The curator's reviewThe 2015 original nightmare machine, still running. Feed it any photo and a convolutional network amplifies whatever it thinks it sees, which is usually dogs. Dogs in the clouds. Dogs in your soup.
Before AI made pretty pictures, it made this. In 2015 Google engineer Alexander Mordvintsev turned a convolutional network inside out: instead of asking what does this image contain, he asked the network to amplify whatever it thought it saw, feeding the result back until faint hunches became hallucinations. The networks of that era had seen a lot of ImageNet, and ImageNet contains a lot of dogs. Hence: dogs. Dogs in clouds. Dogs in soup. Dog-slug chimeras erupting from your vacation photos, garnished with pagodas and eyes.
Deep Dream Generator is the original tooling around that discovery, still running a decade later. Upload anything, pick a style and depth, and the fever dream comes back - iridescent, fractal, unmistakably 2015. The look was everywhere for a year or two (album covers, music videos, one memorable psychedelic era of the internet) and then vanished as suddenly as it arrived, which makes the site a living museum of a dead aesthetic.
It matters beyond the spectacle: Deep Dream was the first time the public saw inside a neural network - proof that these things had internal imagery at all, that recognition ran deep enough to be run in reverse. Feature visualization research descends from it directly.
Historically important, permanently unsettling, and the reason an entire generation cannot look at broccoli calmly. Visit once a year, the way you visit a lava lamp: to confirm it is still going, and that it still sees the dogs.
