The #DEEPDREAM Trip Report

The biggest news of the month is that Google has developed an Artificial Intelligence that is able to paint. The artist Daniel Keller tried it out. In the process, he discovered two things: a Post-Internet art of hallucinogens and an answer to the question of why visual memes can so quickly go from mindblowing to lame.



Such fleeting profundity.... Why am SWIM (“Someone Who Isn’t Me”, in online drug forum vernacular) even writing about #deepdream this week? It’s soooo late June 2015, right? But SWIM won’t pretend to have felt anything but elation last month when they first read a relatively humble blog post about the recent work of a few Google employees who dauntlessly proclaimed a new artistic movement christened in the tradition of the Modern avant-garde: Inceptionism.

It looked neat and showed off Google’s work integrating neural network ‘AI’ into its services. Perhaps more essentially, it was another triumph on the inevitable march into posthumanity and a funky icon for Californian Ideological supremacy. These same kind of neural networks (already valiantly filtering Gmail spam) are deep learning how to SEE (!) and in the process, teaching us about the fundamental basis of perception and pattern recognition. Essentially, #deepdreams are the product of the inversion of the neural networks intended function (which is to identify and classify imagery) and make it output super trippy imagery instead.

Not to diminish it’s impact, the first images and videos of #deepdream SWIM saw felt genuinely shocking, awesome and beautiful. Bewildering alien imagery that seemed distinctly nonhuman in their aesthetic palette and unintuitive decision making, let alone the conceptual ramifications of their creation process. SWIM thought this was the 2015 artwork to beat after The Secret Lodge by Elda Oreto.


It was a welcome antidote to weeks of seemingly endless clickbait controversies and filter-bubbled debates over identity politics along with plenty of actual geopolitical crises (and we can assume subsequently increased revenue for fb and twitter.)

It speaks volumes that the only thing seemingly unifying in our feeds was the freaky outsider art of an artificial neural network endorsed by the world’s second biggest technology company

—The Singularity is gonna be siiiiick.

The original post “Inceptionism: Going Deeper With Neural Networks” was promptly misinterpreted by the media with a bunch of borderline batshit crazy headlines like “Google’s Dream Robot is Running Wild Across the Internet.” and “Google’s Machine-Mind Turns Pics Into Hallucinatory Nightmare Fuel” 

A few days later they made the code open-source….OMG. SWIM was quixotically committed to running the #deepdream code ASAP. This instinctively felt absolutely integral to SWIM’s practice. 


The install instructions were cheerfully advertised as requiring ‘as few dependencies as possible’. But to a code-illiterate artist like SWIM the prerequisites read like teen slang from an unimaginative Margaret Atwood dystopia: Open terminal, download Xcode, install Homebrew packet manager, ready your iPython notebook, get yer google protobufs and a ‘standard’ python science stack (and/or something called Anaconda or Enthought Canopy) and you’re ready to... compile?

SWIM knew they would be hopeless so they enlisted a patient genius hacker friend (who SWIM may have gone to high school with) to handhold them through the install. At one point hacker friend instructed SWIM to inscribe “brew tap homebrew/science; pip install NumPy SciPy Pillow; --Fresh -vd- Snappy” into the terminal window...and they weren’t even trolling.

Nothing seemed to work :-( SWIM’s interest and willpower were steadily waning.

In the meantime, competing New Media artists had already begun uploading the fruits of far more code-literate labor and it was only a matter of time now before this aesthletic territory was claimed and its visual resources exhausted.  

SWIM felt dumb and defeated but figured a day or two would bring an easy online app. Lo and behold! Some kid posted a basic web interface onto his site, but it was so popular that it was unintentionally DDOSed for days. 

But SWIM would be patient. 


The interface was simple enough: upload a picture onto the website and choose between two abstraction settings. One is essentially an impressionist Photoshop filter and the other generates those polycephalic Puppyslugs we’ve all grown to love/loathe. Sure the app lacked the flexibility of the original code and uses the standard dataset of images – but it was something. 

SWIM listens to Soundcloud and waits.

YES!!!!! The page finally loaded, it spits out a #deepdreamt rococo painting SWIM had downloaded from Tumblr! SWIM is Euphoric!!! Finally SWIM has their hands on #deepdream!!! Er well sorta…at least the ability to post pics on social media making it seem like SWIM has mastered the arcane install.  

SWIM decides to ignore all incoming messages asking how they’ve accomplished it. 

