The next camera race will be about understanding

The demo that convinced me happened at an art exhibition I'd built myself.

We couldn't afford a team of docents, so I put together a side project to stand in for one.

There was no prompt box and nothing to type.

You pointed your phone at something and it told you what you were looking at.

Then the show closed, and people kept using it anyway.

They pointed it at buildings, flowers, their kids' toys, posters on the street, none of which had anything to do with the exhibition.

Without anyone asking them to, they had decided a camera should be able to explain the world back to them.

That expectation is now the industry's to deliver on, and it arrives just as AI shifts from a training problem to an inference one, with more of the real work happening on-device while someone points a phone at an uncooperative real world.

A model needs different things from a frame than your eye does: edges to lock onto, and structure it can recover when the shot is full of glare or blur.

The color and contrast that make a photo look good are, to the model, mostly in the way.

The sensor you'd want for a human viewer and the one you'd want for a model call for different designs.

The sensor has become a perception problem, and the people building real-time vision systems now care about it the way photographers once did.

Where the sensor came from Eric Fossum's CMOS active-pixel sensor, developed in the early 1990s, is why a capable camera now sits in nearly every pocket.

His later work heads somewhere else entirely: the Quanta Image Sensor, which counts individual photons.