13 October Mark

AI summary

Showing less, knowing more

On 13 October 2026, the day after a Future Text Lab session, Frode Hegland and Mark Anderson had a long one-to-one call, with Frode working from the quiet upstairs room of a café in town. It began with Mark clarifying that his question the previous day about how references are stored had not been a criticism but an attempt to understand what information was actually available, a thread that came back later in the call. From there the conversation ran across three connected subjects. The first, and the one that took most of the time, was the design of spatial workspaces for knowledge work: how an environment that contains everything in a library, every document, concept and citation, can show almost nothing by default and still let its user reach anything quickly. The second was the relationship between maps and writing, including Mark’s long-held view that a map is for planning rather than composing, and what that means for Author and for the new concept-and-document workspace Frode is building. The third was publishing infrastructure: citation data as ground truth, the Citation Style Language file Mark has built for ACM, and whether academic authoring tools are ready for EPUB, tested live by exporting an EPUB from Google Docs and opening it in Origami Text.

The through-line is a principle Mark returned to repeatedly: don’t decide too early. Don’t name a feature before the design has suggested its own name, don’t force objects into categories before a task needs them, and don’t fix a position for every object in a space when most of them are not being looked at. Frode, who has to build something concrete, pushed back on where that principle meets the screen: at some point something has to be visible, and he wanted to know what. The exchange between those two positions produced the most useful material of the call, including a clearer statement of what the workspace should do (lay out very little; let the user summon, pin and release) and several experiments worth running.

Progressive disclosure and “enclosures”

Frode described where he had got to that morning: all Author documents and all defined concepts now live in one space, but the system cannot render them all at once and, even if it could, the result would be unusable. The design question is therefore what to show, at what level.

Mark’s answer picked up the phrase “progressive disclosure” from the previous day. The old desktop filing-cabinet metaphor feels dated, but if it is loosened into the idea of an enclosure, without a physical picture attached, the two ideas collapse into one: an object’s metadata records which enclosure or enclosures it belongs to, and expanding an enclosure brings its contents forward. He was deliberate about not naming this. If you name a thing first, you end up building something that looks like the name, and the name usually carries legacy meaning. Let the experiment that looks right suggest what it should be called.

Frode noted that Brandel Zachernuk first raised the related idea of seeing more detail as you physically approach an object in a virtual space, and that in his own testing it rarely feels right. In Interatlas, country labels appear as you walk up, but it feels like an old man taking his glasses off to peer at something, and when many similar objects sit together, walking up reveals all of them at once. Mark suggested that only the single nearest object should disclose, and that a better first experiment might use gaze rather than distance, since the headset knows what you are looking at. He also suggested testing disclosure with deliberately meaningless glyphs, so the experiment answers “does looking at a thing and seeing a little more give useful feedback?” without getting lost in “what exactly should it show?” He recalled Peter’s early idea, from the first Sloan-funded days, of a ring of tools appearing around a selected object: neat, but busy. Four widgets work; twenty do not.

Frode’s current stages are: filter by metadata (show only this conference track, only research notes), pull an item out of the wall to reveal it, select it, open it. Disclosure through action, in steps. His concern is that each step must feel smooth rather than like clicking buttons. He also offered to have optical inserts made for the headset so that Mark can see the work properly in XR, since this is the research he most wants to pursue together.

What are the things? Objects, categories and view specs

Frode asked Mark directly: when you put on a headset, what is the stuff you want to move around? New thoughts, your own papers, other people’s papers, concepts? Mark’s answer was that he does not think that way. To him everything is a data object, and what matters is its relevance to the current task, not its type. Tags that can apply to anything are not useful; exclusive categories become long lists or overly prescriptive. What he wants is to concentrate on the stuff relevant to X, and to have a rough sense of how much of the library that is, a sense of “enoughness.” If you expected a quarter of the collection and got five percent, something is wrong: your estimate, your query, or your metadata. When it is the metadata, the right move is to stop, fix it, and come back, which is less work than making the search interface more complex to cope with the problem of the moment. He had written exactly that kind of reply to a user the day before, showing that the “new feature” they wanted was already possible with lesser-known tools, and the user had been pleased rather than put off.

He also argued that virtual objects do not need positions until they are looked at. They could all sit at the same point until summoned; the only thing the system needs to know is what they are. Deciding in advance that this is a document and that is a note, this is blue and that is a triangle, is “making decisions I don’t need to make for the convenience of a certainty I’m not yet using.” This recalled a conversation some years ago about Fabien Benetou’s environment where the two of them had not quite met.

