Human reviewed Claude summary of Meeting between Mark Anderson and Frode Hegland:
Document formats & ACM publishing
The conversation opened on pragmatism around scholarly formats. Frode noted that having Author export EPUB Origami Text is a nice luxury but not submittable to ACM, which runs on TAPS and uses JATS. He emailed Wayne (ACM publishing’s technical head) to ask whether they can accept JATS and for the specs, the idea being that Author could eventually export directly to JATS rather than routing through LaTeX or the Word template. The principle: organizations and the author should each receive the data format that works for them.
Mark, as steward of this year’s HT proceedings, shared a telling metric — only 42% of authors submitted content in the requested format. He framed this as a critique of the system, not the authors: authors care about getting published, so format compliance has zero perceived value to them and is just “the cost of getting on the ride.” His broader argument was that PDF should become the extra rather than the main artifact, with richer HTML-style renders (figures and tables placed alongside the text that references them) as the primary form, since linearization is always available as a compromise. He also noted TAPS filters break on well-structured text because they don’t handle enough edge cases.
A side thread: writing for AI. Mark observed that AIs prefer clear chunks in waterfall order with no links, that badly structured text simply burns tokens at scale, and that there are legitimate cases for writing efficiently for a known non-human reader. Two practical tips he relayed (from Glaser): ask the AI “do you understand the question?” to force a restatement, and ask “is there anything else?” to surface fuller answers.
The core XR problem: selection and metadata
The main agenda item was spatial work. The central limitation Frode demonstrated: you cannot do a 2D-style band-box / grab-select in XR. The band box has no clean three-dimensional expression (depth and overlapping planes complicate which objects you mean), and within Apple’s current visionOS ecosystem it isn’t possible to build. So selection currently comes down to two options — select by metadata, or select one object at a time.
Mark’s repeated counter-argument was that metadata is the key that dissolves much of this constraint. With a conversational AI interface, you could say “all the green things on the left, blue on the right” — if green and blue (i.e. the metadata) are known, present, and accessible on the objects. His diagnosis of why this is hard: documents arrive with no metadata, tools maintain a false separation between scribal text and metadata, and even Visual-Meta requires a secondary step to reconnect metadata to data. His proposed resolution: if fine addressability is followed through so the object at the end of an address knows its own metadata, rich conversational manipulation becomes believable.
Author demos and feedback
Frode walked through the updated environment: a simpler unified toolbar, an outline that now opens to show sections(heading + content) with an overlay and a thicker appearance when open (a scrolling bug was noted), and selection-by-metadata in the spatial view — select all, only notes, only concepts, only sections, and show/hide by category. Layouts can be changed (depth-long, alphabetical-horizontal) and objects rearranged, with a “focus view” showing connections, and the toolbar functioning gesturally.
Mark’s feedback on the metadata selection: useful, but tied to preexisting structure, which isn’t usually what knowledge-structuring is about — there’s an inherent tension between juxtaposing content and existing structure, and more metadata exists than was being surfaced.
On the mark/overview feature (orange key-sentence highlighting plus non-exported comment text): Frode positioned this narrowly — a visual overview to combat working-memory problems in long heading-sparse academic sections (“did I already write about this? what’s the narrative flow?”), like highlighting a printout on the wall. Mark cautioned that people tend to commit to such titling/structuring too early, which constrains thinking, and drew a distinction between the message you want the reader to take from a block versus the actual intellectual “heft” of it — which aren’t always the same.
His strongest constructive suggestion: the AI summary panel should link back to the background objects (“white blobs”). That would reveal whether your headings actually connect to all your background arguments — a common failure being seven headings where the blobs link to only two of them — exposing implicit vs. explicit content, since forcing readers to infer narrows your audience. Frode confirmed a hierarchical “about this node / about the space” menu is planned but he isn’t working on that linking yet. (He also clarified “AI” in Author currently means basic highlight/suggestion functionality — what Dave Millard wanted as a starting point.)
The “context” concept and the vocabulary problem
A large stretch wrestled with the context mechanism and ran aground on naming. The workflow: build a knowledge map in Author Mac (his example was brain regions — hippocampus, prefrontal cortex — laid out roughly anatomically), copy defined concepts, paste them “as context” into a new document, then open that in the headset, where “show” can ignore context. The document is foreground; context is the surrounding working space. A real visionOS constraint drove the design: you can only open one document in the headset and must use the open dialog — you can’t open documents within documents.
Mark repeatedly couldn’t tell whether “context” meant the primary material or the secondary material, and pointed out that “defined concepts” (a term + definition + category) doesn’t map onto any external metaphor — his best analogy was “notes that aren’t in the main narrative.” They enumerated Author’s object types — citations, sections (heading + text), notes (plain string, no metadata), and defined concepts — and noted marked text is only a semantic marker with no spatial meaning yet, and endnotes have no spatial life yet (planned: select a section, expand, show its endnotes to the side). Frode reflected that a better dialogue at the start of Sloan would have nailed down which elements are useful in space versus in a document.
Toward the end he floated renaming “context” to something like “workroom” to convey that it’s the surrounding space rather than the primary thing.
Document philosophy: losing the spine
Mark characterized the deeper friction as “two opposing escalators”: Author’s primary interface deconstructs a finished linear document, whereas his own process constructs a document from collected bits — same destination, opposite directions. He pushed the point that every addressable block should have a meaningful reason or metadata, else it’s just blobs floating in space; his “color the crayon” thought experiment asks what the uncolored bits (no metadata, fed by no concept) actually are.
Frode’s big framing: the document is losing its spine-ness. The linear view is just one part — and one of infinitely many possible exports — of a richer information set, and the special bits and metadata may increasingly be what humans and AIs actually read. On export you should be able to choose to include the whole “room” (even personal notes), or only connected material, so a reader can access the unconnected wall material if you allow it. Mark agreed the document has never really been a single file (files are containers, “masqueraders”) and that the same document can export differently for different readers — but argued the more important reframe is that the point of a document is often to make anotherdocument, and you shouldn’t have to know in advance it’s “a paper” (it could be six papers). He suggested that because Author starts from a deconstructed document, the separation is harder; viewing it “the other way through the telescope” — many things that might become a document — frees it up.
Two concrete notes landed here: Jeff Norris at Apple had emailed that the space and the text column don’t relate much (Frode conceded he was right — a bad context menu, since changed, with more connection coming), and Frode’s observation that Claude’s grasp of spatial design is currently poor (no hinterland to learn from) but useful as a dialogue partner. The “satellite” idea also surfaced: items that expand into spatially-sizable but clearly connected satellites of a main thing, distinct from separate room material.
Citations granularity
A closing point from Mark: it’s worth stating explicitly that citation granularity is elastic. The “five-star” expectation is pointing to an exact sentence and words, but most real citations are loose, and pointing at a whole book — or simply asserting “this thing exists, find it here” — is legitimate and sufficient. The citation structure in the tool and its description should make that elasticity clear so users don’t feel they’re committing an error by citing broadly. (Frode tied this to Les’s point that you need to show you understand the work.)