The Serendipity Machine

In the first TikTok that I can recall watching, a teen girl was using a mirror effect to pretend to eat her disembodied, floating fingers to the Kidz Bop version of Evanescence's 2003 classic “Bring Me to Life.” On screen, she could barely keep from breaking into laughter. Mundane, zany, completely contextless, endless replicable: perfect internet humor. It was the summer of 2019, and TikTok had that heady, diffuse feeling that social media companies struggle to cultivate in their early days—the sense that people are gathering and coalescing into communities because something is happening on their platform.

What is TikTok? What are any of the major social media platforms, these days? They are never-ending A/B tests, where a single piece of content slides onto the screen and all interactions with it—milliseconds of watch time, number of watches, likes, shares, comments, even things like phone orientation—are combined with contextual information—IP address, device type, time of day, location—to become a remarkably powerful information extraction machine.

In his New Yorker essay “Why the Internet Isn’t Fun Anymore,” Kyle Chayka argues that the internet has neither the optimistic shine nor the informational functionality that it used to, despite the fact that content creation is at an all-time high. More than a year of video is uploaded to Youtube every 18 minutes, 120 zettabytes (a million million gigabytes) of data were created in 2023; at this scale, content quickly outstrips the bounds of human intuition and experience. All this, Chayka claims, means that the core model of the web has changed:

The platforms that have the most traction with young users today—YouTube, TikTok, and Twitch—function like broadcast stations, with one creator posting a video for her millions of followers; what the followers have to say to one another doesn’t matter the way it did on the old Facebook or Twitter… [C]onversation isn’t strictly necessary, only watching and listening.

The social web is dead; in its place is the publisher-consumer model, where content is hyper-targeted, self-branding is critical, audiences are bigger, stakes are higher, and rewards are greater. I’m not an exception to these trends. I, like most users, am a lurker.

Recently, I’ve often found myself losing time on TikTok. Decontextualized videos fall into my feed, plucked from an unknowably vast realm of content, each struggling to fit me. I get a cat playing piano, a critique of the colonialism of the Mercator projection, an AI generated filter of “how I would look in the 1930s,” and a gloss of the Swedish engineer who created the first moose crash test dummy, one after the other in an improbable chain of manufactured serendipity. But when I close the app, brain crashing from a screen-induced dopamine high, I often have a sense of my own mortality, that I just wasted half an hour (or, let’s be real, an hour, or an hour and a half) of time that I will never get back. Seen in that light, the name TikTok becomes a macabre double entendre.

Ultimately, that feeling—endless serendipity—is what social media platforms need to keep eyes on screens (and the advertisements contained therein). In order to survive, platforms need to give consumers a sense of things falling effortlessly into place; that, somewhere in the slice of human experience captured in photos, videos, and text, there is something resembling individual meaning.

TikTok is the culmination of a multi-decade process of knowing-at-scale, of endless customization. That road began in the wake of the dot-com bubble in the early 2000s, when Facebook was just a twinkle in Mark Zuckerberg’s 16-year-old eye. It began with a company that struggled to survive the ups and downs of the social media market for decades; with a cast of founders, investors, and engineers who have shaped our era of technology; with a seemingly simple technology—the customized algorithm—that has become the hallmark of our era. It began, counterintuitively, with StumbleUpon.

---

StumbleUpon rose from the smoldering ashes of the dot-com bubble. In November 2001, Garrett Camp (one of the future co-founders of Uber), Eric Boyd, Geoff Smith, and Justin LaFrance, all University of Calgary students, started building an extension for the Mozilla Firefox browser that let you semi-randomly discover internet content. One click of the “Stumble” button, and you’re on a new page.

The internet of the mid-2020s is an almost entirely frictionless place. I slide from one piece of content to another—auto-generated playlists, auto-recommended videos, auto-summarized essays; screen-recorded TikToks reposted on Twitter, screenshotted Tweets on Instagram. It’s often easy to forget how difficult the internet could be in its early days. When StumbleUpon was created, the web was mostly static images and text. Social media beyond listservs didn’t really exist (MySpace and Facebook were founded in 2003 and 2004, respectively). “To google” wasn’t yet a verb (it was added to the Oxford English Dictionary in 2006).