Frode pressed: to build any of this, something has to appear. Once an object is summoned, what should it look like? Mark’s preference is for abstractions. At the first level of zoom, shape and colour are more useful than text, because text costs cognitive load and you usually know roughly what the things are; you decompose when you need detail. His atomic unit is a “note,” a habit from Tinderbox, but a note might hold a chapter, two words, or no words and a pile of metadata. His Future of Text slide of scattered coloured boxes worked because colour came from prototypes, websites blue, protocols yellow, so recolouring two hundred objects is one change. Frode said he already does this: a document’s kind gives it a coloured frame, in 2D and in XR, but only about seven colours are reliably distinguishable. Mark: if you need fifty categories you are already in a different problem. Less is more. “Colour is not a property of the object; it’s a temporary view spec.” Metadata is the substrate; you choose a few things to visualize at a time and change them constantly, keeping positions and changing looks, or keeping looks and rearranging.

Planning is not writing

When Frode asked for a concrete scenario, you are about to write a paper, what is in the room, Mark said he only uses such a space for planning and would never write in it. Maps help most when relationships are not yet clear; once the structure is known, the map falls away. Frode recognized this as the point Mark made ten years ago, that Author is most useful when planning a document rather than writing it, and said it clarified Mark’s position considerably.

Frode’s own framing: the shaping phase is different from writing, and writing should be relatively linear even when it is in connected blocks, a distinction he and Dave Millard drew in their paper on hypertext for thinking versus hypertext for communicating. Author now supports the shaping phase through the map and the writing phase through blocks. The missing third piece is preserving some spatiality for the reader. Shipping a paper with a built-in knowledge map is neat but, as a community, nobody has yet shown it to be useful.

Mark agreed, and found it galling, since intuitively it ought to work. A memory palace is beautiful and deeply meaningful to the person who built it, and not necessarily to a visitor. Early Future of Text books put material on timelines because why wouldn’t you, and the result was bulky and the strands never joined where you expected. Adaptive hypertext tried decomposition twenty years ago and did not go far, because what the recipient takes from a document differs from what the author planned. Knowledge tools are free-form, and the job is to keep them that way for the reader.

On the planning side he described the early stage of a paper as having a boundary but no form: you know you need a bit on this, references for that, a side quest to understand something else. Modular assembly suits this, where a piece may be a sentence, a chapter or a captioned figure, and where an axis of the map could carry order so that chapter four becomes chapter two without cutting and pasting. He had done it the hard way, in LaTeX, when restructuring his Tools of Thought paper. But it depends on the writer: people who write front to back do not need this; people who yo-yo around a document do.

Maps, he added, are less intuitive than they look. Moving from “I can move things around” to doing knowledge work in a map takes practice and a kind of mise en place, populating a category in advance because you know you will need it later. Spatial hypertext kept appearing at the edges of papers at this year’s Hypertext conference. Frode’s maps are in the tradition of Storyspace and early Aquanet, where you draw wherever you like, unlike systems that decide placement for you. And often the map’s job is transient: once the relationship he could not visualize is on screen, he goes back to writing, much like the scribbled paper beside his desk that is thrown away when full. After twenty years he mostly does this in his head, which he takes as evidence of how close maps are to the way we think. Whether a map is 2D or 3D is the wrong question; what matters is being able to see that Tim Berners-Lee sits closer to the three-month report than to something else.

The demonstrations

Frode showed several things live during the call.

Focus mode isolates a selected concept and its connections; leaving focus mode shows the full mess. Mark’s test for the mess: do the positions of all the extra things have salience? If not, they could be a single stack in the corner. If bottom-left versus top-right means something, that is a different game.

The new workspace has concepts controlled from a bar at the top and documents from a bar at the bottom. Each document carries three user-set attributes: type or kind, state (in progress or published), and category (note, proposal, research and so on). The aim in XR is that choosing “research” removes everything the user has not marked as research. Mark said he understood the visualization but the categorization did not sit with him: a finished paper is also research for the next paper. Frode replied that a transcript is a transcript, but conceded the larger point that one document is many things to the same person, and that the language of the system should be careful about the future while being clear about the past. Origami Text already offers a “papers I have written” view for exactly this reason.

