I | The Web was Built for Eyes

When I was fifteen, I started a music blog on WordPress. This was 2009, 2010, the dead space after the financial crisis, that liminal window between Limewire and Spotify when nobody was quite sure how music would travel anymore. I'd find MP3s of artists who didn't exist in any catalog, artists emerging online before anyone with a record deal had heard of them. Avicii, Wiz Khalifa, Mac Miller, dozens of others. I'd write a few paragraphs about why they mattered, update the files with custom album art and metadata so that when the tracks travelled across the internet they'd carry attribution back to the blog, and post them. I was fifteen. I was rushing between classes to get drops up before the other blogs. Within a few months, tens of thousands of people were reading it.

What had emerged, without anyone designing it, was a small distribution network. Artists sent us tracks. Fans found them through us. Other blogs linked to us and we linked to them. The whole thing ran on WordPress templates and MP3 files and the fact that the web made the cost of connecting supply to demand effectively zero. Anyone with curation instincts and distribution channels could find an audience because the infrastructure made it free to participate.

That infrastructure was built, every layer of it, for a specific kind of user. A person sitting at a screen, with eyes and a cursor and roughly eight hours of browsable time per day. It shapes everything. And it's the only way to understand what's about to change.

In March 1989, Tim Berners-Lee submitted a document to his manager at CERN titled "Information Management: A Proposal." His boss, Mike Sendall, scribbled "vague but exciting" on the cover and let him spend some time on it.

The problem Berners-Lee was trying to solve was local and specific: physicists at CERN used different computers running different software, and information about experiments, personnel, and systems was scattered across incompatible machines. Knowledge was constantly being lost. With a two-year average stay, researchers would leave and take their institutional knowledge with them. "The introduction of new people demands a fair amount of their time and that of others before they have any idea of what goes on," he wrote.

His insight was to combine two existing ideas. The internet, a network of connected computers exchanging data using TCP/IP, had been growing since the 1970s. Hypertext, the concept of documents containing links to other documents, had been explored since Vannevar Bush's Memex proposal in 1945 and Ted Nelson's coinage of the term in 1965. Berners-Lee married them: hypertext documents, served over the internet, addressable by a universal locator, formatted in a standard markup language. HTML, HTTP, URLs. The entire architecture of the web, in its initial form, was three protocols and an addressing convention.

The simplicity was the point. Working alone on a NeXT workstation with no team and no budget, he needed something that could be built fast and adopted without institutional permission. He didn't build something complex. He built something minimal and let the complexity emerge.

And CERN released it royalty-free in 1993. The protocols were open. Anyone could build on them. This decision, more than any technical one, is why we have the web we have.

Every layer of those protocols encodes information into a form that a human decodes with their eyes.

HTML encodes the layout of a document into markup that a browser renders visually: text, images, and links arranged for a person scanning a screen. CSS encodes presentation decisions into rules that produce something legible to human eyes. JavaScript encodes interactivity into scripts that respond to clicks and keystrokes and scrolling. The entire frontend stack is a rendering engine for human visual perception.

HTTP encodes the request cycle into a pattern that maps to how a person browses. A client sends a request, a server returns a document, each interaction stateless and independent. Cookies were bolted on later to simulate continuous human presence across page loads, the illusion of a session, because the protocol needed to track a person navigating from page to page, spending seconds or minutes on each, following a thread of interest.

URLs encode addressing into a single string a person can read, share, bookmark, or type into a browser. Before URLs, accessing information on a remote computer required knowing the machine's address, the file path, and the access protocol separately. URLs collapsed all of that into something that could be embedded in a link and followed with a click.

These are encoding tools tuned for human decoding. Every design choice optimizes for a person with eyes processing a two-dimensional screen. That fact is worth holding onto, because most of what came next emerged from it.

Nobody planned the web's emergent properties. They arose from simple protocols, open access, and millions of people acting on local information and local incentives.

A link in HTML is trivial: a tag pointing to a URL. But links in aggregate create something no individual linking decision intended. They create a graph, a vast network of relationships between documents where the structure itself encodes collective human judgment about what matters and what connects to what.

Larry Page and Sergey Brin recognized this in 1998. PageRank treated links as votes. A page that many other pages linked to was probably important. A page linked to by important pages was more important still. But the insight ran deeper than counting links. The entire relevance infrastructure that Google built, and that the industry of search engine optimization grew up around, was reading human behavioral signals through the browser. Click trails revealed what people actually found useful after they landed on a page. Session duration indicated whether content delivered on its promise. Bounce rates exposed the gap between what a headline claimed and what the page contained. Links showed what authors considered authoritative enough to reference. And the structure of the pages themselves, header tags, meta descriptions, content placement above the fold, was optimized for how human eyes scan a rendered screen. SEO was the science of encoding pages for human visual decoding and for the algorithms that measured human response to that decoding.

This is where "built for eyes" stops being a technical observation and becomes something deeper. The web's protocols encoded information for human consumption. But the web's intelligence, the collective signal about what was true and important and trustworthy, depended on humans producing the signal. Every link was a human judgment. Every click was a human choice. Every second spent on a page was a human signal that the content delivered or didn't. PageRank, and the relevance systems that followed it, were theories of collective human discernment. They worked because the agents generating the data were people who could read, evaluate, and decide. Later, social feeds shifted the model further, with Facebook and Twitter and TikTok curating not by link structure but by engagement signals, likes and shares and watch time, but the underlying dependency was the same: algorithms reading human behavior to determine what mattered. The web didn't just display information for eyes. It thought through them.

