What’s permanent and what’s not

The book has surveyed, at this point, the alternatives the web displaced (Parts II and III), the way the web’s victory was produced (Part IV), and the recoveries that are underway (Part V). The closing chapter is the assessment. Some of the losses chronicled in the earlier parts are being recovered; some are partially recoverable but not yet recovered; some are not on a trajectory to be recovered and probably will not be. Naming which is which is necessary if the recovery effort is to know what to aim at, and naming which is which is also a way of being honest about the contemporary web and the constraints it operates under. The chapter takes the losses in three categories: what is being recovered, what is partially recoverable but where progress is slow, and what is permanent in the sense that the conditions for recovery do not exist and probably will not.

What is being recovered

Several of the losses this book has chronicled are being meaningfully recovered, in working form, by communities that have made substantial progress on the necessary work.

The personal hypermedia graph. Chapter twenty-four documented the recovery of the working pattern of NoteCards, KMS, and the broader knowledge-base lineage in the form of Roam, Obsidian, Logseq, and the surrounding ecosystem. The recovery is real and is operating at non-trivial scale. The audience is in the millions; the tools are mature; the practice is being adopted in education, in research, in professional work, and in personal use. The pattern that the academic hypermedia community had specified by 1985 has, after thirty-five years of dormancy, returned to consumer software. The recovery is not perfect — the federation between tools is missing, the user base is still narrower than HyperCard’s was, the artifacts’ long-term durability depends on the specific tool — but the core property, that a user can build and maintain a graph of linked notes that supports their thinking, has been restored for the audience that wants it.

Local data ownership. Chapter twenty-one documented the local-first movement. The recovery is more recent than the personal-knowledge-tools recovery and at smaller scale, but it is real. Applications that store the user’s data on the user’s machines, that work offline, that survive the disappearance of their vendors — these exist in production form (Obsidian being the clearest example) and the architectural patterns that support them are maturing in the open-source library ecosystem (Automerge, Yjs, the CRDT research broadly). The recovery is partial in the sense that most consumer software is still cloud-canonical, but the path forward is visible and the necessary technologies are at the point where building local-first applications is a viable choice for new development.

The document with its address visible. Markdown plus static site generators plus the various stable hosting platforms have produced, for the substantial population of people who want them, working tools for publishing documents on the web that they own, that have stable addresses, that can be archived, and that survive their authors’ decisions to stop maintaining them. The static renaissance has been more cultural than technical — most of the underlying technologies are not new — but the cultural turn has been substantial, and the population of people who have chosen this path is growing. The web of personal sites that the platform decade had nearly extinguished has not returned to its 1997 prominence, but it has stopped contracting and has begun to grow again in specific communities.

Federated social communication. Chapter twenty-two documented the fediverse and the broader recovery of federated alternatives to the centralized social platforms. The recovery is the most visible of the recoveries in numerical terms — millions of users across the various ActivityPub services, with continued growth — and is the one that has, in specific moments (the post-Twitter migration of 2022-2023), entered the mainstream conversation. The recovery is incomplete: the user base is much smaller than the platforms’, the user experience is rougher, and the federation is fragile in specific ways. But the recovery has reached a scale at which it is structurally relevant, not just notional, and the trajectory is upward.

These four recoveries together address substantial parts of what the platform decade consolidated. They are not, individually or collectively, a return to a pre-platform web. They are, more modestly, a set of working alternatives for the users who want them, operating alongside the platforms rather than displacing them. The user who in 2026 wants to manage their personal knowledge, keep their data local, publish their writing as documents they own, and participate in federated social networks can do so. The tools work. The communities exist. The economic models, while imperfect, are sustaining the work.

What is partially recoverable

A larger category of losses is recoverable in principle but where progress is slow, the obstacles are real, and the current state is partial in a way that does not currently feel like a stable equilibrium.

Two-way linking and the full citation network. Chapter four (Xanadu) and chapter seven (the academic hypermedia tradition) documented the loss of rich two-way linking. Webmentions, backlinks in personal-knowledge tools, and the various ad-hoc citation systems recover pieces of this, but the broader web’s citation network remains one-way and lossy. The recovery in personal-knowledge tools, where the system manages the link database, is more complete than the recovery in the open web, where the link database is in the publishers’ hands and the publishers do not generally cooperate. A working web-wide two-way linking system would require infrastructure (a notification protocol, a registry, or both) that has not been built and has limited momentum behind building.

Transclusion at the document level. Chapter four also documented Xanadu’s transclusion: the ability to quote a passage of one document inside another by reference, with the quoted passage being the live original. Personal-knowledge tools have recovered this at the within-tool level (Roam’s block references, Logseq’s block embeds, Notion’s synced blocks). The web has not recovered it at the document level. The closest the broader web has is the iframe, which is a degenerate transclusion (no granularity below the document, no integration with the citing document’s text), and the various text-fragment-based partial implementations (URL fragments that target specific passages). True transclusion of the kind Nelson specified would require an infrastructure for stable document addressing, version permanence, and active reference-following that the web does not have. The recovery is plausible but would require significant work.

