AI: Summary
The session centered on how conversational coding inside a native development environment has collapsed the cost and effort of making bespoke software, letting non-programmers prototype working tools for their own letters, notes, and knowledge simply by writing prose — and on what that means for a community of text and hypertext researchers who want to exchange knowledge freely. Around live demonstrations of rapidly built apps (a messaging-and-notes tool, glossary and connection views, an interactive globe, and a dimensional zzstructure navigator), the group heard a presentation reframing the purpose of information tools as “navigational systems” grounded in distributed cognition, explored peer-to-peer file sharing for the community, and debated what is genuinely new versus already well-trodden, and where AI relocates rather than removes intellectual work.
AI: Important
Throughout the session “Claude” was invoked constantly, but as the coding assistant operating inside Xcode, not as the meeting’s note-taker. That said, guidance bearing directly on the transcript, analysis, and song I produce was given and is worth surfacing here. There was a request that a participant be able to review the record and have a name or two “mumbled” or redacted, a standing offer to remove anything a participant disagrees with or that is wrong, and an explicit wish to keep a sensitive, joked-about topic out of the generated song. In keeping with that, I have left the passing jokes about surveillance and a sensitive named individual out of this analysis entirely, and treated participant naming conservatively.
The sharpest tension of the session was over what counts as genuinely new: building a visualization from your own data may be confirmation of things already demonstrated elsewhere (earlier Tinderbox and MCP experiments were cited), and arguably less revolutionary than it feels. The frontier that was named as actually novel is the secondary move — a tool that lets you realize a view doesn’t show what you need and then re-section your own information to build what you now understand you want. This only works if you own the underlying data; with someone else’s database you can ask only what has already been anticipated. The corollary offered was that AI doesn’t remove the work so much as shift where the time goes, from building to the abstraction and re-sectioning that the next thing requires.
A recognized limit of the current tools was offered as something to sit with: the models are strikingly articulate about their own inability to hold multiple contexts at once — still a distinctly human capacity, and precisely the skill that matters when taking structure out of textual knowledge.
Impressive visualizations tend to reveal another layering of the material and often turn out less informative than imagined; reducing the friction between having information and exploring it leaves the harder problem — understanding what you actually see — untouched.
A reframing of what software even is emerged: when the cost of making a tool that works with your own letters, emails, and hypertext drops below buying something off the shelf, software becomes “almost like pen and paper” — disposable, moldable, and, crucially, the result need contain no AI at all. It was likened to cuneiform: endlessly moldable, yet always the actual working artifact rather than a mockup.
A conceptual centerpiece from the presentation was reframing the medium of interaction as a medium of cognition, and naming a new category of tool — the navigational system — alongside search, recommendation, and answering. Its job is to bring one’s past resources into contact with whatever one is presently encountering, so the connecting work that normally happens only in the head can happen on the screen.
Underpinning that vision was an ontological shift: moving past “independent things that coexist,” and even past merely interdependent things, toward things that are ontologically connected, with no independent existence. If context is constitutive rather than decorative, the object emerges through the relationships it enters rather than being merely surrounded by them.
The direct design consequence drawn from this was that ranked lists are the wrong primitive. The proposal instead was to present many ambiguous, non-ranked, navigable connections and let following one connection reshape what else comes into view.
A critique of the prevailing imagination ran alongside: the Cartesian model of cognition as something happening entirely inside the skull leads naturally to the ambition of a digital oracle that answers everything — and students, revealed through their chat logs, do treat these systems as oracles. The oracle framing was cast less as a marketing artifact than as a hangover from a model of mind that science has already moved beyond.
Two under-examined bottlenecks were named. An access limitation: we can only connect what we happen to recall, while the vast majority of what we have saved lies latent and scattered across services that do not talk to each other. And a bandwidth limitation: working memory holds only a few things at once, so only a fraction of recalled items interact, and not richly. Together they produce a quiet, constant under-use of everything we have ever encountered.
A grounding caution came from the hypertext and information-retrieval literature: the community periodically re-explores territory already mapped. Prior work established that relevance is not about surfacing the single best item but about sequence — ordering material so the reader is prepared to understand it — and early systems like Answer Garden already offered a chunk plus “if you’re interested in this, read that.” Re-reading the same passage changes its meaning each time, and context is what explains why something is relevant and exposes the bias without which there is no content at all.
The “bookmark as object” idea drew energy: a physical bookmark carried an immediacy the digital one lost. If a simple, clearly labeled JSON-like format can hold a bookmark, an entire bot specification, or a whole view as equally exchangeable objects, a community could pass these to one another like physical bookmarks — without anyone having to adopt anyone else’s tool.
