The Signal Was Already in the Stream

A model smaller than your autocorrect

Google Research published GlucoFM last week, a foundation model for continuous glucose monitors, the small sensors that sit under the skin and take a reading every few minutes. The headline number is not how big it is. It is how small: GlucoFM runs on 0.72 million parameters, according to the arXiv paper, a rounding error next to the language models everyone argues about. It was pre-trained on 109,066 hours of unlabelled glucose data from 477 participant records, and on a single day's window it outscored every CGM model that came before it, including GluFormer, across 14 evaluations on seven clinical tasks: diabetes risk, insulin resistance, beta-cell dysfunction, hyperlipidemia, hypoglycemia, obesity and glucotype.

Google Research is careful to call it a research prototype, not a diagnostic, and it has not been cleared by any regulator. Read that honestly. This is not a product you can buy. It is a pattern you can learn from, and the pattern is the point of this piece.

Slow state, fast events

Here is what the team actually did, because the architecture is the lesson. A glucose trace is not one signal. It is two, braided together: a slow baseline that drifts over hours and days and says something true about your metabolism, and short sharp deviations that say something about breakfast, a run, or a sensor glitch. Every model before this one read the trace as a single line and tried to learn both at once.

GlucoFM splits them. A causal Gaussian filter separates the slow state stream from the fast event stream before the transformer ever sees the data. The recordings are aligned to a 24-hour clock so the model knows that a spike at breakfast means something different from a spike at three in the morning, and the gaps where a sensor dropped out are kept as a mask instead of being smoothed over. Then it is trained with JEPA-style latent objectives, predicting hidden patterns rather than raw numbers. Google Research reports that the two-stream design beat the single-stream versions in every ablation, and the event-only version did worst of all.

Read that last line twice. The spikes alone, the dramatic part of the chart, were the least informative view. The slow baseline underneath them carried the diagnosis.

Results that transfer

The numbers are modest in size and large in meaning. On the paper's task average, GlucoFM scored 58.8 PR-AUC against 54.7 for the strongest baseline retrained on the same corpus, per Google Research. It predicted a two-hour post-meal glucose curve with a mean error of 21.88 mg/dL where the best competitor managed 22.90. It transferred across datasets it had never trained on, winning 11 of 12 cross-cohort tests. And it kept its lead when the labelled data was cut to one patient per class.

That few-shot result is the commercial story hiding inside the medical one. Labels, the confirmed diagnoses that let you train a supervised model, are the expensive, slow, human part. The raw stream is cheap and already flowing. GlucoFM learns from the cheap thing so it needs almost none of the expensive thing. No new blood test, no new sensor. Everything it knows, it learned from data that was already being collected.

Every business already wears a monitor

This is where we stop talking about glucose. Every operating business is wearing a continuous monitor and reading it like a fingerstick. Your CRM is a glucose trace. So is your inbox, your ad account, your cash position, your calendar. They record something every few minutes, and most operators look at a single reading, panic or relax, and move on. The Monday sales spike gets a Slack message. The slow six-month drift in reply rates, the one that actually predicts whether the quarter closes, gets nothing, because nobody separated it from the noise.

The GlucoFM move is the CYSTEMS move: decompose the stream before you interpret it. Pull out the slow state, the baseline health of a pipeline or a channel, and treat the fast events, the launches and the outages and the one viral post, as deviations from it rather than as the story itself. Do that and you need far fewer labelled outcomes to know where you are heading, because the shape of the baseline is already telling you.

We built our audit practice on exactly this premise. The free audit reads the surface. The deep one pulls the baseline out of the data a client already has, before a euro of new spend, and forecasts where it goes. Prediction from existing streams, not new instrumentation. Same shape as Google's, smaller universe.

The operator takeaway

Three things to carry out of this. First, the most valuable model in your business may be tiny, if the representation is right; parameter count is vanity, decomposition is leverage. Second, your labels are the bottleneck, not your data, so build the thing that learns from the unlabelled stream you already have. Third, the spikes are the least informative part of your chart. Stop managing to them.

Google taught a model with fewer parameters than a phone keyboard to read a person's metabolism from a sensor they were already wearing. The signal was in the stream the whole time. Yours is too. Build the Cystem. Watch It Work.