I worked at Carl Zeiss for two years, selling consumer optics across six states. This is one of those assignments that ends up mattering more, professionally, than it was supposed to. The reason is that Carl Zeiss, the company, has spent over a century publishing in an area that the modern computer-vision field mostly does not read. Their internal technical literature contains a genuinely impressive inheritance of thinking about how humans actually see, why lenses do the specific things they do, and what it would mean to build an artificial visual system that respected these constraints rather than abstracting past them.

This notebook is a three-month reading through that literature. I started it because I kept noticing, in my AI evaluation work, that problems I was helping diagnose had specific parallels in Zeiss's published optics research from the 1920s, the 1960s, and the 2000s. The parallels are not aesthetic. They are structural. I wanted to see how deep they ran.

The short answer is: deeper than I expected. Deep enough that I think there is a book-length argument here that I am not going to be the one to write, but that someone should, because the argument is true and it is not being made.

The three main threads

The Zeiss literature, read carefully, carries three threads that I think should be part of the contemporary conversation on computer vision.

First: perception is not pixel-reading. Zeiss's work on how human vision actually functions — as distinct from how lenses and sensors function — is unusually sophisticated. The eye is not a camera. The eye is a sampler that builds a coherent visual experience from an enormous number of partial glances, reconstructions, and prior expectations. Any artificial system that treats vision as "read the pixels, produce the answer" has already made a philosophical error that Zeiss identified and articulated clearly several decades ago.

Second: lens character is information. Every lens has specific character — how it renders bokeh, how it handles flare, what its chromatic aberration looks like. The field, led by Zeiss and a few contemporaries, understood that this character was not noise. It was the mechanism by which the image became legible as an image. A computer-vision system trained on images without any model of lens character is, in effect, trained on images stripped of one of the signals that made them meaningful.

Third: the observer matters. Zeiss's work on consumer optics, particularly on eyeglasses and on cinema lenses, took seriously that different observers need different optical solutions. There is no single "correct" rendering of a scene. There are renderings that serve particular observers for particular purposes. This observer-specificity is, I think, one of the most underrated facts in the current AI discourse, which is still largely organized around the idea of a universal evaluator.

Why I'm genuinely excited about this material. Most of this literature was published in German, in technical journals aimed at working opticians and engineers. It was never translated into the vocabulary of computer science. Reading it three decades later, with the vocabulary of a computer-vision practitioner, it feels like finding a library that was closed before the building next door was built, and discovering that it already contains half the answers the neighbors are struggling to find.

The full notebook walks through each of these three threads with specific sources, worked examples, and...