Nobody currently measures local AI adoption. Cloud providers publish usage statistics; the local ecosystem's footprint must be inferred from download counts — signals of scale, but silent on the questions that matter: who runs models locally, which models, for what tasks, on what hardware, and why.
We are designing the first longitudinal study of local AI adoption, with methodology published before data collection begins. Design summary:
Component 1 — Stratified survey panel
Recurring quarterly survey sampling three strata: general computing users, professional knowledge workers in privacy-constrained fields (legal, healthcare, education, government), and self-identified local-AI users recruited from practitioner communities. Instruments cover hardware owned, models used, task mix, frequency, motivations (privacy, cost, reliability, offline need), and abandonment reasons. Prior community research suggests reliability, local control, and privacy dominate motivation — our design tests whether that holds beyond the enthusiast population.
Component 2 — Opt-in telemetry cohort
Volunteer participants running instrumented open tooling contribute anonymized, on-device-aggregated usage data: model families invoked, task categories (locally classified — raw content never leaves the device), session energy draw where measurable. The telemetry design will itself be attestation-protected — the study eats our own cooking.
Component 3 — Model-ecosystem observatory
Quarterly public tracker of the local model landscape: releases, download-weighted popularity, capability-per-parameter trends, and lag-to-frontier measurements on non-saturated benchmarks.
Outputs
An annual State of Local AI report, an open dataset, and peer-reviewed submissions. All instruments, code, and aggregation methods published openly.
Participate. We are recruiting survey panelists, telemetry volunteers, and academic collaborators in survey methodology and HCI. → support@yforest.ai, subject "Adoption study."