that fits.
Comms-denied is also a legal posture, not only a physical one. Mission data that can’t cross a classification or national boundary can’t be shipped home for retraining either, which rules out the entire train-elsewhere-and-push pattern before anyone evaluates the model. Learning on the unit avoids the issue entirely.
Horizontal
Edge-AI silicon and IP
Every major vendor shipped edge-AI acceleration in the last two years, and it is inference silicon. Where learning is marketed at all, it’s a last layer bolted onto a frozen trunk. The learning half of the stack doesn’t exist.
This one fits the cheapest parts. The structural bound sizes a fixed-point format that can’t overflow, so the complete learning step, not just inference, runs in 16-bit integer arithmetic on silicon with no floating-point unit. That’s the license: bounded on-device learning as a horizontal primitive, not one vertical’s app.
Next
Robotics and autonomy
Embodied, non-stationary, and full of faults nobody tuned for. Bounded live adaptation rides on top of the stack that’s already there.
Safety-critical and medical
This is the ceiling: a bounded change envelope is exactly the shape regulators have been drafting pathways around. It is high-margin and real because the primitive exists.
There is a second blocker here that has nothing to do with model safety: the data itself. Patient signals are the most governed data on Earth, and rightly so. The usual improvement loop of collecting, shipping, retraining centrally, and pushing back has to account for every place those signals land, every party that touches them, and every border they cross. A learner that adapts on the device removes that loop rather than documenting it. The recording never leaves the machine.
The constraint
Every industry on this page runs the same policy: train, freeze, ship. Aviation is just the one place it’s written down.
EASA, Europe’s aviation safety regulator, covers models “frozen at the time of approval, but not adaptive or online learning processes” and states that “such applications will not be accepted by EASA at this stage.” DARPA’s Assured Autonomy program names the assumption directly: “current assurance approaches are predicated on the assumption that once the system is deployed, it does not learn and evolve.”
Everywhere else the freeze is unwritten and paid for anyway: in retrain pipelines, truck rolls, baselining campaigns, and models fitted to the fleet instead of the unit. Remove the divergence and you remove the reason to freeze.
Read on