Embodied Trust

What passes for safety today is mostly a patch — a rule layer bolted on after the system is built, a checklist added before the product ships. Engineers predict what the AI will encounter, anticipate how users will behave, model the edge cases they can imagine, and call it done.

I've spent enough time in robotics, computer vision, and safety research to know this approach is wrong. And I spent years at TikTok watching it fail at scale.

What social media taught me about AI safety.

At TikTok, I led Minor Safety — the work of protecting children from a platform being actively exploited by people trying to reach them. Not hypothetical bad actors. Real ones, working in real time, reverse-engineering algorithms to do things the engineers never imagined anyone would try. Every time we patched a gap, someone found a new one.

Human behavior at scale is fundamentally unpredictable, and any safety framework built on predictions will eventually be surprised by reality.

Robots are next. And the stakes are physical.

We have two choices.

We can deploy expensive robots slowly, gather limited data from a narrow slice of users in controlled settings, and hope we've anticipated the important failure modes before something goes wrong in the real world.

Or we can deploy cheap, fast, breakable — but responsible — robots to everyone.
simulation → Robot as a Service → Origami.
And watch what people actually do. The more we observe, the better we can train our systems to recognize and mitigate misuse before it escalates.

We're not trying to sell robots. We're trying to earn public trust before humanoids are everywhere. Because by the time they are everywhere, it will be too late to start.

Robots & kids — is it safe?

We work on World Models, Vision-Language-Action architectures, and Reinforcement Learning with humans in the loop — not to make robots more capable, but to make them safer when capability runs out. We don't want robots that perform well in the demo. We want robots that behave responsibly in the ten thousand situations nobody planned for.

Solving the data problem and Origami v1.0 aren't separate from this research. They're how the research gets done. The combination of our data platform and Origami will give us something no robotics lab currently has: a continuous, diverse, real-world stream of human-robot interaction at scale.

The goal isn't a better robot.

It's a generation of robots that humans can genuinely trust. Safe by design, not by patch. Companions that grow alongside people, rather than tools deployed at them.