Adding a concept: if it exists, the action is a find; if not, it is created and the view moves to show it. Mark raised the collision case, where a hidden object already sits at that spot, and described the Tinderbox approach of nudging into free space to the right or down, which came from Mark Bernstein. Connections in the map are text-based, not drawn; you cannot drag-connect two objects.

Pinning: double-clicking the desktop and defining a concept makes it stay visible, but there is no way to unpin yet. Mark offered “pinned” or “foregrounded” as language, lightly held.

Selecting Brandel Zachernuk brought connected items forward. Mark asked for a query interface, the intersection of Brandel and Peter Dimitrios, and noted that users will need both “show only these” and “show these against everything else,” because there is rarely a right default and anyone doing non-trivial work will quickly want both. Frode said the dialog at the bottom was designed as an LLM interface to the map, and that analysis of the kind Mark was asking for is a different thing. He has been deliberately slow to add AI research inside the environment because he thinks the human should do the research; the ideal is controls that start local, spread outward as far as you want, and come back in.

In Origami Text, reordering sections from the outline strip at the bottom now works without breaking line spacing on export, which had previously been the problem. Mark flagged the edge cases: footnotes, tables, figures.

Frode’s spatial idea coming out of this: what you are making now sits frontmost, and everything that supports it sits behind, so that “reading” can be pulled forward when it matters. Mark connected this to enclosures, an object belonging to several sets at once (people; people at Southampton), and asked whether that is manageable in XR. He thinks maps are where XR shines, since the screen is as large as you want, and primitives like a circle versus a triangle are legible at distances and sizes where text is not.

The library as its own thing

Mark observed that the map was originally designed inside a document and now lives outside one, and that “document” is overloaded: the linear, Word-like thing on one hand, a bundle of concepts and metadata on the other. Author began as an elegant form of word processing; the maps came later; and there is now a reference manager in all but name shared between Author and Origami Text. His suggestion was to let the library become a peer of Reader and Author, so a user can simply see their library without first opening a document to write or read. There is also a scaling question: when a library grows from 500 to 5,000 items, an Author document should not be polling the whole database. Frode noted there are two libraries today, Author’s and Origami Text’s (PDFs, EPUBs, and references to books on paper), and that Reader’s remains PDF-only. Mark’s reading was that the most developed thinking is in Origami Text and should be folded back into the other tools. Frode also showed the citations across all his documents rendered either as lists or as computed shapes, still highly experimental.

Citation data as ground truth

Mark’s working reference database holds about 2,900 papers, around 60 percent of which he has needed only once but categorized properly so he never has to again. A separate Hypertext conference corpus runs to roughly 2,500 items, about 770 of which (workshops, keynotes) were never part of proceedings and so cannot be cited. His example was Halasz’s “Seven Issues,” which had three versions; the middle one was a conference keynote never published, and a librarian at Texas A&M happened to hold a quasi-transcript that Mark has put back into the public record, but there is no process for ACM to add it to the digital library. Scott McCloud’s Hypertext keynote is in the same position, and McCloud himself told Mark he no longer has any record of it.

Since the conference Mark has used Claude to build a CSL file (Citation Style Language, the open XML format that Zotero and other reference managers use to render references) for ACM. The existing one assumes everything is a journal paper. Zotero cannot hold all the metadata ACM wants, so he is working within what it can hold. His conclusion is that the whole thing needs flipping: the citation data is the ground truth, rendering is whatever you like, and if you do not have the data you do not have the reference. This was what lay behind his question about how references are stored. An EPUB, he argued, should carry its BibTeX. Frode confirmed that the Origami EPUB specification’s preamble declares the BibTeX conventions used, such as name order.

Mark also argued that legacy conventions, lowercase three-letter months, abbreviated journal names, solved typing problems we no longer have, “a buggy whip holder in a brand new car,” and that information should pass through a one-way gate rather than dragging that cruft forward. Names remain hard: family name, given names, suffixes, particles like “de” or “von,” Chinese names with family name first and anglicized names in brackets. The format is only the transit envelope; the receiving system still has to know how to unpack it, and no citation standard exists. Word makes documents deliberately hard to work with; Mendeley is no longer a serious tool; Zotero is the main multi-platform player; Bookends is next on his list to try. Why would anyone starting today choose ACM’s current format? He would not. And the reason people do not bother with citation data is simply that it is so hard.