Versioned, durable references. Chapter four (Xanadu) and chapter fifteen (the technical decisions) documented the loss of version permanence. The Internet Archive, persistent identifier systems, and content-addressable storage (IPFS and similar) provide partial recoveries in specific domains. The broader web’s relationship to its own past is still fragile, and the recovery would require either substantial archival infrastructure that operates as public goods or a structural shift to content-addressable storage. Neither is currently on a trajectory that would change the broader web in the next decade.

End-user authorship at HyperCard’s scale. Chapter five documented HyperCard’s specific contribution: a mass-market tool for end-user authorship, distributed free with the platform, in which any user could make interactive media. The personal-knowledge-tools recovery has recovered some of this for note-taking; Twine and the broader interactive-fiction community have recovered some of it for narrative; the dynamic-document tradition has recovered some of it for technical exposition. None of these has reached HyperCard’s mass-market scale, in part because no current platform ships a HyperCard-equivalent free with every device. The recovery is plausible but would require a platform decision (Apple, Google, Microsoft, or someone) to ship such a tool by default, and no such decision is on the visible horizon.

The federated peer model at the protocol level. Chapter three documented the original internet’s federated peer architecture. Email survives in compromised form; the fediverse is a partial recovery for social networking; various peer-to-peer protocols (IPFS, BitTorrent, the various blockchain-based systems) are partial recoveries for specific use cases. The broader question of whether consumer internet usage can return to a peer model — where any user with a machine on the network is structurally an equal participant — depends on infrastructure decisions (consumer ISPs allowing inbound connections, asymmetric residential bandwidth becoming symmetric, NAT being replaced by IPv6) that are slowly happening but on long timeframes. The recovery is plausible but partial and depends on factors largely outside the control of the recovery’s participants.

The dynamic document for the mass audience. Chapter twenty documented the dynamic-document tradition. The tradition has produced excellent tools for technical audiences (Jupyter, Observable, Mathematica). The mass-audience version — explorable explanations, interactive textbooks, computational essays for general readers — has been demonstrated in specific instances (Bret Victor’s essays, the Distill.pub corpus, various Observable notebooks) but has not become a dominant mode of technical communication for general audiences. The recovery is plausible but would require either an authoring tool with HyperCard’s accessibility (which does not yet exist) or a cultural shift in how technical communication is done (which is slow).

What is permanent

A smaller but real category of losses appears unlikely to be recovered. The reasons are several, and the items in this category should be named with care.

The pre-platform consumer internet’s specific cultural moment. The web of roughly 1996-2005, with its personal pages, geocities neighborhoods, blogs, web rings, and substantial fraction of users who were also authors, was a specific cultural moment. The static renaissance is recovering some of this in pieces; the personal-knowledge-tool community has its own version of it; the fediverse has aspects of it. But the original moment had specific conditions — a small enough user base that conventions could form, a homogeneous enough demographic that those conventions could be inherited, a slow enough pace of platform development that the user culture could keep up — that are not coming back. The web’s user base is now in the billions and the platforms’ development pace is faster than any user culture could absorb. The specific texture of the late-1990s personal web is, structurally, a moment in time rather than an architecture that can be reinstated.

The genuinely peer internet at consumer scale. Most consumer internet connections are, as discussed in chapter three, asymmetric, behind NAT, in violation of the original peer-to-peer architecture. The recovery to a fully peer-to-peer consumer internet would require structural changes to the consumer infrastructure: symmetric residential bandwidth, IPv6 deployment, ISPs that permit inbound connections, operating systems that present running a server as a normal user activity. Each of these is on its own slow trajectory; the cumulative shift to make consumer machines genuine peers is, on current trends, decades away if it happens at all. Most users will, for the foreseeable future, experience the internet as clients to servers operated by other parties.

The pre-advertising web. The web from 1991 to roughly 1995 was not advertising-supported. Most of the content was produced by people who were not paid for it; the economic model was the same as for academic publication and personal correspondence — produced because the producers wanted to produce it, for the audiences they could reach. From 1995 onward, the advertising-supported web grew and now dominates. The various subscription models, paid newsletter platforms, and creator-economy services that have emerged are partial alternatives but have not replaced the advertising-supported web as the dominant model. The structural commitment of the web to advertising as a revenue mechanism is, at this point, deep enough that it is unlikely to be reversed at scale, even though individual users and communities have produced alternatives for specific contexts.

