Location-aware computing

Geospatial signal is one of the richest features a system can use — and one of the most dangerous to mishandle. The invention here is the discipline of using location as input without it becoming a dossier.

Location data is the single most valuable feature and the single most dangerous feature in almost every consumer system. It predicts intent, availability, fit, and risk better than any purchased dataset. It also reconstructs a person's life with chilling fidelity when kept in the wrong shape.

The discipline the 2014 habit-index patent proposed — append location to a tokenized profile, compute a scalar, discard the trajectory — is not the only answer, but it's the right shape of answer. Keep the aggregate, throw away the trail. Compute on-device, emit a differentially-private signal. The clever part is proving your signal is calibrated without exposing the input.

This is also the axis where the shift from "one dashboard per human" to "one agent per human" cuts hardest. An agent with live location can be asked to make decisions ("reserve a parking spot near where I'm heading") in ways that would be creepy coming from an app. The difference is whose permission envelope the decision lives in, and whether the agent is accountable to the user instead of the platform.