Fake the Funk

Computers can’t guess. They can’t generate random numbers or output anything that doesn’t come from following a rule or fulfilling a pattern. They can only calculate. This inability to create the unexpected matters because digital information is only as secure as the encryption used to protect it from spreading, and encryption requires random numbers, similar to the way a random password is much harder to crack than a password based on someone’s birthday or a sequence like 1-2-3-4. Computers fake the funk by running complex programs that generate pseudo-random numbers whose patterns are very difficult to discern, and that pseudo-random is random enough for, say, safe online shopping.

To come up with a truly random number you need to sample the world outside the computer, which is an actual and mathematically unpredictable mess. White noise is pure randomness, every moment patternless and unforeseeable. It’s what happens when equally intense information exists across all measurable frequencies. Randomness admits unruly information into old systems. This allows for change. It is also fun.

In 1980, Ikutaro Kakehashi went looking for superlative white noise. He found it in a box of reject electronics. Economic pressure had forced Kakehashi to abandon sonic realism: the memory chips required to make a drum machine that used actual recordings (as his competitor LinnDrum did) were prohibitively expensive, so he decided to synthesize the audio instead, using tiny analog transistors as noise sources to be shaped into high hat, snare, and other percussion hits. Scientific equipment couldn’t differentiate between different transistors’ amplified static, but Kakehashi’s ears could. He listened to dozens before he found a batch of 12,000 transistors whose white noise made for the best drum sounds.

Every unit got one. The transistors’ manufacturing errors were impossible to recreate, so when the supply ran out, in 1983, Kakehashi discontinued the Roland TR-808 Rhythm Composer. Forty years later the name “808” remains synonymous with powerful low-end bass and whipcrack snares, used everywhere from Marvin Gaye’s 1982 hit “Sexual Healing” to Kanye West’s album 808s & Heartbreak (2008; the old Kanye, mind you). Kakehashi’s choice of transistor—and by extension the legendary 808—is responsible for some of the world’s most popular drum sounds.

In a typical 808 beat, people use bits of white noise, the snare and the high hat, to mark the moments of swing, to underscore the offbeat, to fine-tune the tensions that encourage dancers to negotiate with the grid by bending themselves in and out of relation to it.

In 1985, Earl “Spanky” Smith paid $40 for a silver box from Japan. This secondhand “computerized bass machine,” manufactured by Kakehashi’s company Roland, lacked a manual, but no matter: in Chicago, Smith met up with his partners in the group Phuture to record a half-hour live exploration of the squelchy, chewy alien tones they were able to twist out of the TB-303, set against a spare kick/high hat/cowbell beat (digital drums from two other Roland boxes). The Windy City was still figuring out the parameters of its new invention, house music, and the scene was capacious enough to accommodate disco, gospel, and, hopefully, this avant-garde anti-anthem. Phuture played their tape to Ron Hardy, the most adventurous DJ in Chicago. He didn’t say a word until the song ended, thirty minutes later, then asked, “When can I get a copy?”

Phuture’s concoction cleared the dance floor when Hardy debuted it in a club, the Music Box. But he knew how shocking the sound was, and he knew how much his audience trusted his forward-thinking mixes too, so Hardy waited awhile, then played it again. And again. And again. By the fourth time it came out of the speakers, people were primed and the dance was on.

Within a few years the sound would spread across to the UK, sparking the genre quasi-eponymously called acid house. The 303’s quicksilver alien voice became synonymous with smiley-face iconography and cheap pills, the halcyon days of rave. Phuture’s “Acid Tracks,” which was released on vinyl in 1987 following its debut in 1985, was a revolutionary event: the song established and epitomized a genre at the same time as it exhausted its aesthetic possibilities, leaving everyone else to do endless iterations.

The 303 (as freaked by Phuture) frustrates comparison: it doesn’t sound like anything else. A chaotic ensemble of factors contributes to the bass synthesizer’s infamous sound, starting off with interactivity itself. The machine was intended to be painstakingly programmed, not played live. Yet the knobs and internal circuitry are highly interactive and respond dynamically to subtle changes. Phuture took immediate advantage of this.

The “Resonance” knob is particularly important. A regular low-pass filter works similarly to the Low-Mid-High EQs on home stereos. As you “sweep” the filter, the sound gradually loses high frequencies to become more muffled, until only the bass frequencies remain. The 303’s resonant low-pass filter employs a technical quirk: the Resonance knob boosts the sound at the frequency where it gets filtered into silence. The exact place of disappearance calls attention to itself. It does this by injecting the 303’s output signal back into its own input. This makes an unpredictable feedback circuit where minute variations can have a huge sonic impact. The description is technical but the effects are immediately audible: acid house is dedicated to exploring the slithery harmonics of an unstable system.

The moment of feedback is the moment when control threatens to become its opposite. It’s when relatively small aesthetic and technical decisions can snowball into enormous repercussions. None of this would have mattered if the 303’s engineer, Tadao Kikumoto, hadn’t cut corners to avoid getting sued by Bob Moog. Moog synths were (and remain) renowned for the “creamy” and “deep” sound of their patented filter design. Kikumoto stole Moog’s schematic layout but, to prevent lawsuits, swapped transistors for cheaper, less precise diodes. The diodes in such close proximity influence each other electrically, in nonlinear, complex ways. Hardwired instability makes the machine sound alive. These imperfections gel especially well with subsequent processing, like the overdriven 303 that enters at 2:30 in Daft Punk’s “Da Funk” (1995).

