Deciding to Leave: Four Years of Building

I'm leaving Magic Leap. After four and a half years building perception systems for mixed reality, it's time for the next chapter.

Why Now?

A few reasons stacked up at once. The learning curve flattened: years one through three were steep, year four and beyond have been execution - still valuable, but the growth rate slowed. The company changed too. Enterprise pivot, layoffs, leadership changes; it's not the same company I joined. In December I wrote that V2 was on track and that I had made no decision about staying. The second half resolved itself. The first half I would now soften, for reasons that show up under regrets below. There's also opportunity cost, because at some point staying is a decision to not do something else. And then an offer emerged that checked every box.

What I'm Proud Of

Shipping, first. Magic Leap One is a real product that real people use, and I led the perception system that makes it work. The synthetic data team I founded now generates millions of training images and has become essential infrastructure. I filed four patents along the way, technical contributions that will outlast my tenure. The part I'll carry longest is the team: 45+ people who grew in their careers, several of whom now lead their own teams. And I arrived knowing nothing about AR. I leave having architected state-of-the-art perception systems.

What I Regret

Not pushing harder on calibration earlier. We scrambled before ship, and earlier investment would have helped. Some hiring decisions: a few people I championed didn't work out, and I should have assessed fit more carefully. US-Israel coordination could have been better - I should have visited more. I wrote last June that quarterly trips in both directions were worth every dollar. They were, and they were also not enough. And some of the shortcuts we took for V1 limited our V2 options. I should have held the line.

Lessons Learned

Hardware timelines are unforgiving. Decisions made in year one determine what's possible in year four, so think further ahead than feels reasonable. I wrote in 2016 that the compute architecture we chose then would set the ceiling for five years. That one held. Predictions about constraints usually do. Data is strategy: the team that generates and curates training data faster wins, which means investing in data infrastructure before it feels urgent. Shipping teaches what planning can't - all the analysis in the world doesn't compare to the crucible of launch. And culture beats process. Good people with shared values figure out the right process; process can't fix cultural problems.

The Next Chapter

I'm joining Meta's Reality Labs to lead Visual Positioning Service development. Different scale, different challenges, same underlying problem of helping computers understand the physical world. Magic Leap taught me what's possible. Meta will teach me what scale looks like.

Grateful to everyone who made these four years what they were. The work continues, just from a different seat.