LLMs treat places as frozen words, but Google Research's ME-POIs framework layers in anonymized mobility patterns, arrival times, stay durations, daily foot traffic, to teach models a place's functional rhythm, not just its metadata. The payoff is concrete: up to 81.9% relative gain predicting visit intent, 75.1% for price level, and 24.7% for busyness, across unseen places. A self-supervised approach that makes embeddings encode when a place is alive, not merely what it is.
Wearables already capture floods of physiological data, but turning it into trustworthy biomarkers is the real bottleneck. Google's Biomarker Discovery Framework tackles that with a human-supervised multi-agent loop that generates hypotheses, runs adversarial validation, and grounds results in the literature before anything becomes a candidate. Across three cohorts (9,279 participant-observations), it recovered known clinical signals and surfaced convergent markers across mental-health and metabolic datasets. The design keeps statistical rigor front and center, addressing exactly where naive LLM pipelines leak spurious correlations.