The Last Word On Nothing

If you could record your dreams, would you ever be able to stop watching? 

That’s a central plot point in Until the End of the World, Wim Wenders’ three- to five-hour arthouse flop/work of staggering genius (depending on which cut you watch) from 1991. He threw cyberpunk, detective noir, and road movie into a blender with a civilisational apocalypse. It cost him $20 million to make and did not earn it back. Love it or hate it, this movie is canon, in part because it was prescient about the technological future to the point of being oracular.

In the near future as he imagined it, a scientist has invented a machine that lets you re-watch your dreams – including all the dreams you never knew you had (most of them become irretrievable within minutes of waking up). People become instantly bewitched by the technology, some sliding into “a disease of images” in a way that is immediately legible from 2020s discourse on phone and social media addiction. They’re simply unable to tear themselves away from the beautiful, narcissistic universe this technology has opened for them, and are no longer able to engage with reality. “How many times must I have dreamed this?” one character mumbles reverently, lost in the grainy footage of her childhood memories.

Wenders got a startling number of things right about the technological future. Ubiquitous wireless devices, search engines, voice-to-text, an early version of Google maps and retinal prostheses; the characters in the film are saturated in tech.

And now Will Heaven has written in MIT Technology Review about a company that has been able to reconstruct lost childhood memories using generative AI. The dream camera is nigh.

The Synthetic Memories project, run out of a design studio in Barcelona, turns memories that were never captured on camera into tangible photos. It works a little bit the way a forensic artist illustrates a police sketch. A subject with a specific memory sits with an interviewer and recounts the precise situation, their perspective within it, and its factual contents. Then a generative AI model is used to stitch those elements into an image. It’s an iterative process: the image is shown to the subject, who corrects what’s wrong with it until it is right. 

Doing this, they were able to recreate a photo that never existed, from the memory of an 84-year-old woman seeing her father when she was 6. It was, Heaven writes, “a glimpse of him through the small window of his cell, where he was locked up for opposing the dictatorship of Francisco Franco.” They just took a tweezer and pulled it from her memory.

Intriguingly, the subjects were only happy with the results of the studio’s alchemy if their memory-images were ‘painted’ using older versions of generative image models (like DALL-E and Stable Diffusion) rather than with newer ones.

That’s because the old ones could “produce images that are glitchy, with warped faces and twisted bodies,” Heaven writes. “But … [images created with] the latest version of Midjourney (another generative image model that can create more detailed images)… did not click with people so well.”

“If you make something super-realistic, people focus on details that were not there,” says Garcia. “If it’s blurry, the concept comes across better. Memories are a bit like dreams. They do not behave like photographs, with forensic details. You do not remember if the chair was red or green. You simply remember that there was a chair.” 

“For us, the glitches are a feature,” the researcher told Tech Review. That was also the most incisive insight in Until the End of the World – Wenders’ intuition that any future technology that looks inward would need to move away from realism.

He went full Kubrick in his fussiness, endlessly deliberating on how to depict a person’s inner world. Wenders and his little AV team stayed in the editing suite for almost two months, even sleeping there, in the obsessive quest to produce the footage that is meant to look like it was extracted from brains.

“Some of [the footage] was my own 8-millimeter stuff of my own childhood because I felt that the dreams that appear in the film, I was getting so personal I felt I needed to use my own,” Wenders said in a later interview. He “terrorised” the technicians because even as they had been refining their sophisticated equipment to become capable of ever-greater feats of realistic depiction, Wenders “did everything in the book to sort of pull the images apart so that something else would come out. Those technicians were so glad when we left because we had basically ruined all their equipment.”

The result of all this effort render the inner dream landscape in the visual language of a Vermeer. I can’t improve on The Guardian description of the emotional effect: “The lengthy sequences of pixelated, almost inscrutable images that project ecstatic glee and existential frustration on to the characters are breathtaking.”

Perhaps predictably, however, the memory camera isn’t eliciting quite such rapture online. One guy on LinkedIn (I know, I will go wash my mouth out with soap) announced that we were seeing the end of believable documentary evidence. “Looks like “Minority Report” is getting real,” said another. “Only a matter of time until some quack with a PhD offers this as a service to law enforcement,” said a third.

I couldn’t help noticing that this is the same valence of response that has greeted deepfakes. A litany of warnings has told us that soon we will no longer be able to believe our eyes. I’ve lost count of the number of variations on “seeing is no longer believing” I’ve seen in headlines. Photography, evidence, documentation, it’s all dead, they say, destroyed by this new technology. But I wonder if that’s backwards. Photography was never a perfect medium for representing reality. Long before deepfakes were even a twinkle in Photoshop’s eye, two Edwardian-era teenage dirtbags caused Arthur Conan Doyle to lose his entire mind with doctored photos of fairies in his garden.

The uncomfortable truth, I think, is that what photography shows has always been more art than science – there is a whole world outside any shot that the photographer chose not to show, privileging only what is in the shot. The editorial decisions that go into the image have been invisible to us for so long that their revelation by deepfakes has been shocking. Photography has never been reality (see also: “reality” shows. What distinguishes a credible documentary from a garbage truck reality show isn’t the verisimilitude of the images but the reputation of the filmmaker). Future historians might look at our unquestioning belief in objective images during the 20th century as a curious artifact. Or they might be in their own dream loops, the only history they wish to record being their own.