Google has spent the last few years teaching computers how to see, understand, and appreciate our world. It's an important goal that the search giant hopes will allow programs to classify images just by "looking" at them, but there's one problem — to computers, our world looks like a horrible nightmare. The first images from Google's project, shown in June, reveal an organic hellscape inhabited by slug monsters and disembodied tentacles. The pictures show leaves that become two-headed birds, horses that grow dogs' heads from every inch of their skin, and humans whose bodies erupt in furry growths. To Google's computers, every surface of our world is covered in eyes.
The artificial neural networks used by Google work out what's in an image by analyzing layers until they finally decide on what it shows. "For example," Google engineers wrote in June, "the first layer maybe looks for edges or corners. Intermediate layers interpret the basic features to look for overall shapes or components, like a door or a leaf. The final few layers assemble those into complete interpretations — these neurons activate in response to very complex things such as entire buildings or trees." But while Google's code might be able to ascertain that a tree is indeed a tree, the image it builds from the original picture looks less like a harmless tree, and more like a glowing space temple for a cult of silently screaming bipedal monsters.
And now that code's loose. Instead of doing the sensible thing and destroying it, Google put everything you need to create your own neural network art on programming repository GitHub, calling it Deep Dream. Already it's capable of making art out of almost nothing. By plugging in a piece of random noise to a network tuned to look for specific places, the program can produce snapshots of science fiction worlds, building bridges, pagodas, and impossible spires out of wild colors as part of its attempts to find meaning in the randomness it's shown. Its creators say it'll be "interesting" to see what kind of art people can create with it.
"Interesting" is one word. "Terrifying" might be more accurate, especially when the code doesn't just work on still images — it can now be used on video too. Perhaps the scariest result so far has come from applying Deep Dream to Fear and Loathing in Las Vegas. The movie's acid scene, already a pretty convincing case not to over-indulge in high-grade hallucinogens, becomes a vision of hell itself when viewed through the computer's eyes. Faces twist and morph as walls melt, blurring the lines between animate and inanimate, organic and inorganic. A man groans into the camera as a jet black eye rises from his throat. In the bathroom, Dr. Gonzo drops acid, his shoulders and arms rapidly turning into warped chickens. As he leaves, the floor becomes a giant salmon.
The visual effect is Deep Dream trying to find patterns it recognizes, trying to get a bead on exactly what it's being shown, but the result is mightily unsettling. It makes me feel like all that money spent by Elon Musk on making sure we don't invent murderous AI is a good idea. If I was a computer, and I too discovered I'd been created by a race of rapidly metamorphosing pustule monsters, I too would try to destroy the planet.