The unification Plan 9 proposed. Chapter fourteen documented Plan 9’s proposition that the operating system, the network, and the user environment should be a single coherent system. The recovery in pieces (FUSE, containers, namespaces) does not constitute a return to the unification. The current computing environment is, structurally, layered and disunified, with each layer having its own abstractions. The Plan 9 unification would require a coordinated redesign across operating systems, networks, and applications that no party has the position or the inclination to undertake. The pieces will continue to be recovered; the whole will not.

The richness Xanadu specified in its full form. Chapter four documented Xanadu’s specifications. Several of them have been recovered (transclusion in narrow tools, version permanence in narrow systems, the rights model not at all). The full system Nelson described — with universal transclusion, universal addressing, integrated rights management, and the full set of properties working together — has not been built and probably will not be. The technical conditions are gradually becoming more favorable (content-addressable storage, cryptographic identity, distributed systems research have all matured), but the political-economic conditions that would support a global infrastructure of the kind Xanadu specified are not in place. The recovery will continue to be partial.

What the recoveries together suggest

The recoveries this book has chronicled, taken together, have a shape. They are, with rare exceptions, recoveries at the application layer of properties that the web’s protocol layer threw away. Local-first software adds, at the application layer, the data-ownership property that the cloud-application model removed. Personal-knowledge tools add, at the application layer, the linked-notes pattern that the web’s flat document model omitted. The fediverse adds, at the application layer, the federation property that the platforms removed. The static renaissance adds, at the application layer, the document-orientation that the application web abandoned.

The protocol layer underneath is not changing. HTTP, URLs, HTML, JavaScript — these are largely what they were in 2010, with incremental improvements. The big changes happening are above the protocols, in the patterns that applications adopt and the communities that adopt the applications. This has implications for what kind of recovery is possible and at what pace.

The application-layer recovery is, in some ways, faster than a protocol-layer recovery would be. The protocols have enormous deployed bases and any protocol-level change requires coordination across millions of operators. Application-level changes can be made by individual application teams and adopted by individual users; the friction of change is lower. The personal-knowledge-tool wave has been able to move from “nobody has heard of this pattern” to “millions of users” in about six years, faster than any protocol-level change would plausibly happen.

The application-layer recovery is, in other ways, more fragile than a protocol-layer recovery would be. Applications can disappear. Companies can fail. Cultural moments can pass. The recovery that depends on a specific set of currently-popular applications is, in a sense, only as durable as those applications. The Markdown-files-in-a-folder pattern is more durable than this because it has been deliberately designed to outlive any specific tool. Many of the other recoveries are less robust on this dimension.

The pattern, then, is: application-layer recovery is the path of least resistance, is producing real progress, and is not as durable in the long term as protocol-layer recovery would be. The recovery that lasts is the recovery that the protocols themselves enforce. The recovery that depends on application-layer adoption is one application failure away from dissipating.

What the recoveries do not suggest is possible

The book has not been arguing that the web should be replaced. The argument has been about what has been lost and what is being recovered, with the implicit position that recovery happens on top of and alongside the web rather than in place of it. The web has been a substantial success as public infrastructure; the costs of replacing it would be enormous; nothing on the current horizon is poised to do so. The recoveries are not, in any case, alternatives to the web. They are uses of the web (or, in Gemini’s case, of an adjacent protocol) that take more seriously some of the properties that the web’s protocols permit but do not require.

The recoveries do not suggest, then, that a counter-revolution is possible. The platforms will continue to dominate consumer internet usage for the foreseeable future. The application web will continue to be the dominant kind of web for the foreseeable future. The advertising-supported business model will continue to be the dominant business model for the foreseeable future. The recoveries operate in the space these dominant patterns leave room for, which is real but is not the majority of consumer internet activity.

The recoveries also do not suggest that the losses can all be made up. Some of what was lost is gone. The HyperCard generation’s specific texture, the pre-platform web’s specific culture, the various small communities that depended on infrastructure that no longer exists — these have, in many cases, been replaced by other things, sometimes worse, sometimes better, but not by the same things. The recoveries should be read as recoveries of properties, not of specific cultural artifacts. The properties can come back; the cultural moments are gone.

The condition of the substrate

The substrate underneath the web — the internet’s protocols, the institutions that maintain them, the operators that run the networks — has held up better than this book’s argument might suggest. The IETF continues to operate. The DNS root continues to function. The major internet protocols (TCP, IP, DNS, BGP, TLS) continue to be developed and improved. The infrastructure has, despite various pressures, remained genuinely federated at the protocol level.

The pressures on the substrate are real and worth being explicit about. Consolidation of infrastructure into a small number of cloud providers has put pressure on the federation at the operational layer. Regulatory capture and the related political pressures on the substrate have been significant. The economic incentives that favor centralized platforms over federated alternatives have shaped what gets built on top of the substrate. The substrate’s continued operation depends on the continued willingness of operators, standards-body participants, and various other stakeholders to maintain it, and the willingness is not guaranteed.