All this meant that the frictionlessness of the “Stumble” interaction was novel, as was the quality of the content. When the University of Calgary group released a beta on the Mozilla add-on store in February 2002 it had instant virality (another novel concept in the early internet era). “I think we were the 138th extension for Firefox ever written, and we got a top rating very early,” Camp said in a 2010 interview. “People liked it, and we followed the comments and fixed anything that was wrong. After a while we had 4.8 stars out of five, and we started to get thousands of installs a day, for free.” The business model was simple—every 20th stumble took a user to advertiser content, and advertisers paid 5 cents per visit.

The rest of the story reads like an exhausting Silicon Valley side-plot, with a menagerie of half-formed business ideas that could have succeeded but never quite took off. There was exponential growth through the 2000s and early 2010s, with millions of users and billions of stumbles; a $75 million acquisition by eBay in 2007; a few years of platform stagnation; a $29 million counter-acquisition to take the company independent again in 2009; a massive app redesign in 2011 that flopped, shedding 25% of the app’s traffic in a month; mass lay-offs and a poorly thought-through purchase of an up and coming video start-up in 2013; and, finally, the slow death of the community, culminating with the end of the app in 2018. For a brief moment in 2011, StumbleUpon drove more traffic to websites than Facebook—then it came tumbling down.

StumbleUpon used the simple predecessors of the mathematical constructs that we think of today as The Algorithm—collaborative filtering, pagerank, basic social graphs—to make the internet fun and accessible. It was distinct from the rising monoliths of 2000s social media in that it was explicitly focused on content consumption rather than profitable co-creation. StumbleUpon was one of the first serendipity machines; it was ahead of its time, and it was addicting.

---

StumbleUpon was also my pre-teen gateway to the internet—to my developing sense of self and my rapidly-expanding sense of the world. Upon signing up, I ticked a few boxes pertaining to my interests. Then, after a day of school, I would come home and log onto my older sister’s hand-me-down iMac and open the site. As the short New England days grew into evening, I would while away the hours. There was a seemingly-interminable pool of content, an endless number of creative individuals who poured their souls into a bewildering curio shop of websites, animations, photos, videos, articles, and music. Not interested in looking at something anymore? One click and it was gone, replaced by something else that was just about as interesting.

StumbleUpon was among the first platforms to algorithmically order content, even though it maintained the patina of randomness. StumbleUpon used social media algorithms that ranked content from the very beginning—Facebook started the first edition of its newsfeed (a reverse chronological list of posts from friends) in 2006, and only started inserting content algorithmically into newsfeeds in September 2011. StumbleUpon, then, was a novelty and an industry leader. The platform’s success came from its simplicity and serendipity, the algorithm’s ability to, for the first time, maintain a fine balance between unpredictability and delight.

These phenomena are quotidian now; they are the water of the internet that we all swim in. But at the time, they were revolutionary. We know now that my developing brain was being flooded with neurotransmitters associated with positive emotions on a semi-random, intermittent schedule—the slot machine effect, which has been shown to be correlated with addictive tendencies. Dominant social media platforms have been hauled into congressional hearings to address social media addiction and the resultant increase in adolescent depression, based on the formula that these engineers figured out in the early 2000s.

---

2023 was a uniquely crappy year for the dream of large-scale online communities dedicated to the organization and discussion of high-quality, engaging content on the web. The signs have been there for years—platforms incapable (and unwilling) of solving problems of effective moderation, the use of the coronavirus-induced economic downturn to gut trust and safety teams, whatever the hell is happening at Twitter/X, the proliferation of AI for automated scamming and non-consensual deepfake porn. The internet has become fragmented, and the largest fragments (Facebook, Instagram, TikTok, Twitter, etc.) are increasingly enshittified as they pursue ever-more elusive profit margins.

