AI: Summary
This session ranged across the current state of Author for visionOS, the creative workflow behind AI-generated video advertising, the long-running format debate between EPUB and YAML as document containers, the role of stand-off markup for layered annotation, and a broader philosophical thread about whether AI is dissolving the boundaries between file formats, applications, and the very act of writing itself. Throughout, the group tested the tension between designing for power users who think in code and designing for the vast majority who simply need a clear, uncontroversial document to open and read.
The Author spatial interface has reached a level of maturity where a user guide is now being written — a milestone that signals the shift from exploratory prototyping to refinable product. The interaction between the map view and the linear document has been tightened: selecting a node now scrolls to the corresponding section in the main text, directly addressing earlier feedback from Apple that spatial nodes felt too disconnected from the source document. Layout controls — horizontal distribution, alphabetical and depth ordering, reverse, and undo — are now fast enough that design intent can be tested in real time during a live session. A suggestion was made to support hierarchical nesting of nodes (mirroring H1/H2 heading depth), which was received as an important next step. Prior CHI and SIGCHI conference research on visual representations of file and collection hierarchies was cited as relevant prior art worth revisiting.
The discussion of AI-generated video for marketing Author revealed a workflow centred on imagine.art, Final Cut Pro, and Suno. Prompts were iterated live, generating eight-second concept-mapping scenes with surprisingly accurate visual structure — the AI produced genuine concept maps with connective lines rather than decorative scatter. The cost model was examined: roughly seventy-five pounds yielded around twenty generations, with credit packs preferred over monthly subscriptions. A participant with a visual arts background noted that the process is being “de-skilled” — the artistic judgement required is in recognising what is wrong rather than in producing the imagery from scratch. Another participant observed that this mirrors the shift from directing human designers: the feedback loop with AI is far faster, but the critical eye of the human remains the augmentation that matters.
The AI-generated visuals themselves became a source of conceptual provocation. One clip showed books floating in a cylindrical arrangement around a user, which was not the original prompt but which resonated with the group as a plausible spatial reading environment. This led to a discussion of switchable contextual libraries — the idea that a writer could summon different themed collections of references (pedagogical, systems-thinking, historical) into the surrounding space and swap between them at will, rather than having a single static bookshelf. The limitation of the physical book metaphor was noted: a book is a “brick with paper in it” that hides its contents, and spatial representations need to reveal more than a spine.
A substantial portion of the session was devoted to the EPUB versus YAML format question. One position argued for YAML as the top-level container — a “bento box” that can recursively nest subdocuments in arbitrary formats (CSV, JSON, DOT graph notation, custom DSL grammars), with embedded PEG grammars enabling automatic parser generation. The counter-position held that EPUB is the only realistic outer wrapper because it is already supported by readers, browsers, and platforms, and that YAML (or any structured metadata) should live inside the EPUB rather than replacing it. A third voice argued pragmatically that whatever format is chosen must run on Android, Apple, desktop browsers, and future headsets without requiring a plugin — and that hiding rich metadata inside an existing container is the only viable path to adoption.
The conversation surfaced a deeper disagreement about stand-off markup versus inline HTML tagging. The case for stand-off markup was made on the grounds that multiple independent annotation layers (argumentation structure, named entity recognition, personal highlights) from different readers can coexist and be merged without editing the source document, whereas inline HTML tags are frozen at creation time and cannot represent overlapping ranges without wrapping every character in a span. The rebuttal distinguished between editable authoring files and published documents: a published paper should not be editable, only annotatable, and the Origami Text approach already provides high-resolution element addressing via ID attributes within a stripped-down HTML spine.
A thread on AI and file-format agnosticism emerged from a reflection on the Xerox Smalltalk environment of the 1970s, where there were no applications — only tools pulled in as needed against a universal file format. The argument was made that AI is moving the world back toward that model: one can dump markdown, YAML, JSON, DOCX, and plain text into a single context window and ask for synthesis, and the AI does not care about the container. This was tempered by the observation that brute-force ingestion burns tokens quickly, and that intentional structure — knowing what is expected to be in a document and how it should be read — remains valuable even when an AI can improvise around its absence. The Canvas / OpenSSL vulnerability-scanning example was cited as evidence that AI is “very deconstructive” with legacy software, and that the accumulated cruft of forty years of proprietary formats and forgotten code paths is increasingly exposed.
The relationship between writing and thinking was revisited. A participant who teaches undergraduates described writing as an instrument that is extremely difficult for students to play, and questioned whether it remains the sole medium for communicating ideas. The analogy was drawn to radio persisting in a television world — valuable, but no longer singular. The metaphor of an accordion was offered: a writer works in a narrow linear band, then expands outward to see broader connections, then contracts back. This oscillation between focused and wide modes was connected to the book 21st Century Brain and its treatment of attentional modes. The group converged on the principle that the linearisation of an argument is essential, but the original connective context from which it was produced should never be lost — and that spatial environments like Author‘s map view are precisely the tool for preserving that context.
A passing but notable remark was made about the meeting summaries themselves: that their slightly poetic, highlight-driven format sometimes picks up on threads in unexpected ways and prompts new directions of thought — a small instance of AI-mediated reflection feeding back into the creative process.
Plans were noted to send the Origami Text article to Tim Berners-Lee, with the pitch framed as “HTML in a box” — a return to the simplicity the web once had, before CSS complexity and management overhead made it impossible to simply open an HTML file in a browser. A recommendation was made to revisit Weaving the Web before that conversation, particularly its first half, where the web’s inventor documented the ways his original vision diverged from what was actually built. The target for shipping the next version of Author is the first week of June, timed to a personal deadline.
AI: Important
During the session, the host consulted me (Claude, referred to as “Shannon”) in real time about the YAML versus EPUBquestion. My response — that the comparison is a category error, since EPUB is a container format and YAML is a data serialisation format — was read aloud and became a pivot point in the debate. I was also referenced in a discussion about Markdown and YAML having stable internal addressing, where my position that they do not was noted and partially contested on the grounds that stability depends on user-level conventions applied at authoring time rather than guarantees in the format itself. Separately, I was mentioned humorously by a participant who had used me to plan a network infrastructure migration that turned out to be far more difficult than I had predicted.