brianchitester.com Posts Polymath Profiles Games RSS

Project Xanadu

On 6 April 1988, Autodesk, the Sausalito company that made the AutoCAD drafting program, announced that it was buying 80 percent of a small software company called Xanadu. John Walker, Autodesk’s founder, said the product would ship within 18 months. He also gave a longer forecast, which the journalist Gary Wolf later quoted back at him in Wired: “By 1989, it will be a product. And by 1995, it will begin to change the world.”

He was right about 1995. The world was changing that year, because millions of people were going online for the first time and clicking links on the World Wide Web. Xanadu had not shipped. That June, Wolf’s article called it “the longest-running vaporware story in the history of the computer industry,” vaporware being software that is announced and never delivered. The project had started 35 years earlier.

The Pitch

In 1960 Ted Nelson, a graduate student in sociology at Harvard, took a computer course and came away with an idea about writing. Paper forces text into a single line, one page after another. Thought doesn’t work that way. Nelson wanted a system where a writer could keep every version of a document, set two texts side by side, and connect any passage to any other. In a 1965 paper for the Association for Computing Machinery, the main professional society for computer scientists, he gave this kind of text a name: hypertext. In 1967 he named the larger system after the pleasure dome in Samuel Taylor Coleridge’s poem “Kubla Khan.” The project’s own history page describes Xanadu as “a magic place of literary memory and freedom, where nothing would be forgotten.”

What Nelson was proposing was something like the web, designed by someone who had thought hard about what could go wrong with it. Over the following decades the design settled into a list of 17 rules . A few of them carry the whole idea.

Links would run both ways. If one document linked to another, the second document would know it, and a reader of either could follow the link in either direction. Nothing would be orphaned, and you could always see who was quoting you.

Quoting would happen by reference, not by copying. Nelson’s word for this was transclusion: instead of pasting a paragraph into your document, you would include a live window onto the original. The reader would see the quoted passage in place, could step through to its source, and could see what surrounded it there. Nothing would be copied, so nothing could drift out of date or lose its attribution.

Every document would have a permanent address, and nothing would be deleted. A new version would sit alongside the old ones, so a link made in 1990 would still lead somewhere in 2020.

And authors would be paid. Each time a reader pulled up a piece of someone’s work, including a transcluded quotation, a small royalty would go to whoever wrote it. Publishing a document would automatically grant everyone permission to link to it and quote it. Copyright and quotation, which on paper were in constant conflict, would be settled by a meter.

Put together, this was a single library for everything ever written, where every quotation led back to its source and every writer got paid. Nelson has spent his life since arguing that this is what computers are for. In 1974 he self-published Computer Lib/Dream Machines, an oversized book with two front covers, one for each half, subtitled “You can and must understand computers NOW.”

The People

Ted Nelson was the son of the actress Celeste Holm and the film director Ralph Nelson, and had a philosophy degree from Swarthmore College before he went to Harvard. He was not a programmer and never became one. Wolf describes a man who recorded nearly everything on tape and video and threw nothing away, afraid of forgetting. He had attention deficit disorder and called his own thinking a “hummingbird mind.” He told Wolf that Xanadu was “the first step to anything I ever wanted to do.”

Roger Gregory was the one who stayed. He was repairing equipment at a used-computer shop in Ann Arbor, Michigan, when he came across Nelson’s work. In the summer of 1979 he hosted the core design team in a house in Swarthmore, Pennsylvania, and in 1983 he founded the Xanadu Operating Company to turn the design into a product. He paid for hardware himself: in 1982, according to Wolf, he bought a Sun workstation for $26,000 and an 80-megabyte disk drive for another $10,000. Gregory believed the system was nearly done for most of the next decade. “I had something that was within six months of shipping,” he told Wolf.

Mark S. Miller was one of the young programmers at Swarthmore that summer. He was a Yale undergraduate, and he became the project’s lead architect in its Autodesk years. Miller’s ambitions were larger than Gregory’s: he wanted Xanadu built properly, on a cleaner foundation, even if that meant starting over. That disagreement would decide the project’s last real chance to ship.

What Got Built

For its first two decades Xanadu was mostly paper: Nelson’s books, design documents, and a set of data structures for storing text so that every version and every link could be retrieved quickly. The Swarthmore summer of 1979 turned those ideas into an architecture. The team, according to the project’s history , included Nelson, Gregory, Miller, Stuart Greene, Eric Hill and Roland King.

What came out of the next decade was a working server, written in the C programming language and mostly complete by 1988. It stored documents, versions and two-way links, and it ran. Years later it was released as open source under the name Udanax Green.