Acid offers entanglement at every scale: the weird signal fluctuations resulting from the knockoff design, the squelchy knob-twiddling liveliness, the way the squiggly new sound encouraged bodies to move in equally sinuous ways.

How does anything emerge out of life’s chaos to form a lasting pattern? Here, you needed Phuture’s Spanky and DJ Pierre and their manualless TB-303. But in order for it all to come together, you needed Tadao Kikumoto and of course Ron Hardy, who had changed the course of house music by the time he died of AIDS, only a few years later. You also needed the community of Black and queer Midwestern club kids who, after some coaxing, didn’t reject that alien sound but instead kept dancing to it. You even needed all the lounge acts who didn’t buy what was marketed to them as a “replace your bassist” box. The Roland TB-303 was a flop, discontinued in 1984 after a three-year run. Lack of commercial success kept it cheap enough for the experimentalists to afford.

Phuture’s 303 and Kakehashi’s 808 are prime examples of the thrilling things that can happen when musicians bend their (analog) tools into shapes the designers could never have imagined. John Cage upended the grand piano’s classical legacy simply by placing random objects on its strings to alter the sound, transforming his “prepared piano” into an unpredictable atonal instrument more like an eccentric gong. Try doing that with an iPad! Software is so much harder to hack, customize, or productively “break.” Yet over the last few years, a new wave of digital tools that offer truly new functionality have emerged, made possible by machine learning and artificial intelligence (AI).

AIs are calculators cursed with the ability to only see patterns, whether or not they are there. If you tell them what patterns you want (the pattern that forms around “Kate Middleton and her kids but Kate is the Monkey Christ, in the style of Picasso,” say), they will produce them, in constantly sharpening detail. The way they do this varies, although, like Kakehashi’s 808, most AIs start with white noise.

“Diffusion models” add random noise to images and, over time, learn how to remove noise from images—until they get so good at noise-removal that you can give them random noise and they’ll “remove” the noise until a completely new image is created. (Pareidolia, the tendency to perceive an image or pattern where there is none, such as hearing voices in radio static or seeing objects in a cloud, is the closest thing we have to experiencing an AI worldview.)

“Generative Adversarial Networks” (GANs) pit two AIs against each other. The “Generator” AI creates a cat pic, for example, then passes it to the other AI (the “Discriminator”), which decides whether the cat image is real or fake. A winner is announced, both AIs sharpen their skills, and the game repeats. The setup requires a reference database of human-labeled cat photos, but with that in place, the dueling AIs evolve very quickly. The Discriminator discerns with increasing sophistication, the Generator crafts ever more plausible cats. White noise is crucial to the Generator; without it as the seed value for each game, the exact same cat would be calculated every time.

ChatGPT works differently. The technology is only good at one thing: guessing what word might come next in any given phrase. Even its developers were shocked by how potent and wide-ranging this monotalent proved to be. Any prompt you give ChatGPT gets analyzed by approximately 175 billion parameters, which translate it into an array of numbers. These are called embeddings, simultaneously compressed information and specific locations in the AI’s ever-growing multidimensional information space. The next word is whatever embedding lies nearest. Chances are that most of the AI content you’ve experienced ran a diffusion model, GAN, ChatGPT, or some combination of the three.

Deepfake pop songs showcase AI’s mimetic ability (if it doesn’t sound like a passable Taylor Swift/SpongeBob SquarePants duet yet, give the tech a few months). This works by resemblance: analyze the known, then make more of it. To witness an artist’s self-defined style get deployed as statistical probabilities is uncanny, tiresome, previously unimaginable, and miraculous. Yet there’s no opportunity for meaningful feedback. You tell it what you want and get whatever the black box spits outs. Don’t like it? Write another prompt. The tool is far bigger than the user.

While the ability to generate an entire pop song off a text-based prompt is impressive, the tech is the opposite of hands-on. For that one must turn to the more modest, and more tweakable, AI-enhanced instruments. Take the Synplant 2 software that was released a few months ago: you give it a snippet of any recorded audio and it will re-create that sample as a dynamic synthesizer—an algorithm—you can edit and play. Plus, the interface is gorgeous. Before Synplant 2, you’d have to be an extremely skilled sound designer to come close to approximating any sound by electronic means. This is like having a bite of a restaurant meal and going home to a robot that has written down the recipe and cooked it from scratch.

Or, try Life by XLN Audio. Feed this app twenty seconds of audio or video and it immediately transforms your sounds into a beat. And the beats are good! As with Synplant 2, Life lets you refine or reshuffle with a fantastic level of detail.

These examples represent a “best of AI”: extremely focused; emphasizing bidirectional feedback and ways to shape what is being made; resource light, with the ability to run on one’s own computer and work offline. The tool is smaller than the user.

Music lives close to us. We feel it, we carry it. It beds down into memory and paces the day. (If music felt less free, would we pay musicians more?) Music is a public intimacy that celebrates the small scale and the anomalous. As AIs get better at spoofing us and each other, they necessarily improve at “Anomaly Detection”: identifying the outliers, the deviators, the rare and the noncompliant. Not only does AI require a massive scale of material resources, digital data, and energy to process it all, that scale relies upon, then reinforces, a concentration of power into the hands of a few people running a few companies. In no known history have a few dozen people holding that much computational leverage over everyone else turned out well for the small numbers. Serendipity doesn’t exist in digital spaces. For that you need white noise. True randomness means the next moment hasn’t yet been decided.