What the Camera Sees
A meditation on perception, and the long road from pixels to meaning.
A camera does not see anything. It registers light. The seeing — the noticing, the recognition, the care — happens somewhere downstream, in tissue or in code, and that is where things get interesting.
I've spent enough years working on vision models that the gap between registering and seeing has stopped feeling like a technical problem and started feeling like a philosophical one. The pixels are easy. The meaning is the hard part.
A small experiment
Take any trained segmentation model. Show it a photograph of an empty kitchen. Ask it what is there. It will give you a clean list — counter, refrigerator, window, floor — and that list will be correct, and it will also be wrong in some way you can't quite articulate.
Because what the photograph contains is not just objects. It is also: the hour of the morning, the residue of last night, the imminence of breakfast, the absence of someone who used to be there. The model doesn't know about any of that, and we don't yet know how to ask it to.