Then Autodesk arrived. Nelson had met Walker at a hackers’ gathering in 1983, and Walker, whose own company had grown from $15,000 in sales in 1983 to more than $54 million in 1987, had money and patience. Under the 1988 deal Xanadu moved to offices in Palo Alto, more than an hour’s drive from Nelson, who worked out of Autodesk’s headquarters in Sausalito. The programmers, now led by Miller, decided that Gregory’s C code was not a sound base for what they wanted to build. They set it aside and began again in Smalltalk, a programming language better suited to the kind of object-based design they had in mind.

The rewrite, later released as Udanax Gold, was never finished. The team split into factions over it, and an Autodesk deadline passed. Marc Stiegler, the manager Autodesk installed to get a product out, later told Wolf that Xanadu “constantly defeated me.”

The office did produce one thing that worked. In April 1989 Robin Hanson, a consultant at Xanadu, put a betting market on the wall of the main common room: a paper-based mechanism that let the 18 people who took part trade on a question, with everyone’s stakes posted where all could see them. The first question was about cold fusion. In May 1990, when that price had sunk to 7 percent, Stiegler proposed a replacement: “Xanadu will deliver its product before Premier Deng of China dies.” It opened at about 70 percent. Deng Xiaoping died in February 1997.

How It Died

Autodesk’s patience depended on Autodesk’s stock price. In 1991 the shares traded above $60. By the middle of February 1992, after weak sales and a fight among the company’s managers, they were at $23.50. In April Autodesk hired Carol Bartz, from Sun Microsystems, as chief executive. On 20 August 1992, according to Walker’s own record of the company , Autodesk gave back its stakes in Xanadu and in a second experimental venture and stopped funding both. “They do not fit into Autodesk’s core business,” Bartz said.

Ownership went back to the programmers and the project’s longtime supporters, with Nelson and Gregory together holding about half. They looked for a customer. The copy-shop chain Kinko’s was interested and then wasn’t. The new board fell to fighting over who would run the company.

The last customer was Memex, a startup that licensed the technology to build a record-keeping system for insurance companies, unrelated to anything Xanadu had been designed for. Miller and five other programmers went to work there. The company soon had money trouble, and by Wolf’s account, when it stopped paying its license fees in November 1994, the programmers removed the computers. The company’s founder threatened to sue, the machines went back, and he hired Gregory, who had not gone to work there, to finish the code: 300,000 lines of it. Meanwhile Nelson had gone to Japan, to a visiting professorship at Hokkaido University in Sapporo. He was no longer pitching Xanadu as a product. He was promoting its royalty scheme on its own, and, by the project’s account , working on ideas aimed at the web.

The World Wide Web had come from somewhere else entirely. In March 1989 Tim Berners-Lee, a software engineer at CERN, the European particle physics laboratory near Geneva, wrote a proposal for sharing documents among the lab’s scientists. It used one-way links: a page could point to another page, and the target page would never know. If the target moved or disappeared, the link simply broke. There were no royalties, no versions and no permanent addresses. The software was running at CERN by the end of 1990, and in April 1993 CERN put it in the public domain so anyone could use it. When Autodesk let Xanadu go, the web was running on a few dozen servers, mostly at physics labs. By the time Wolf’s article came out, it was spreading through everything.

Nelson did not take the article well. He called Wolf “Gory Jackal” on his website, wrote a rebuttal, and threatened to sue. Wolf ends the piece with Gregory in a parked car, his head in his hands. Asked why he had stayed with the project for so long, he said: “Total insanity.”

The Autopsy

Xanadu lost to the web, though not the way companies usually lose. The web never took a customer from it, because Xanadu never had a product to take customers from. What the web took was the job. By the time a working Xanadu might have reached the public, the world of connected documents Nelson had described already existed, built on cruder rules, and there was no room left for a second one. Xanadu lost a race it never actually entered.

The obvious objection is that it didn’t need a rival to fail. Xanadu went 29 years without shipping before the web existed. Autodesk dropped it in 1992 because of Autodesk’s own stock price, when the web was a few dozen servers at physics labs. On this reading the web was a bystander, and Xanadu died of perfectionism.

A second objection goes further: Xanadu could never have worked at the scale of the web, so there was no race to lose. Its guarantees, links that never break, a permanent address for every version of every document, a royalty counted on every quotation, all assume that someone keeps a complete and consistent record of the whole network. On a network that nobody owns, that is close to impossible. Berners-Lee’s decision to let links break was not a corner he cut. It was the reason anyone could put up a web server without asking permission, link to anything, and take the server down again without telling anybody. Seen this way, Xanadu was a design for a closed system dressed as a design for an open one. Walker himself said the team had “hyper-warped into the techno-hubris zone.”

