The Backrooms, and Parallels to Our "Almost-Things"

My latest obsession has regrettably been The Backrooms movie. Like a moth crashing into a flourescent lightbulb, I am gravitated towards the amount of technical craft (and amount of yellow) shown on screen. The movie also holds a mesmerizing narrative where, although the characters wrestle with their own complexities, the shining star is this cosmic horror space disguised with the almost-familiar fluorescent lighting, objects “clipping” outside of physical bounds, and the presence of a moist interior.

If you’re not entirely familiar with the Backrooms’ cultural significance, it is worth the deep dive. Many during the COVID pandemic gravitated to fan-generated content for a number of reasons, one of the most frequent being that these stories and spaces captured nostalgia in an uncanny way. This space looks familiar… yet something feels eerily, terribly wrong. Chiwetel Ejiofor’s character chillingly describes this feeling in the movie:

“Imagine describing a dog to someone who has never seen one, and then asking them to draw it. They might get some things right, but there’s no way they get everything right?” -Clark

The devil’s in the details

Much of the glitchiness explored in the Backrooms lore always finds reference to something in the real world, but without the 1:1 resemblance. Almost as if trying to render a half-remembered dream. Some things right, but not everything right.

There’s obviously a place in this story where the allegories of the Backrooms can be compared to the characters’ traumas or manifestations of reflection. But as I found, the spotlight isn’t on the characters — it’s on the endless, mysterious chasm. And when interpreted that way, there are many more parallels to identify, but for the sake of time on this topic let me just share one parallel about distortions of reality.

There’s an undeniable view that much of this restructuring looks like the AI imagery we’ve seemingly let “grow on us”. We’ve been asking machines to draw the dog for a very long time:

The most compelling example of our systems trying to materialize from what exists is something you can try at home yourself. There are examples across socials showing what happens when you prompt for the exact image in-place 100+ times, and the results are a bit jarring. Where did the image model get this representation from? What was determining its “thought process” for the next image? WHY did it create this?

What is it like to imagine?

Latent feature space is the territory our machines inhabit today: it is a coordinate system with billions of axes (more than the 3 we visualize), where everything the model has absorbed sits at some address, and everything between two addresses is a place that never existed but plausibly could. It is much like the way we map neural networks onto the biological architecture of the human brain, except randomness is so much at play here that we cannot seem to determine what happens outside of a feature level. So our visual analogy remains a smaller black box.

Each output is a projection, billions of dimensions folded flat so a human visual system can perceive it as such, and what comes back always has the same quality: corridors of “almost-things.”

In 1974 Thomas Nagel wrote an essay titled “What Is It Like to Be a Bat?” Knowing everything about echolocation, he argued, gets you nothing of the experience of it. Some vantage points aren’t closed to us for lack of data, we’re just the wrong kind of thing to stand there. While not the primary goal, The Backrooms provides a modern-day commentary on our greatest primal fear during the Information Era: not something that can jump-scare us, but something we cannot understand at all.

• The Backrooms (Found Footage) — Kane Pixels, 2022. The nine-minute short the movie grew out of, made by a then-16-year-old in Blender.
• The Backrooms — Know Your Meme. The full lore rabbit hole, from the original 4chan post onward.
• What Is It Like to Be a Bat? — Thomas Nagel, 1974.
• Latent Diffusion Models tutorial — NeurIPS 2023. Source of the pipeline diagram above, and a good on-ramp if you want the math.
• Harold Cohen and AARON: A 40-Year Collaboration — Computer History Museum.
• Will Smith Eating Spaghetti test — Wikipedia. Yes, it has its own article now.