EPUB and the road from HTML

An EPUB is XHTML underneath. Publishers live in JATS, which is XML and deliberately hard-edged; the web was built to be forgiving. Mark’s proposal, refined with Claude, is a server-side service that takes HTML in, returns XHTML (or asks the author for clarification in the rare cases it cannot), and then produces an EPUB. If ACM said tomorrow that EPUB is the format, readiness is low, because few authors know what an EPUB is. Origami Text does the heavy lifting, but it is “too narrow a needle” for mass adoption on its own.

Frode agreed: more import is needed, a proper community-based open authoring tool is needed, a Windows port is under way, and alongside the plumbing there must be interactions and visualizations compelling enough to persuade people to try it. The benefit must be demonstrable. He recalled his advertising professor’s half-joke that the benefit of Author is that you can go home earlier on a Friday. Within five years someone will ship a much better academic writing tool, and the opportunity is to work out what the baseline should be. Mark added that the friction of adopting a new tool is always underestimated. They agreed the converter might belong in a separate app or service rather than overloading Origami Text as a reader, and Frode proposed a recorded session on what it would involve.

The live test: Google Docs exports EPUB. Frode built a test document with headings, a table, an image and an equation (Google Docs has an equation tool), exported it, and opened it in Apple Books, which applied its own styling, and in Origami Text, where the scroll view kept the source font because that view follows the document, while the horizontal view is themed. The styling was a single line of basic CSS. Headings came through; tables and images did less well on the first pass. Mark’s points: semantic structure matters so that styles can be swapped; for formulae he would suggest inline SVG (scalable, no external asset) over MathJax, with MathML if interactivity is wanted; and a checker could steer first-time authors toward the cleanest route. On ACM’s two-column layout, he believes it survives only because the publications office and the conferences each think the other wants it; the journals have dropped it. More broadly, PDFs full of non-breaking spaces and layout cruft are bad for the machine reading that is clearly coming. Better not to put garbage in the water than to filter it out afterwards, and continuing to produce PDFs puts us “on the naughty step.”

Working with AI

Several threads about AI-assisted work ran through the call. Mark proposed a reverse experiment: take a finished paper, Frode’s spatial hypertext paper would make a tight canvas, have Claude decompose it into its parts and relationships in map form and choose its own categories, then compare with the categories Frode already uses. It is “victimless”: it might confirm them, suggest more, or show that categories are contextual, and it would also expose the problem with a category like “person,” where someone central to a paper and someone mentioned once in passing are very different things.

He described a contact from the Tinderbox forums who, without being a programmer, has built with Claude a browser clipping tool that records the provenance of highlights and notes, pushes them through MCP into Obsidian and Tinderbox, monitors his own data exhaust, and produces reports distinguishing his intellectual input from the AI’s and feature work from fixes. It is open source, kept in git, and has been rebuilt from scratch three times to shed accumulated cruft. Mark drew the line between vibe coding, where you ask and accept, and a human-in-the-loop style where you keep asking “what did we just make?” and compare the answer with what you thought you were making. He connected this to Dave Millard’s framing: not dogmatic human-in-the-loop, but wanting to be asked before certain things enter your system, as quality control.

Two further suggestions for Frode: have Claude explain the workspace’s concepts and buttons to a third party, as a check on whether Claude’s understanding matches Frode’s intent (a mismatch is a tell, not a fault); and think about where design intent is captured, since Claude logs what it changed but not necessarily why. He also observed that AI sees the whole context at once and struggles to serialize it for humans, presenting things thematically rather than in page order and drifting back to that habit even when asked otherwise.

Takeaways

  • Lay out very little. Everything else should be accessible, not visible.
  • Visual properties are view specs, not properties of objects, and only a few should be active at a time.
  • The map is for planning. Writing is linear or block-based. Spatiality for the reader is still an open problem.
  • Experiments worth running: gaze-based disclosure; disclosure with meaningless glyphs; reverse-decomposition of a finished paper; Claude explaining the interface to a third party.
  • Needed in the workspace: unpin, collision avoidance, query by intersection, and both “only these” and “these against everything” views.
  • The library should become a peer of Author and Reader.
  • Citation data is the ground truth; rendering is downstream. EPUBs should carry their BibTeX.
  • An HTML-to-XHTML-to-EPUB converter should live outside Origami Text, as a lightweight service, so that early mistakes happen there and Origami Text stays clean.

Leave a comment

Your email address will not be published. Required fields are marked *