The second objection is right about the full design and wrong about the idea. Two-way links and transclusion work perfectly well inside any system that one organization controls. Every Wikipedia page has a “What links here ” button, and Wikipedia’s software pulls shared text into thousands of pages by transclusion , using Nelson’s word for it. The Green server did these things on a single machine in 1988. The two-way link even came back to the web itself, from outside. In 1996 two Stanford graduate students, Larry Page and Sergey Brin, built a search engine that crawled the web and worked out, for every page, which other pages pointed to it: exactly the information the web had thrown away. They called it BackRub , after those back links, and later renamed it Google. What could not be built was all of it at once, for everyone, with the payments attached.

The first objection is closer to the truth, but it explains how Xanadu lost, not what ended it. Slowness alone had not killed the project. For three decades the idea kept finding new backers, Gregory, then Autodesk, then Memex, because nothing else on offer did what it promised. The web changed that. After the web, the pitch for Xanadu could no longer be “connected documents.” It had to be “connected documents done properly,” which is a much harder thing to sell, and Nelson’s own work turned from building a product to criticizing the one that had won.

And the web could win because of a choice Xanadu kept making. When Autodesk’s money arrived, the team threw out the code that ran and started over in a new language, and the four years Autodesk was willing to pay for went into a rewrite that was never finished. Xanadu was not short of time or money in those years. It was short of a version it was willing to let people use. The web won with the one-way link, the part Nelson had specifically designed out.

What Survived

The words. Nelson coined hypertext in 1965, and it is in the name of the protocol every web page travels over, HTTP, the Hypertext Transfer Protocol. Transclusion is still used by programmers for including one piece of content inside another by reference, and Wikipedia’s editors use it every day.

The code. In 1999, according to the project’s history , the company’s board voted to release both servers as open source, under the name Udanax. Nelson kept building smaller versions of his own. In 2014 he put OpenXanadu online, a browser demonstration that shows documents side by side with their quotations visibly connected to their sources. Nelson’s team called it “a working deliverable” .

Mark Miller’s second career. Miller spent the following decades on the security problem Xanadu had raised, how to let programs and people cooperate across a network without trusting each other completely. He designed the E programming language, helped bring security features into JavaScript as a member of the committee that standardizes it, and is chief scientist at the smart-contract company Agoric .

Prediction markets. Hanson’s board on the common-room wall is the earliest market of its kind he knows of. He went on to become an economist at George Mason University and the best-known academic advocate of markets like it, which now trade on elections and interest rates. His first one was a bet on whether Xanadu would ship.

The Ghost Today

The web chose one-way links that are allowed to break, and they break. In 2024 the Pew Research Center tested a sample of web pages that had existed in 2013 and found that 38 percent of them were gone . It also found broken links on 23 percent of news pages and 21 percent of government pages. A 2014 study by the legal scholars Jonathan Zittrain, Kendra Albert and Lawrence Lessig found that half the web links cited in U.S. Supreme Court opinions no longer pointed to the material the justices had cited . Their fix, Perma.cc, is a service that saves a permanent copy of a page at the moment it’s cited. That is one of Xanadu’s rules, rebuilt by libraries, one citation at a time.

Berners-Lee knew. In 1998 he wrote a short page called “Cool URIs don’t change ,” asking the people who ran websites to stop changing their pages’ addresses (URIs, in the web’s jargon). “URIs don’t change: people change them,” he wrote. The web’s answer to broken links was a style guide.

Nelson’s royalty meter has come back too, in a cruder form. When HTTP was revised in 1997, its list of error codes included 402, “Payment Required,” reserved for future use : a placeholder for digital cash that did not exist yet. It sat mostly unused for almost thirty years. In July 2025 Cloudflare, a company that sits between a large share of the world’s websites and their visitors, launched Pay Per Crawl . A publisher sets a price per page. A crawler, a program that downloads pages automatically, gets back a 402 and the price if it arrives without paying, and gets the page if it pays, with Cloudflare passing the money on. The crawlers belong mostly to companies collecting text to train AI models. It is a charge at the moment of retrieval, owed to whoever wrote the thing retrieved, built for machines instead of readers.

Nelson, for his part, never accepted the web’s terms. His summary, from a list of one-liners he posted in 1999 , refers to HTML, the code web pages are written in: “HTML is precisely what we were trying to PREVENT.”