From Intention to Action

The most interesting problem in human–AI interaction isn't making the machine smarter. It's closing the gap between what a person means and what the system understands them to mean. We call that the alignment gap, and it's where most of the friction in everyday computing actually lives.

Think about how you interact with a device today. You form an intention, translate it into a physical action — a tap, a click, a spoken command — and the system reacts to the action after it happens. Every step in that chain adds latency, ambiguity, and effort. The richer and more capable AI becomes, the more this old input model strains under it.

A wrist that reads neuromuscular signals changes the shape of the problem. Instead of waiting for a completed action, the system can read the command as it forms — and, paired with AR, it can show you how it understood you, so you can correct course in real time. The interaction stops being a sequence of instructions and starts becoming a conversation that converges.

Our research roadmap is built around three pieces. The first is the Large MUAP Model — turning raw EMG into a shared "alphabet" of canonical neuromuscular patterns, so intent can be represented and predicted the way a language model handles tokens. The second is cardiac-locomotor coupling — fusing EMG, IMU, and PPG so the system reads not just what you're doing but your engagement and state. The third is AR as the feedback channel, the surface where the machine's understanding becomes visible and correctable.

To be clear about what's shipping and what isn't: today, our hardware reads wrist EMG reliably and turns it into action, and it's been doing so in a consumer product for years. The Large MUAP Model and the full alignment loop are active research — a direction, not a finished feature. We think the honest version of a roadmap is more useful than an inflated one.

What grounds all of it is that we earned the right to the ambition the hard way. Reading EMG at the wrist, reliably, across thousands of different people and physiologies, is genuinely difficult — and we've been doing it commercially for years. The expertise that took generalizes: the harder problem we already solved is what makes the bigger one credible.