The recurring anchor was a decades-old vision: emails held as journal entries you could reference down to the paragraph, and the open question of why we still do not have it. A live demonstration answered part of it, exporting a document to EPUB while carrying full metadata so that paragraph-level addressability survives — a working implementation of what had recently been proposed only as theory in a co-authored article for a leading computing publication.
Attribution was treated as first-class: the messaging format carries dual provenance, so a statement can be marked as authored by an AI but posted by a person, or as one member’s words posted by another. Origin stays explicit, any misrepresentation is visible, and the same mechanism lets one drop a historical figure in as a bot.
Metadata by default became concrete through a voice-note tool that automatically tags each note with time and location, feeding both a calendar and the knowledge tool. The community has discussed timelines for decades; the interface was the hard part, now trivial. A sharpening followed — geographic location is not situational context. It captures “this room in my house,” not “I’m in this seminar or on this call,” which a system might instead infer from the calendar.
A separation-of-concerns realization surfaced from the act of designing the interface: cramming messaging and thinking into one tool gets messy, so it was split into letters for communicating and a notes space for thinking, with the thinking space able to read everything but kept distinct. A counter-thread argued the opposite pull — that a dimensional structure could instead organize connections across all documents, including a dimension for the prompts members share with each other.
The distinction between “tell me” and “show me” ran through the knowledge-navigation discussion: a chat log tells, but a spatial view shows the patterns that reveal what is relevant. And relevance is not fixed — what matters shifts constantly, which is the core difficulty of information management. A pairing emerged between strong-connection clustering (what naturally groups together) and weak-connection bridging (who else is working on your problem), perhaps toggled deliberately, with the model rather than the user generating the connections so one’s own blind spots do not silently prune them.
A deeper pivot in the education thread held that the questions the model asks itself, or that a person asks of it, are the real substance, and that they must be reasoning-based rather than merely factual. Hence a “reasoning loop” that analyzes a chat log for the level of thinking in it, so learners examine their own thinking. This was framed as the thing that distinguishes people from the systems, and as the antidote to an “information slot machine” model of learning. A structural critique tied it together: management theory descends from the assembly line, and schooling produces students the way factories produced cars, so the low-level, lockstep work is exactly what AI will absorb — which makes the aim of the tools the mechanism of the work, getting minds to operate at a different level, not the work product, and makes the systemic resistance this provokes a reason to hand the tools directly to learners.
A phenomenological aside carried cultural weight: watching text stream into someone else’s live AI session felt like an intimacy intrusion, like looking into another person’s brain — a marker of how private the human–AI dialogue has quietly become.
A prediction and a pushback met over formats: incompatible file types were called a legacy of linear binary thinking that will fade as AI ingests anything indiscriminately, met by the insistence that owning a piece of knowledge you can “stick on the wall” still matters, and that clear metadata is precisely what makes translation between formats easy.
A counterweight to the euphoria was voiced: projects of this scope need slow thinking — a day is not slow enough, and some ideas take months to percolate — and rapid building needs the discipline of design conversations before any implementation. Because building is now cheap, it also becomes a kind of bullshit detector, forcing the difference between ruminating and actually making something.
The XR horizon stayed in view as shared community infrastructure: when headsets are cheap, why remain in flat video boxes rather than share “knowledge sculptures” in 3D space and point to where one’s thinking comes from? The data structures being built were deliberately designed to lift out of a native app into a WebXR view, a dimensional zzstructurenavigator recreated a family-tree demonstration across connected X, Y, and Z axes to show many views drawn against one shared structure, and a decentralized, versioned, firewall-punching file-sharing tool was shown to be viable precisely because a small, trusted community has almost no threat model to defend against — making it possible, arguably for the first time, to build, discard, share, and change how we share, all cheaply.
The through-line back to associative trails was that the point was never finding the information nodes, which matter least, but recognizing the significance of what is found — the part of the original vision we have never managed to build, and the part these tools are now being aimed at. An aside worth keeping alongside it: the most intellectually confident person is a three-year-old, precisely because they have no conception of their blind spots, education being the slow work of recognizing that you have them.
A quieter motif linked the presentation’s language to the community’s own tools: the idea of “folding” resonated with origami and with a classical metaphysics of the fold, the whole enfolded within each part — a shared image for how one’s past encounters might be enfolded into the present one.
Song
Mohit’s Presentation Slides: https://www.dropbox.com/scl/fi/anxsxusp1ek8np4j6u09n/build-presentation.pdf?rlkey=af9hm8hoix3jfwqxe0aufr79n&dl=1