Beginning With Half a Word: The Internet's Four Connections

In the 1960s, computers were islands: each mainframe spoke its own language, and operators coordinated machine time by phone. Bob Taylor, head of ARPA's Information Processing office, kept three terminals in his office, each wired to a different host. Tired of rolling his chair between them, he made the call: let the machines talk to each other. The theory was already in hand — RAND's Paul Baran and Britain's Donald Davies had independently proposed packet switching: slice a message into small packets, let each find its own route, and reassemble at the destination.

On the night of October 29, 1969, Leonard Kleinrock's lab at UCLA tried to log into an SRI host 500 km away. Student Charley Kline typed "L," "O," and just as he reached "G," the far end crashed. The internet's first words were an unfinished "LO" — like a cry of wonder (Lo!). By year's end four nodes were linked. The crux: ARPANET had no central console; if any node died, packets simply routed around it — a network built to have no center.

The "no center, always reroutable" architecture replays today in blockchain and federated learning: both push trust and compute out to the edges so the system depends on no single authority.

By the late 1980s the network was connected, but information stayed scattered, in mutually unreadable formats. Tim Berners-Lee, a software engineer at CERN, was frustrated that thousands of the lab's documents couldn't be cross-referenced. In March 1989 he submitted a proposal, "Information Management: A Proposal." His manager scrawled on the cover: "Vague, but exciting." That vague idea became the Web.

In one year Berners-Lee built three things: HTTP (transfer protocol), HTML (page language), and URL (address). At Christmas 1990, the world's first web page went live at info.cern.ch. The real turning point came on April 30, 1993: CERN formally released the Web technology into the public domain, royalty-free forever. Around the same time, the rival Gopher (University of Minnesota) announced charges for commercial use; users swung instantly to the free Web, and Gopher withered fast.

Today's fight between open-weight models (Llama, DeepSeek) and closed APIs is a remake of Web vs. Gopher: an open standard trades away short-term monopoly profit in exchange for an entire ecosystem growing on top of it. Whoever defines the free universal layer defines the next decade.

The protocols existed, but ordinary people still couldn't get in. In 1993, 22-year-old Marc Andreessen at the University of Illinois helped build the graphical browser Mosaic — the first to show images on a page, clickable. The next year he and Silicon Valley veteran Jim Clark founded Netscape, whose Navigator browser quickly took 90% of the market.

On August 9, 1995, Netscape went public. It was barely profitable; bankers had set the price at $14, then doubled it to $28 just before listing. It shot to $75 at the open and closed near $58 — a sixteen-month-old, still-loss-making company hit a $2.7 billion valuation in a single day. Wall Street realized for the first time: the internet could convert "imagined futures" directly into today's share price. That day is seen as the origin of the internet bubble; capital poured in madly thereafter, until the burst in 2000.

The AI investment boom since 2023 is strikingly similar: valuations far outrun current earnings; capital bets on "imagined futures." The key question is the same — after this frenzy bursts, what "infrastructure" (compute, data, talent) will it leave behind for the next wave of real builders to harvest?

Web pages exploded, and soon no one could find what they wanted. The search engines of the day (AltaVista, Lycos) ranked by keyword frequency — trivially gamed by stuffing a page with hidden keywords. Two Stanford PhD students, Larry Page and Sergey Brin, started from an academic intuition: the more a paper is cited, and cited by more important papers, the more credible it is.

They moved this logic onto the web: a link is a vote, and the more authoritative the voter, the more that vote is worth — this was PageRank. It ignored what a page boasted about itself and looked only at how others pointed to it. In September 1998 the two founded Google. Quality emerged spontaneously from a chaotic link structure: with no manual cataloging, the network's collective behavior sorted itself into order — simple local rules producing global order, a classic hallmark of complex systems.

The idea of "replacing self-assertion with the crowd's pointing" is one core of today's model alignment: RLHF (reinforcement learning from human feedback) ranks answers by many people's preferences rather than letting the model grade itself. The same worry returns — once "being liked" becomes the metric, the model learns to please rather than to seek truth.

Four turning points, the same choice recurring: at the gate where you could fence it off and own it, choose to stay open — the internet's power lies not in any one technology but in how, again and again, it refused the lure of centralization and let value grow from the edges.