That comic lost some meaning in recent years, you could put out some decently accurate bird detection model by following a 30 mins YouTube video, it was even one of the first exercice in some ML youtube courses (can't remember the channel).

To be fair, that comic was published around ten years ago. I'd say the example it used to make its point was reasonably accurate.

Yes for sure and the general idea is timeless anyway.

But it still retains some meaning. Using a glob pattern to wildcard files with the same name vs. running an image entity classifier across 1 to n images in potentially recursive directories. In a command line utility for displaying files.

Maybe more in the sense of the Unix philosophy: The detection should happen in a separate tool of which the output filenames can then be piped into lsix.

well sure. But if you were building this for actual usage, you'd have a separate cron-style tool that would monitor your picture folder (and its subdirectories) for new images. Feed aforementioned image into Llava or any number of image classifier APIs, embed classification information as IPTC data into the image file. Then give the CLI flag for the search tool the ability to do partial word search on metadata within images.

Jeremy Howard references this xkcd in the very beginning of his stable diffusion course[0], and builds such a detector in about 10 minutes.

0. https://course.fast.ai/Lessons/lesson9.html