cd ~/
RetrospectiveMetaVPSYear Review

One Year at Meta: VPS Progress Report

A year of building Visual Positioning Service: what shipped, what stalled, and what I would do differently.

Evyatar Bluzer
4 min read

18 months since I joined Meta, 12 since I moved onto VPS full time. End of year, so here is the honest ledger.

What We Shipped

The VPS developer preview is live in 15 countries, covers 500+ mapped locations, and is in the hands of 50+ developer partners, with integrations running in Quest experiences. On the technical side we hit sub-meter outdoor localization with under 1 second query latency, on a privacy-preserving architecture with a learned feature pipeline. The system now processes 10M+ images per day and serves 1M+ localization queries per day, built and run by a team that grew to 50 people.

What Went Well

Team building, mostly. We hired strong people, promoted good managers, and the culture held. Shipping velocity too: the MVP launched in 6 months and we iterated quarterly after that. Cross-functional partnership worked better than I expected - the Quest platform, privacy, and policy teams behaved like partners rather than gates. And the early architecture choices are paying off; they are what let us scale and iterate at this pace.

What Didn't

Indoor localization is behind plan. It is harder than we expected, and I underestimated it. I gave it three months in June, then a 2022 beta in September, and it appears again below as a 2022 launch. Three dates in six months. Mapping is still too expensive per location - the economics do not scale, and that is the number I stare at most. We hit the 500 locations I called in July and not the $500 per location that was supposed to arrive with them. The SDK needs polish; onboarding confuses new partners. And accuracy is inconsistent: great in some conditions, poor in others, and developers notice the second kind. The learned feature pipeline shipped with a 40% cut in failures from internal testing; "fewer failures on average" and "works in every condition" turned out to be different promises.

Year-one calls and their December gradesFive horizontal rows, one per call, each a line from the month the call was made to a dashed December line with a verdict at the right; the fourth row, the cost-per-location call, is highlighted.FebAprJunAugOctDecLearned features: 40% fewer failuresIndoor localization in three months500 mapped locations$500 per locationIndoor beta in 2022this postpartial: not everywheremissedhit: 500+ livemissedre-dated: 2022 launch
The year's calls as this post grades them: the July location count landed, the July cost target and the June indoor date did not, and indoor's September date has moved again; dates from the linked posts.

What did a year of VPS teach me?

Privacy is a design constraint. Treat it as a feature to bolt on later and you end up rearchitecting; every decision has to account for it from the start. Coverage beats accuracy - developers would rather have it work everywhere than perfectly anywhere, and I resisted that conclusion longer than I should have. Operations at scale is its own discipline: what works for 10 locations does not work for 500. And Meta's resources are real. The infrastructure, talent, and reach here enable things impossible elsewhere.

Looking to 2022

Four priorities: launch indoor localization, cut mapping cost by 10x, ship developer self-service tools, and integrate with Quest 3, the next hardware. The team goes from 50 to 80 - focused growth, not exponential. The technical bets I care about most are neural rendering for map compression, multi-modal fusion (WiFi + visual), and crowdsourced mapping at scale.

Personal Reflection

What energizes me: building systems that ship to millions, solving hard technical problems, growing talent. What drains me: meeting overhead, cross-org alignment, politics. The growth edge for next year is delegating more, thinking bigger, and influencing without authority - the last one especially.

Overall

VPS is real and growing, and the path to ubiquitous spatial computing looks clearer than it did a year ago. We are not there - indoor is a gap, rural areas are a gap, edge cases abound - but the foundation is solid. Year two is where the mapping economics either start working or they don't.

Comments