05 On-device AI · Sensing · Systems design
Latent Compute
Devices that learn what they can sense.
The question
Does a small language model with context from the device it runs on outperform the same model without it?
- Role
- Research direction
- Years
- 2026
- With
- H3M research
- Status
- Research
- Disciplines
- On-device AI, Sensing, Systems design
- Constraint
- A device’s sensors, emitters and telemetry carry information that operating systems and apps routinely filter, average or throw away.
Method
01/04
On install, the runtime finds what the device can sense and characterizes how reliable each source is.
Illustration, not measurements
02/04
Each family of sensors and telemetry becomes a module with the same shape.
Illustration, not measurements
03/04
Readouts specific to each device are learned by self-supervision, without labels.
Illustration, not measurements
04/04
Local agents read the result through one context interface of stable signals that don’t depend on the hardware.
Illustration, not measurements
Results
- Test designed, not yet runLatent Compute Runtime — research paper and test and evaluation planDesign documents and an early code scaffold. The primary test they define has not been run.The go/no-go test — the same small model, with and without runtime context, on the same hardware — is specified before any platform work.
Carried forward
Decide how you’ll know it works before you build it.
Credits
- Ryan BrantResearch direction
- H3MStudio