MIT researchers introduced FTTE (Federated Tiny Training Engine), which transmits only a subset of model parameters per device, uses asynchronous aggregation weighted by data freshness, and reduces the communication payload by 69%—enabling smartphones, sensors, and smartwatches to participate in privacy-preserving AI training without sending raw data to a server. On standard heterogeneous-device benchmarks, FTTE completed training 81% faster than baseline federated learning and cut per-device memory overhead by 80% with near-equivalent accuracy. The underlying system addresses a known failure mode of federated learning: capable devices idling while waiting for resource-constrained peers, which FTTE resolves by accumulating updates asynchronously. The full paper is on arXiv.