The rest of the emergent properties followed the same pattern. Wikipedia, Stack Overflow, open-source collaboration, the entire creator economy, none of it was planned. It emerged from simple rules, open access, and millions of agents pursuing their own interests within those rules.

The web's economic model is itself an emergent property, and it rests on the same foundation: human attention is scarce.

A person has roughly sixteen waking hours, and only a fraction of that is spent online. Advertising works because human attention is finite and valuable. Content in return for eyeballs, with advertising as the monetization layer. The web inherited television's business model not because anyone chose this but because the structural conditions, scarce human attention consuming media, were similar enough that the same economic logic applied.

When you monetize attention, you create selection pressure toward whatever captures it most efficiently: engaging, provocative, emotionally resonant, addictive, regardless of whether it's accurate or useful. Clickbait is the logical endpoint of an economy that pays per eyeball. Paywalls represent a different model, one where the consumer pays directly, but paywalls also assume a human who will make a purchase decision, enter credentials, and consume content in a browsing session. Both models orbit the same constraint: a human with limited time in front of a screen.

Even the web's pathologies are emergent properties of its human-centric design. CAPTCHAs exist because protocols built for human users need a way to distinguish humans from automated scripts. Anti-scraping measures exist because content encoded for human eyes is being consumed by machines in ways the economic model didn't anticipate. The robots.txt convention, a text file asking automated agents to respect certain boundaries, is a gentlemen's agreement layered on top of protocols that have no native concept of user type.

The web is the largest non-parametric knowledge store in human history. That distinction, between parametric and non-parametric, matters for everything that follows.

A parametric system compresses knowledge into a fixed architecture. A trained language model, for instance, encodes everything it has learned into a set of numerical weights, billions of them, determined during training. Those weights are a lossy compression of the training data: patterns, relationships, and statistical regularities distilled into numbers. Once training ends, the knowledge is frozen. The model can't learn new facts without being retrained. It can't tell you what happened yesterday. It can't update when the world changes. And because the compression is lossy, information is inevitably lost in the process. The model may generate something that sounds right but has no grounding in any specific source, because the source was dissolved into weights during training. This is the fundamental nature of parametric knowledge: it's compressed, frozen, and unattributable.

A non-parametric system stores knowledge in its original form and grows without bound. Add more data and the system accommodates it. A search index is non-parametric: add documents and it expands. A library is non-parametric: add books and it grows. Crucially, the original sources are preserved. You can trace a claim to the document that made it. You can check when it was written, by whom, in what context.

The web is non-parametric and alive. New pages go live every second. Existing pages are revised, extended, corrected. The knowledge base reflects the current state of human understanding, not a frozen snapshot from a training run. When something happens in the world, the web knows about it within minutes.

The open web is a miracle. Anyone can publish, learn, and collaborate. It is the closest thing to humanity's living memory: a continuously evolving representation of what we collectively know, believe, argue about, and care about, right now. Its quality is uneven precisely because its coverage is universal. And that universality compounds over time from two design decisions: open protocols, meaning anyone can publish, and universal addressing, meaning anything published gets a stable identifier that anyone else can discover and link to.

The web succeeded not because it was the best hypertext system. Ted Nelson's Xanadu was more ambitious, with built-in versioning, attribution, and micropayments for content. Douglas Engelbart's NLS demonstrated more sophisticated features in 1968 than the early web offered in 1991. The web won because it was simple enough to deploy without permission, open enough to evolve without coordination, and minimal enough that complexity could grow from the bottom up.

Its genius was in what it left out. No native identity layer. No payment system. No access controls. No built-in concept of trust or reputation. Every omission has created problems we've spent decades patching with layers on top of the original protocols. But the omissions are also why the web grew as fast as it did. Every feature that requires coordination slows adoption. Berners-Lee shipped the minimum viable set of protocols, and the world built everything else.

Those omissions all share a root cause. The system was designed for a user who could evaluate, judge, and decide. A person with eyes, reading a page, deciding whether to trust it, whether to click, whether to pay. No payment system because a human would decide to pay. No trust layer because a human would evaluate credibility. No attribution system because a human would follow a link to the source. The protocols offloaded judgment to the human. They didn't need a trust layer because the human was the trust layer.

The web is extraordinary infrastructure built on a single assumption that was true for its first thirty years.

Its protocols encode information for human eyes. Its intelligence, the links and rankings and relevance signals, depends on human judgment. Its economics run on human attention. Its trust model relies on a human reader who can evaluate what's in front of them.

That assumption is embedded so deeply in the architecture that it's invisible, like asking a fish about water. And the assumption is becoming less true.

Not because the web is going away. The web for human eyes will persist, the way physical mail persists after email, the way vinyl persists after streaming: as something valued precisely for being crafted for human experience, intentional and designed. But the utilitarian function of the web, finding and retrieving and synthesizing information, is migrating to a different kind of user. AI agents that don't have eyes. They have context windows and process tokens. They don't browse a page or scan a layout. They consume structured data, millions of requests per day, at a speed and scale no human user could approach.

When the telephone network was built for voice, nobody anticipated modems. The system worked, poorly, for data transmission because the infrastructure had been designed around assumptions about signal type and session duration that data traffic violated at every level. The answer wasn't better modems on voice lines. It was broadband: new infrastructure for the new user.

The web's new users are arriving now. What happens to a knowledge system whose intelligence was built by human eyes, when the primary users are machines?