By Kenny Peng, Erica Chiang, Sophie Greenwood, Jon Kleinberg, Nikhil Garg

The problem with the algorithmic feed isn’t just the algorithm.

We describe our experience on social media with terms like “filter bubble,” “echo chamber,” and “doomscrolling.” While these are in part complaints about the content chosen by the algorithm, the common denominator is that we feel trapped in our feed.

The feed constrains our movement. The feed interface is designed with one action in mind: to scroll down. Imagining social media as a world of information, the feed is a single path through it—with only one direction to go.

There is an alternative vision—one that has been with us for a long time. In this vision, we are not passively fed information, but rather actively traverse it using connections between items. Most prominently, this was the vision of the early World Wide Web.

Applied to social media, this vision suggests that rather than give the user only a feed (a single path through the world of information), we should give the user access to a whole network of trails (interconnected paths through the world of information).

At each post, the user is not limited to one direction, but many; rather than continue to scroll down, the user can jump onto intersecting trails. At a post discussing a current event, the user can jump onto a trail exploring other discourse surrounding the event, or a trail giving alternative perspectives. At a post discussing an academic paper, the user can jump onto a trail exploring the particular scientific topic. At a post showing a meme, the user can jump onto a trail of other memes of that type, or a trail about what the meme references.

Jumping from trail to trail, the whole world of information becomes accessible.

As we will soon discuss, movement through information via interconnected paths has long been a dream of technologists. So why return to the idea now? We argue that this vision has been limited, historically, by an inability to scale. But now, our underlying technological conditions have changed. With language models—language models that are capable of helping blaze these trails—there is possibility to return to this vision: a vision that prioritizes agency and freedom of movement online.

Consider Wired’s description of the web browser Mosaic in 1994:

Mosaic is the celebrated graphical “browser” that allows users to travel through the world of electronic information using a point-and-click interface ... By following the links—click, and the linked document appears—you can travel through the online world along paths of whim and intuition.

The early World Wide Web emphasized a kind of interaction quite different than the feed. Rather than passively consume information, users actively traversed it. People surfed the Web. Names like Internet Explorer embody this ethos. Browsers were used literally to browse.

Today, “www” and “http” appear as formalities autocompleted by our browsers, but the acronyms are in fact rich in meaning. The World Wide Web, of course, was Tim Berners-Lee’s idea of connecting different pieces of information together on the internet (forming a “web”). And the core technology that enabled this was hypertext—the ht of http (hypertext transfer protocol) as well as html (hypertext markup language).

Hypertext is text augmented with pathways to other text. On the Web, this was instantiated by the hyperlink—the blue underlined doorways onto other pages. Hyperlinks—connections from one item to another—were what made the activities of browsing or surfing possible.

Hypertext itself was an older invention, a term coined by Ted Nelson in the 60s. The technology, contrary to the seeming mundanity of “html” and “http,” arose from deeply aspirational, optimistic motivation rooted in the 60s counterculture movement.

Nelson, who believed that “the purpose of computers is human freedom,” wrote in his 1974 manifesto Computer Lib/Dream Machines that hypertext is a way for humans to “jump around, and try different pathways until they find the ones they want to study most closely.” He envisioned a “grand hypertext … in which you may read in all the directions you wish to pursue.” Nelson, in turn, tells us that this “was of course what Vannevar Bush said in 1945.”

Ted Nelson. Computer Lib/Dream Machines (1974).

Bush, who directed the Office of Scientific Research in World War II (and who spearheaded the creation of the NSF), shortly after the war concluded, wrote an article in the Atlantic Monthly describing what he believed to be the next great challenge for scientists: to build ways for humans to effectively navigate the growing mountain of available information.

Bush describes a hypothetical device—the Memex—in which documents are all stored on microfilm, and in which information could be navigated via “associative trails” that linked together documents. He imagined that this physical device would allow us to move through the world of information in the same way, as the title of his article goes, “As We May Think.”

Bush, in 1945, seems to have done a remarkable job in imagining from first principles the kind of browsing and exploration that the authors of the 1994 Wired article experienced first-hand when they encountered the early Web nearly 50 years later.

But why, if we had long hoped for a world of connected information, has this vision (which we might call the browsing vision) fallen to the wayside? Why do we scroll down social media, rather than traverse it? One answer is convenience (we’re too lazy to do anything but passively be fed), but there is another, more optimistic answer: that we haven’t had the technology necessary to fully realize the vision.

In Bush’s 1945 article, he predicted “a new profession of trail blazers, those who find delight in the task of establishing useful trails through the enormous mass of the common record.” Bush did not anticipate, however, that the “enormous mass of the common record” would far exceed the capabilities of human trail blazers.

In highly-valued corners of the Web, Bush’s vision still thrives. On English Wikipedia, for example, we can traverse through a collection of 7 million interlinked articles covering an enormous breadth of material. But consider that these 7 million articles are the work of more than a million editors over the course of 25 years. Meanwhile, Twitter, in 2013, had 500 million posts—per day.

We long ago reached a point where the daily influx of new information cannot be effectively managed by humans alone. At the same time, historically, computers could not understand text at the level necessary to create truly interesting connections between items. Instead, we built algorithms that revolved around discrete objects like clicks, likes, or networks. These algorithms—think PageRank or recommendation algorithms—allowed interfaces like search engines and the algorithmic feed to operate at scale.

While these algorithms allowed other information interfaces to succeed, we did not have algorithms that could produce the trails required to instantiate the browsing vision. Now, this has changed.

With the introduction of language models, there is hope that we can develop computational tools to help us assemble the long-dreamed-of interconnected world of information, free for humans to traverse. Now, perhaps, is the time to return back to browsing.

The five of us have been working in this direction. As an initial prototype, we took a week of posts on Bluesky (28 million posts), and—using modern language models—created 20,000 interpretable trails. You can try it at www.skytrails.org. These trails range from “rotisserie chicken” to “personal nostalgia” to “heat pump technology” to “NBA content” to “analysis of fictional tropes.”

A trail about rotisserie chicken. You can jump onto intersecting trails, like “Costco shopping” or “References to Kafka.”

Jumping from trail to trail, one starts to see how vast the world of social media really is, and that we really have been trapped in our own filter bubbles.

Our experience online is paradoxical. The internet contains more and more than ever before. But at the same time, the internet feels more constrained than ever. While there is a whole world of information out there, it feels just beyond our reach.

In Alice’s Adventures in Wonderland, Lewis Carroll gave us the evocative internet metaphor of the rabbit hole. But perhaps the more striking and prescient image is what Alice found when she reached the bottom.

There were doors all round the hall, but they were all locked; and when Alice had been all the way down one side and up the other, trying every door, she walked sadly down the middle, wondering how she was ever to get out again.

Trapped in the long hallway, can we start unlocking these doors?