Google Research dropped something genuinely interesting this week: SensorFM, a foundation model that chews on more than a trillion minutes of raw, messy sensor data from over five million Fitbit and Pixel Watch users. The goal isn’t just another step counter. It’s a general-purpose intelligence layer that understands human physiology and behavior well enough to tackle 35 different health tasks, from sleep staging to cardiovascular risk, using a single shared model. That’s a sharp departure from the siloed, single-purpose algorithms currently running on your wrist.
The secret sauce is scale and a clever training trick. Drawing from over 100 countries and 20 device models, the team fed the model 34 features pulled from heart rate, motion, skin conductance, temperature, and altitude sensors. Instead of needing perfectly labeled data, SensorFM was trained in a self-supervised way, learning to reconstruct segments of data that were deliberately masked. The method, called “Adaptive and Inherited Masking,” explicitly teaches the model to handle both real-world dropouts and artificial gaps. The results were stark: the largest 100-million-parameter model saw a 31 percent lower reconstruction error than its smallest sibling, proving that, for this task, bigger really is better.
When put to the test on unseen data from nearly 14,000 participants, SensorFM’s learned representations, even with simple decoding heads, beat hand-crafted, supervised baselines on 34 out of 35 tasks. It was also far more label-efficient, needing fewer examples to adapt to tricky, subjective traits like depression and anxiety symptoms. In a fascinating meta-experiment, the researchers set up a “classroom” of LLM agents that competed to write the best downstream prediction code, running over 30,000 experiments. The models these AI agents discovered outperformed standard linear probes on 28 of the 35 tasks.
The practical payoff became clear when they plugged SensorFM into a personal health agent. Health summaries augmented with SensorFM’s predictions were rated significantly higher than a baseline version across the board. This hints at a future where your wearable isn’t just logging data but providing a continuous, deeply contextualized picture of your health that a personal AI assistant can actually reason about. The real test, of course, will be whether this lab performance survives contact with the chaotic reality of daily life outside a study.
