Essay / 18 Oct 2026
Notes on Looking
On attention, apertures, and the difference between a pixel and a thing.
The first camera I owned had a light meter that was better than my eye. It did not know what I was looking at. It only knew how much light was arriving, and it was almost always right about that. I spent a summer learning to disagree with it in useful ways.
A vision model is a more elaborate light meter. It receives an array of numbers and returns a description of the world. Most of the time the description is good enough to act on. The interesting failures happen at the edge of that "enough": a landing marker seen at the wrong angle, a bounding box drawn around the shadow instead of the object, a sentence about velocity that sounds fluent and is not grounded in the frames.
The meter and the eye
I learned to look at images twice. First as a photographer: where is the light coming from, what is being excluded by the frame, what happens if I wait four seconds. Then as an engineer: what is the input distribution, what does the label actually mean, what would make this example hard.
The two habits disagree productively. The photographer wants the frame to hold a feeling. The engineer wants the frame to hold a measurement. A good dataset needs both: images that are specific enough to be real, and labels that are precise enough to be wrong in a way you can detect.
The annotation is a claim
In the landing project, 367 of 1,999 bounding boxes were flipped. The detector was not hallucinating. It had learned the annotator's mistake with great fidelity, because the mistake was consistent. This is the part that is easy to forget: a label is not a fact. It is a claim, made by a person, under a convention, at a particular time.
Once I started treating annotations as claims, the audit became obvious. Check symmetry. Check whether the box contains the thing or the thing's reflection. Check whether two classes are being used interchangeably. Check the frames where the annotator was tired. The model's confidence is not evidence about the world; it is evidence about the training set.
What the model cannot say
In the Qwen2.5-VL study, position was easy to decode. A linear probe could recover it with an R² of 0.97, and language grounding was perfect. Velocity was different. The information was there, but it lived in an entangled manifold the model did not use by default. Asking the model to reason about speed helped, but it helped the way a hint helps a student who has the answer somewhere and no procedure for reaching it.
This is not a story about a model being stupid. It is a story about representation being selective. The model had learned to say the right kind of thing about motion without building the thing that motion requires: a stable, separable estimate that can be used in a new situation.
A practice
Looking closely is not a personality trait. It is a set of small procedures: compare the frame to the thing, write down the uncertainty, look again after a day, ask what would change your mind.
The notes on this site are those procedures made public. They are not a feed. They are closer to a lab notebook: a place to keep the gap between a metric and a meaning visible, so that the next time the model is confident, I remember to check what it is confident about.
That is the whole practice. The rest is equipment.