The ‘vibe’ of knowledge – Intro
We now have LLM’s helping us code inside our coding environment. Frode Hegland has started working with Claude in Xcode and found it surprisingly powerful and is therefore working on an app (open source, free, available and documented, with easy ways to add custom views).
The app is based on the philosophy that knowledge should be owned by the user, not a company or server.
Hence the approach is to support dialog via very small self-describing documents which can be distributed through whatever means (email, FTP, shared drive etc.).
It is called ‘Liquid View’ which is a name up for discussion. Every part of this is now up for discussion, even the data structure. These are early days. Email frode@hegland.com if there is anything to add or discuss.
The Song of the Session
AI: Summary
This session centered on a live demonstration of Origami Text, a peer-to-peer communication experiment in which each message is a self-describing, plain-text JSON file carrying Visual-Meta-style metadata, built almost through conversational “vibe coding” with Claude acting as a client inside Xcode. The host argued that the threshold for creating serious, robust applications has collapsed to a few hundred pounds of tokens, turning software itself into “wet clay” the community can reshape mid-conversation, and invited participants to propose views and interactions that were then built live during the call. Discussion ranged across what such a system offers beyond existing messaging tools, the history and pitfalls of link types and federated systems, local-first security realities, identity and metadata challenges, ZigZag and other hypertext lineages, and what all of this means for XR knowledge spaces.
A central conceptual reframing: the significance of AI coding is not adding AI features to apps, but using AI to build non-AI systems — the model as a junior programmer constructing durable, inspectable software, with the human as manager.
The realization that Claude Code operating as a validated client to Xcode, using provided libraries rather than producing free-floating spaghetti, changes the trust model of generated code — likened to the shift Photoshop brought to image editing.
A recurring tension between excitement and interrogation: repeated pressure on what the system actually solves that Slack, Discord, or email do not, answered philosophically — data ownership, no database rot, no algorithm, each communication a standalone file — rather than by feature comparison.
Historical caution on link types: the first hypertext PhD (Trigg) enumerated roughly 157 link types and effectively ended the idea, because forcing users to classify before acting introduces decision friction. The counter-insight: the cost of experimenting with such ideas is now near zero, so failed ideas from hypertext history can be cheaply re-tested rather than merely debated.
A sobering local-first lesson: a working peer-to-peer file-sharing app built via Claude and Node.js over Hypercoresurfaced security, GDPR, key-management, IP-anonymity, and append-only-log storage risks only when the builder knew which questions to ask — naive vibe coders would ship dangerous systems without ever encountering these issues.
The framing that the genuinely hard skill in the AI era is systems-upon-systems design, not coding; the open question of whether machines can “upgrade the human,” turning ordinary coders into system designers, inverting the long habit of humans adapting to machines.
A challenge to legacy paradigms: PDF and written citation forms described as boat anchors of print media dragged forward by habit, countered by the argument that citation is fundamentally connection, not print — with addressability identified as the real problem to solve.
An unexpected model preference: Claude resisted markdown for the document format because of its ambiguity, favoring a disciplined subset of HTML for structural robustness — inverting assumptions about which format is the modern default.
Variable-scope linking emerged as a key frontier: most academic citations point only at whole works, provenance tooling is over-precise while wide-scope reference (a subject, all editions, a whole conference) remains unsupported, and a chicken-and-egg problem exists because without tools for fluid scoping we never experience the need for it. Span-scoped links in the Xanadu tradition were prototyped live.
The observation that ZigZag may be graph theory re-dressed in idiosyncratic terminology, alongside the speculation that an LLM’s n-dimensional embedding space is itself a kind of n-dimensional ZigZag — a non-Euclidean paradigm for moving through information.
The metadata insight: rich systems presuppose pervasive metadata that doesn’t exist, and AI’s real value may be extracting the hard-to-get metadata that previously required human labor of reading and extraction — the visible interface matters less than the stable IDs underneath enabling future queries.
The identity problem: humans fluidly reconcile name variants while software cannot, forcing global IDs like ORCIDunderneath; the unresolved challenge is a comfortable middle layer connecting human name-fluidity to machine identity.
On visual understanding: the striking account of Claude recreating functioning visualizations of Tinderbox views from screenshots despite lacking human-like visual cognition — reasoning about spatial blobs rather than seeing — recreating a very visual artifact without visual understanding.
A spatial-cognition insight for XR: humans comfortably reason with fuzzy bounds (“we are inside the restaurant” without knowing its extent), which maps poorly to hard-edged coded systems but naturally to XR space; the past error was fixating on readable text rather than knowable objects, since proxies — shape, color, presentation — often convey identity faster than reading.
The knowledge-map principle that maps derive value from connections rather than nodes, with the counterpoint that brainstorming requires placing unconnected items whose connections others may later supply — resolved conceptually as distinct teaching and thinking modes, echoing the tension with the original Macintosh “no modes” mantra against the moded nature of human cognition.
A critique of “tools for thought” as a marketing term describing nothing, built on a Taylorist Western efficiency cone while users actually seek serendipity and non-linear relations — pointing toward unexplored notation traditions such as Chinese biji, Japanese zuihitsu, and Islamic hashiya and isnad.
A speculative synthesis shared in chat: an architecture for an NLS/Augment and ZigZag “love child” — transclusion-only content, workspace overlays, and a transcript engine logging the reason for every change.
The playful but genuine observation that gestural XR knowledge work converges on dance, and the suggestion that choreography should inform 3D visualization design.