SWIM processed some dickpics, some install views and a peer’s dumb meme and posted a few. Most surpassed the 100-like threshold that SWIM counts as a ‘successful’ post. But alas, no eureka applications came to SWIM. SWIM wanted to believe in #deepdream but they were losing faith in themselves and the medium.

Three days passed, bringing better interfaces and more stable servers. Cue every new media artist, burner and their moms uploading thousands of #deepdream memes to their social media stacks: “Woah! Someone ran ‘Fear and Loathing’ through #deepdream and you won’t believe how freaky it looks!”

Pretty quickly it seemed unanimous amongst go-to contrarians that this thing was OVER. And only 72 hours after the first user-generated #deepdream image was posted. Kinda tragic.

SWIM had been considering using #deepdream in some work currently in production for a fair in Miami this December. But imagine how dated it will be by then? Hydro-transferred onto CNC-milled reliefs or not, this motif became incredibly gauche in an astonishingly short period of time. How can somebody produce timely internet-paced work around art fair application schedules and sea-freight logistics?


Changa vs #deepdream

Obviously SWIM isnt going to be the first to note #deepdream's similarity to hallucinogens. But in the however-many-years image-makers have tried to depict the qualia of drug experience (in a kaleidoscope of crazy zooms and fractals, lava lamps, colored gels and reverb pedals) this might be the closest approximation yet. iDosing is finally real, y’all!

The #deepdream visuals reminded SWIM of their recent experience with Changa, a relatively new mixture of DMT infused smokeable herbs which supposedly originated in the late noughties Australian festival scene (I mean, obviously – just say CHANGA to yourself in an Australian accent) and purchased with Bitcoin on the fabled darknet. Like #deepdream, it was super cool and visually entertaining but seemingly without any emotional or spiritual effects – and it lasted less than 2 minutes. Think pachinko parlor silk ribbon knots with embossed DNA, gothic windows and ‘figurative’ visuals, but mediated through a next-gen Oculus Rift.

DMT is the perfect spectacle-intense, no risk, TV-ad length trip for the on-the-go ADHD YUCCIE who ‘ain’t got time for that.’ It’s 2015 and in theory you could be reading this article and vaping darknet DMT while shopping for sofas at IKEA. Contrary to Terence Mckenna’s gushing about it being the God molecule,

it’s basically the Post-Internet art of hallucinogens — looks great, doesn’t take up too much of your braintime.

To return to the point, the main similarity between #deepdream and hallucinogens is how predictable and hollow their ‘deep lessons’ become after experiencing them a few times. Let’s #deepdream selfies, Mandelbrot sets, Munch’s Scream. Let’s #deepdream your boyfriend’s butthole, Contemporary Art Daily; say, have you been to the Van Gogh museum on Shrooms?


Novelty Decay

SWIM wondered if any existing theory could describe why something like #deepdream can go from mindblowing to lame so quickly? SWIM was reminded by a friend who was super into this 2008 theory from scientist Jurgen Schmidhuber which claimed that a simple data compression algorithm can explain “all essential aspects of subjective beauty, novelty, surprise, interestingness, attention, curiosity, creativity, art, science, music, jokes.” 

To SWIM It seemed so obviously reductive and universalist in its claims (some German dweeb claims to have mathematically defined what makes Lil B cool. Uh Huh.) But even if it’s undeniably naive or hubristic in scope, his theory does at least seem to describe something about the reception of aesthetic memes in social media.

According to Schmidhuber new stimuli like #deepdream become “temporarily interesting” until the audience “learns to predict or compress the data in a better way.” ie; #deepdream was interesting for SWIM while it was surprising and unexpected and they weren’t yet able to visually compress it into an identifiable trope. Now that it’s known as essentially a puppyslug filter meme there is a ‘curiosity and desire’ to “create or discover more non-random, non arbitrary…data that is novel and surprising.” Which #deepdream ain’t.

So while self-optimizing algorithms like #deepdream are learning to categorize and generate content, other algorithms are trying to define how stuff like this gets boring so quickly. In theory, if you could stack them together, you’d be getting close to a convincing automated content producer.  

Not unrelatedly, SWIM recently saw an NPR-created web app which predicted the likelihood of your job’s automation within the next 20 years. ‘Fine Artists’ faired amongst the least likely to be made redundant at 4.2%— phew!

 #deepdream is certainly a long way away from replicating our MFA-level cleverness, but once they have those novelty-creation modules installed...better watch out.


(All Images courtesy Google)