Gwern retrospective: Xanadu's range transclusions solved a problem few texts have

A party full of Xanadu machines

A retrospective published on gwern.net and surfaced on Hacker News's front page revisits Project Xanadu, Ted Nelson's decades-long hypertext project, and argues it never matured for reasons that had little to do with bad luck: too little design iteration, too few meaningful use cases, and a core concept that failed a basic test of practicality. The author, who writes as Gwern, grounds the essay in a December 2024 visit to San Francisco for a party celebrating the 50th anniversary of Nelson's 1974 manifesto Computer Lib/Dream Machines.

The host, identified only as James, had assembled an extensive collection of Nelson memorabilia: copies of the oversized book, Nelson's later and more obscure 1997 work The Future of Information, mimeograph magazines from his Swarthmore College days, and most impressively several vintage computers running different Xanadu implementations. Nelson, 87 at the time, could not attend, but a former Xanadu programmer from the Autodesk era, which began around 1988, was there and shared war stories.

When a compile took a week

Those stories, Gwern writes, put the project's slow progress in perspective. The team prototyped Xanadu in Smalltalk — a productive choice whose object-oriented model suited hypermedia well — but then had to cross-compile to C++, a build that took roughly a week to complete. Storage was equally punishing: according to the essay, a mid-range PC around 1990 might carry a 50 MB hard drive, while the finished Gwern.net site today weighs in at 221,438 MB, and a meaningful fraction of its individual hosted documents exceed that old drive's total capacity.

Gwern admits he had wondered how the Autodesk years produced so little for multiple full-time man-years, given that client-side transclusion and popup features on his own site typically took developer Said Achmiz a few days each. Hearing firsthand accounts of week-long compiles and constant friction with C++ reframed the question: the compute poverty of the era alone made productive development difficult.

The demo nobody could read

The working machines were themselves a small miracle, but party-goers largely bounced off them. The interface was too alien to explore unguided, running on operating systems most attendees had not touched in 25 years, if ever. Gwern's key realization came from simply watching the screen: the classic Xanadu demo — the opening of the Book of Genesis, with lines zig-zagging off to the right to mark transclusions and commentary — struck him as, in his words, "completely unreadable." The crisscrossing lines turned the display into clutter, none of the sidenotes were legible on the small screen, and at many scroll positions nothing at all could be read. He adds that he would not want that layout even on a 4K monitor.

A solution in search of a problem

The deeper conclusion is that the interface was not the real failure, and neither was the team. The whole concept of side-by-side range transclusions, Gwern argues, is a solution in search of a problem. Line-level commentary fits the Book of Genesis, with its dense scholarly apparatus and debates over redaction from contradictory sources, yet Nelson's own "Examples of Parallel Documents" slide quickly from textual criticism (the Bible, Hamlet, Rashomon) to a stretched case about the Virginia Declaration of Rights influencing the US Bill of Rights, and then to dubious ones: lists of saints, telephone calls inside an organization, comedy TV episodes.

Hardly any text needs to be fisked line by line, Gwern contends; most citations point at a work as a whole, and even the worthwhile minority of comments resist any obvious ordering. He notes that his own site supports sidenotes, range transcludes and bidirectional backlinks — used in his analysis of the short story "Suzanne Delage" — and that they mostly go unused, because few readers or writers actually need them.

Echoes of Cyc

The essay also endorses a comparison credited to Yuxi Liu: Xanadu resembles Douglas Lenat's Cyc, another long-running boil-the-ocean project tied to databases and AI, led by a charismatic figure who never changed his mind, averse to open source, and showing similar signs of what Gwern calls pathological science. Along the way it sketches Nelson himself — someone at the party recalled he had dreamed of directing Hollywood films and regretted going into technology instead — which Gwern says explains an auteur approach to software, including Nelson's insistence, fifty years on, that he is "not a programmer," and his fondness for film-editing metaphors such as edit decision lists.

Why it matters

Xanadu is the standard counterfactual to the web: the richer, two-way, transclusion-native hypertext that lost to a far simpler design. Gwern's retrospective argues the loss was not bad luck. A compelling vision pursued without iteration, without pressure from real use cases, and without willingness to compromise on scope can absorb decades of skilled work and still ship nothing usable — a caution that applies well beyond hypertext, to any ambitious system tempted to design everything before releasing anything.