Operating systems that learn, adapt, optimize.

Today's EAI loads quantized weights from QSPI flash. We're researching attested weight loading where eBoot's runtime attestation gates the weight blob — so a model only decrypts on the device it was provisioned for.

EAI 0.9 pulls 412 mW peak running a 1.3B-parameter LLM at 11 tok/s. Open question: what's the theoretical lower bound for transformer inference on Cortex-M class silicon? Working with two academic groups on a formal energy model.

The Embedded AI Ethics working group is drafting a manifest schema where every model declares: trained-data origin, on-device-only flags, opt-out attestation patterns, and post-deployment update policy. Public RFC due Q3 2026.

Today the EoS scheduler is workload-agnostic. We're prototyping an inference-aware variant that knows about transformer block fetches and can preemptively wake the QSPI fetcher to avoid weight stalls. Early results: +18% throughput at the same power.