The Coaching Wearable Needs Two Brains

Every wearable AI coach shipping today has the same architecture: sensors collect data, a dashboard displays it, and a chatbot answers questions about it. The chatbot is an LLM. The sensors are commodity. The intelligence gap is in between.

Google just filled it.

TL;DR

• Google Research published SensorFM on July 10 - a foundation model for wearable health pretrained on one trillion minutes of sensor data from five million people.
• SensorFM ingests 34 aggregate features from five consumer-grade sensors (PPG, accelerometer, EDA, skin temperature, altimeter) over a 24-hour context window and transfers to 35 health prediction tasks, beating purpose-built models on 34 of them.
• The model powers a Personal Health Agent that interprets physiological state qualitatively rather than dumping numbers - this is the JITAI (just-in-time adaptive intervention) layer that coaching wearables have been missing.
• The ambient AI coaching wearable needs two foundation models: one for language (what to say) and one for physiology (when to say it). SensorFM is the second one.

Why Does a Coaching Wearable Need a Foundation Model for Sensors?

A just-in-time adaptive intervention - JITAI in behavioral science - is a coaching nudge delivered at the precise moment the user is both in need of it and receptive to it. The timing is the entire value proposition. A perfectly worded coaching insight delivered during a stress spike when the user cannot engage is worse than silence. It erodes trust.

The timing problem requires understanding physiological context across hours and days. Not just "heart rate is elevated right now" but "heart rate has been elevated for 40 minutes following a pattern that historically precedes this user's anxiety episodes, and their HRV recovery signature suggests they are approaching a window of receptivity."

That reasoning requires a model trained on longitudinal physiological patterns. Not a rule engine. Not a threshold. A foundation model that has seen millions of people's sensor streams and learned the temporal grammar of human physiology.

SensorFM is that model.

What SensorFM Actually Does

SensorFM processes 34 one-minute aggregate features organized into seven categories, drawn from five sensors available on any modern smartwatch. The context window is 24 hours. The model was pretrained on data from five million people - one trillion minutes of continuous physiological recording.

The results are striking. Across 35 downstream health tasks - cardiovascular risk, depression screening, anxiety detection, sleep staging, activity classification, and more - SensorFM outperforms task-specific models on 34 of them. With 60 contiguous minutes of data missing, it retains 99.7% step-count accuracy and 99.9% deep-sleep accuracy. The representations generalize without task-specific fine-tuning.

The architecture detail that matters for builders: SensorFM's representations are designed to be consumed by a downstream agent. Google's Personal Health Agent demonstration explicitly forbids emitting raw numbers or boolean flags. The model outputs qualitative physiological state interpretations - "your body is in recovery mode" rather than "HRV: 47ms." This is not a dashboard backend. It is a coaching substrate.

The Two-Brain Architecture

The coaching wearable that actually changes behavior needs two distinct intelligence layers:

• Brain 1: The physiological foundation model - understands your body's temporal patterns, detects state transitions, predicts receptivity windows, and surfaces the when. SensorFM is the proof that this can be built from consumer-grade sensors at scale.
• Brain 2: The conversational model - understands coaching frameworks, motivational interviewing, behavioral activation, and surfaces the what. This is the LLM layer. It is a commodity.

The intervention timing research confirms this split. The Annual Review of Psychology's 2026 JITAI review identifies the core challenge: "individuals may not be able to engage with an intervention when they need it most in daily life." The solution is not better copy. It is better sensing of readiness.

When I built coaching systems at Meta, the hardest problem was never generating the right thing to say. GPT-4 class models handle that. The hard problem was always: does the user want to hear it right now? That question requires physiological context that no language model can answer from text alone.

Why This Is Hard to Replicate

SensorFM's moat is the pretraining corpus. One trillion minutes from five million people. That dataset came from Google's Fitbit installed base - the largest continuous physiological recording infrastructure ever assembled. No startup can replicate it. Apple has a comparable sensor fleet but has not published a foundation model over it. Samsung and Garmin have the hardware but not the ML research depth.

The training data is the defensibility. The architecture will be published and reproduced. The weights are proprietary. And the corpus required to match them sits behind exactly three consumer wearable ecosystems.

For independent builders - and I am one - this means the ambient AI coaching wearable cannot be fully vertical anymore. You either build on top of SensorFM (when Google opens API access), license a comparable physiological foundation model, or accept that your intervention timing will be worse than Google's by exactly the margin that a trillion-minute pretraining corpus provides.

Where This Heads

The two-brain architecture reframes the competitive map. The LLM is table stakes. The physiological model is the moat. Google just proved that wearable sensor data follows the same scaling laws as language - more data, more tasks, better transfer.

The ambient AI coaching wearable that wins daily use will be the one whose physiological brain knows you well enough to never interrupt at the wrong moment. SensorFM is the first proof that consumer-grade sensors contain enough signal to build that brain. The race to assemble the second trillion minutes has started.