Thread by @michaelzlin

In case you think cameras can detect cars or people ahead of them through the glare of the sun, or that AI has learned to react to ambiguous situations with the, well, intelligence of humans, read this

It's a frightening and sad story (a grandmother was killed) and I'm surprised to find out about it only now.

There's clear evidence from the Tesla's own cameras that no kind of image processing can beat glare (and that FSD lacks all sense of caution)
bloomberg.com/features/2025-…

Glare happens when light rays from the sun get scattered by water or dust in the air at every lit location to new directions. That means air between an object and the camera can scatter sunlight toward the eye or camera.

When the sun is low, even small scattering angles are enough to redirect lots of light toward a camera, from the air between a person and the camera. This means the photon count from glare will exceed the photon count from the (dark side) of the person.

I drew it out below.

This is the actual physical refutation of Musk's claim to have solved the problem with photon counting. Photon counting may help model the spatial extent of glare, but if there is a relatively dark object in there (say emitting 1% of the photons vs glare) it's not going to help.

That's where LIDAR helps. Sure, humans don't have LIDAR either, but the marketing and regulatory claim for developing automated/robotic driving is that the robots can do it better.

Also human brains compensate for lack of LIDAR by exercising caution when we can't see ahead...

and FSD obviously hadn't learned that, as apparent in the video.

And I don't even think in humans it's learned so much as instinctive: We have an innate fear not to advance at deadly speeds into locations we can't see.