LMM Technologies

Sequence models for human movement

Models that know the domain, not the words about the domain.

LMM Technologies is a small research company. We build sequence models for human movement: models trained on motion and prompted with motion. Joint positions come off ordinary video, and the model returns a continuation. No caption appears anywhere in the loop.

The bet is that movement carries structure the way text does, and that a model trained directly on it learns that structure because nothing less survives the objective. A model trained on descriptions of motion knows the language of motion. A model trained on motion has no notation to hide behind.

Today that means a forecaster that watches a few seconds of a person moving and streams what comes next. Alongside it we are working on a written form for movement, so that motion can be recorded, searched, and compared without keeping the video.

The first place this matters is remote musculoskeletal care, where a clinician needs to see how someone moves without the footage ever leaving the room. Where the work stands.

Writing