Comment by simonw

I'd love to know how they are doing this! Are they running an on-device CLIP embedding model of some sort? Could they even be shipping a multi-modal LLM like Phi-3 Vision?