The recoveries in Part V depend, in significant part, on the substrate continuing to be available. Local-first software depends on the user’s machines being able to talk to each other over the network. The fediverse depends on independent servers being able to interoperate over standard protocols. The static renaissance depends on independent hosts being able to serve content to readers. If the substrate’s federation degrades — if independent servers become harder to operate, if independent hosts become harder to find, if independent users become harder to reach — the recoveries will become harder. The substrate is not just a background condition; it is the precondition for the work that the recoveries depend on.

Maintaining the substrate, accordingly, is part of the broader recovery effort. The work is unglamorous: standards-body participation, infrastructure operation, regulatory advocacy, the various low-prestige institutional labor that keeps the internet’s federation alive. The work is also necessary, and the recovery efforts that ignore it are building on conditions they cannot count on.

What the next layer might look like

A speculative conclusion: what might the next decade of the web’s evolution produce, if the recoveries continue on their current trajectories?

The personal-knowledge-tool layer will continue to grow. Obsidian, Logseq, and the other current tools will continue to mature; new tools will emerge; the integration with adjacent layers (AI assistance, web reading tools, dynamic documents) will deepen. The audience for these tools will grow, perhaps to the low tens of millions globally, with the practice becoming a normal feature of intellectual work for the audiences that value it.

The local-first movement will continue to mature. The libraries will become more capable; the user-experience patterns will be refined; more production applications will adopt local-first architectures. The mass-market consumer software will continue to be mostly cloud-canonical, but the local-first alternatives will exist in more categories, and users who care about the property will have more options.

The fediverse will continue to grow, probably more slowly than its participants hope. The user base will reach the low tens of millions; the user experience will improve; the federation properties will mature. The platforms will continue to dominate consumer internet usage, but the federated alternative will be more available and more visible than it currently is.

The static renaissance will continue. The personal-site population will continue to grow; the tooling will continue to mature; the cultural commitment to publishing on your own site will spread further. The application web will continue to dominate, but the document web will hold its position as a stable alternative for the audiences that prefer it.

The dynamic document tradition will produce, eventually, an authoring tool that lowers the floor enough to reach a broader audience. The current tools are still oriented toward technical users; a future tool that brings the dynamic document to non-technical authors is plausible and would be substantial. The pattern is similar to HyperCard’s: a technical breakthrough waiting for the right packaging.

The AI integration will be substantial and difficult to predict. The current state of LLM integration with the various recoveries is preliminary; the next decade will see much more. The effects on what these tools are used for, who uses them, and how they relate to the broader web will be significant. Whether AI integration helps the recovery effort or further consolidates platform power is one of the larger open questions about the next decade.

A careful hope

The book has been about a contest that ended thirty years ago and what has been done in the time since by people trying to recover what the contest’s losers had been working on. The contest was real, the losers had been working on serious things, the recovery is happening, and the recovery is partial in specific ways the book has tried to be honest about.

The through-line the book proposed in the introduction — the document lost to the application, and authorship lost with it — is mostly intact as a description of what happened to the consumer web. The recoveries are partial reversals: in the personal-knowledge domain, the document is back as a first-class object the user authors; in the static-renaissance domain, the document is back as the publishing artifact; in the local-first domain, the user’s data ownership is back; in the fediverse, the federated authorship of public posts is back. Each is partial and each is real. The cumulative effect, if the recoveries continue, will be a web in which the document and the application coexist as serious categories, with users having more choice than they have currently about which kind of experience to have.

This is a modest hope. It is not the hope that the platform decade can be reversed or that the original federated peer internet can be reconstituted at scale. It is the hope that, for the audience that wants a different relationship with the web than the platforms offer, the relationship can be had. The audience is, in 2026, in the tens of millions globally. The audience’s tools are getting better. The audience’s communities are growing. The cumulative effect is not a revolution, but it is a recovery.

The losses the book chronicled were real losses. Many of them are gone. Some are coming back. The book is, in this final balance, neither pessimistic nor triumphal. It is a record of what happened, what has been done since, and what is still possible to do. The next decade will determine which of the still-possible recoveries become real. The work continues.

The book ends here. The chapters have been about people who, in their various ways, tried to build a different version of networked computing than the one we have. They mostly lost. The contest they were part of has been forgotten by most of the people who use the web they lost to. Naming them, naming what they were trying to do, and naming what is being done to recover their work, has been the project of these pages. The systems are mostly gone. The proposals are still available. What the next layer of the web does with the proposals is the open question, and it is, on the evidence of the past decade, an open question that is being taken seriously by enough people that it might be answered.