Google trains SensorFM on a trillion wearable minutes

Google Research trained SensorFM on more than one trillion minutes of multimodal wearable signals from five million consented participants across over 100 countries. The SensorFM research release says the frozen representation outperformed feature-engineered baselines on 34 of 35 health tasks, while an agentic “classroom” explored more than 30,000 prediction-head candidates. Health-AI builders should file the result as a data-and-label advantage, not a clinical clearance: the archive’s ChatGPT Health shadow-clinic analysis is the reminder that a strong representation still needs a deployment evidence gate.