It’s no wonder that, in this context, I started reminiscing about the era of the internet that is typified by StumbleUpon—characterized (in my memory) by independent websites, where the correct verb to describe what one did when one “went on the computer” was “browsing.” I’m not alone in that sense of nostalgia. The modern-day set of monopolistic companies that impose an almost neo-feudal structure on the web remind me of the libertarian hellscape that is America in Neal Stephenson’s Snow Crash; StumbleUpon’s ethos of randomness, on the other hand, recalls the self-made anarchist society in Ursula Leguin’s The Dispossessed. Of course, there are some living vestiges of randomized platforms that live on, reminding us of a grimier, less-optimized era of the internet: ChatRoulette (ick), the “I’m Feeling Lucky” button on Google (which reportedly costs Google millions in lost ad revenue annually), and the Wikipedia “Random Article” button. But with the deaths of Omegle and StumbleUpon, and the relegation of “I’m Feeling Lucky” to a vestigial feature, an era of randomness on the internet has taken a definitive hit.

StumbleUpon got its start as an weighted index of the cool shit people put on the internet. Its foundational assumption of the platform was that the internet was a splintered, cacophonous, chaotic place, and that if you knew where to look you could find some beautiful things. There was no assumption of platform monopolism, or of pre-existing algorithmic curation. It didn’t (and probably couldn’t, due to the move-fast-and-break-things nature of the early social web) bookmark individual tweets or Facebook posts. It was predicated on the idea that the internet could be outside of the grasp of a small number of actors. Once again, the name is apt. Using StumbleUpon really felt like stumbling upon something out there in the world.

That feeling of coming across, of true serendipity, was both StumbleUpon’s genius and its ultimate downfall. StumbleUpon was one of the first companies to deploy The Algorithm in the real world, which today is discussed with in a tone usually reserved for the supernatural: revered for its ability to make money, hated for its addictive qualities, and bemoaned as an accelerant of social division. At the same time, StumbleUpon never was able to create The Platform—the space of total corporate control in which expression is commodified into content and social nuances are pressed into a small repertoire of approved interaction patterns.

Some potential solutions to the contradictory and deeply unsatisfying legacy of StumbleUpon lie in changes that are already happening in the industry. Instead of simultaneously toeing the line between the publisher-consumer and social sharing models, some platforms (think Substack, Twitch, OnlyFans, etc.) explicitly formalize the publisher-consumer model. To the extent that there’s discovery, it functions mostly on a trusted recommender model. Other platforms (Discord, Signal, Telegram, Slack, etc.) break those older models entirely, largely sidestepping questions of individual relationships to focus on the creation and curation of communities. Both of these approaches, used in the appropriate contexts, can be incredibly useful.

The primary difference between the modern platform of today and StumbleUpon is the modern platform’s ability to monopolize the frame in which content is presented. They proscribe content, stomp out the chaotic weirdness of the internet, limit external referrals—all to squeeze every last ad impression out of their users. Think of TikTok and Instagram injecting keystroke tracking Javascript into every webpage that users open on the internal app browser, or Twitter/X under Elon Musk (briefly) removing titles from all external links. StumbleUpon did not, and could not, ever do that. At the same time, almost every large-scale social media platform of today is made in StumbleUpon’s image. The platform’s ingenuity prefigured (in its early stages, at least), the publisher-consumer web that is the murderous child of the social web of the early 2010s.

---

I think back on the StumbleUpon era of the internet—from around 2007 to 2014—with a deep sense of nostalgia. Many people do, and for good reason. The serendipity machine was still a new phenomenon. The content hadn’t yet gotten so targeted that 55% of American respondents to a survey thought their smartphones were listening to them. Individual companies hadn’t yet monopolized the frame of reference. There was error, and randomness, and the feeling of being able to truly connect with new people.

But let’s be real: In 2012 as in 2024, Twitter was already a locus of hate speech; Facebook was already starting to amplify ethnic tensions in Myanmar; eating disorder researchers were already starting to show that social media had intensely negative impacts on adolescent girls’ body images. Omegle and ChatRoulette were full of guys who got off on flashing strangers over the internet. No one really liked the “I’m Feeling Lucky” button (which was more of a cultural signifier of Google’s soon-to-be-dead “Don’t Be Evil” era than anything else). And the Wikipedia Random Article button usually returned a page about moths, or a random town in Denmark.

And during those long afternoons on my sister’s desktop spent stumbling upon new websites, I was mostly alone, just like I am mostly alone as I scroll through TikTok in 2024. The light would lengthen, the sun would go down, and the room would get dark. In the darkness, I would eventually come back to my senses. Looking back on it, the vaguely sickening feeling of a content-induced dopamine high is oddly familiar.

---