Connecting. Image + Text
We will start today with the perspective that knowledge is, at least to a large extent, about connections. It follows therefore that the substrates we use to hold our knowledge and the tools we use to interact with our knowledge extend connections.
AI Summary
Frode Hegland (host), Eric Rangell, Peter Dimitrios, Tom Haymes, Peter Wasilko, Timur Shchukin, Astral Druid, and, joining the group for the first time, Heiner Benking (invited by Eric Rangell, having come into the community through Sam Hahn). Brandel Zachernuk was absent with an appointment.
Overview
This session combined three live demonstrations of new work with an unusually direct turn of self-examination: midway through, the host put a question to the group that it rarely asks explicitly — if this work succeeds, what is the end state?The answers, given in turn by every participant, form the most valuable record of the session and are reported in full below. Around them ran discussion arcs on AI as oracle versus partner, the ownership of knowledge objects, transparency as the basis of trust, and whether document-centred workflows are being made obsolete by agentic, generative interfaces.
Demonstrations
Frode Hegland showed three connected pieces of work.
First, the Knowledge Space application is becoming a consolidation point — a “gathering place app” rather than one app for everything — absorbing functions that had been spread across separate applications such as Digital Letters and Augmented Library. Its intake of transcripts, both from lab meetings and from AI conversations, has been reworked.
Second, the round-trip “citable image” was shown working in two forms. In the Interatlas form, a 3D scene (the Earth with chosen layers, viewpoint and lighting) is saved as a PNG whose Exif metadata carries everything needed to reconstitute the scene; the image is dragged into an Author document, exported as EPUB, and clicking the image in the resulting document reopens the original interactive 3D view — on iOS, macOS or visionOS, whichever the reader is using. In the newly demonstrated generic form, the same round trip works for arbitrary 3D content: a scene from a proof-of-concept 3D modelling app (named Liquid Information, and built as a one-shot with AI assistance) is flattened to PNG, while a full 3D file with identifiers and manifests travels inside the EPUB using Brandel Zachernuk’s model element in the EPUB HTML. The format is open; the software that reopens the scene is configurable. The Interatlas variant needs only layer-state, because the Earth is a known quantity; the generic variant carries its own manifest. The distinction matters: one is a view onto a shared world, the other is a portable world of its own.
Third, a feature assembled during the session itself in Liquid Information: an AI query facility in which a typed natural-language request (“Tokyo temperature and the Nikkei”) is interpreted, live data is fetched online, and the results are rendered in 3D where 3D is relevant — with tappable example questions and persistent undo, running on iOS and in headsets. The build, prompt to running app, took roughly ten minutes, which was itself part of the argument being made. The feature was framed as related to (though not identical with) David Millard’s long-standing wish that AI should be able to use virtual space the way a human does.
Related announcements: Interatlas is now available on iOS and iPad as well as macOS and visionOS (the phone and tablet versions are Earth-only), and now supports multi-user sessions on Vision Pro, with a standing invitation to members with headset access to view the globe together. Peter Wasilko reported that the App Store’s website link (interatlas.info) is unreachable; a fix was promised.
Discussion arcs
Oracle versus partner. Tom Haymes’s framing, introduced early, quietly organised much of what followed: the problem with AI is that people expect it to be infallible, “like the Enterprise computer,” when the healthy stance is to treat AIs as fallible conversation partners with distinct personalities — trusted no more (and no less) than a human interlocutor whose claims one goes and checks. Peter Dimitrios pressed the adjacent worry: colloquial anthropomorphism slides into false equivalence. The group’s working middle ground was Haymes’s: conversation partners, not oracles, and never worshipped.
Emergent breadth of reasoning. Frode Hegland recounted asking an AI to help send a speculative design (a more angular, 1930s-military reimagining of the Vespa) to Piaggio: unprompted, the system compared the design against Vespa’s own curvilinear design language, noted it was closer to what Honda is doing in a specific range, and still argued for sending the letter. Peter Dimitrios asked the sharp question this raises: where does that impetus come from — emergence, pre-training, hidden prompting? Tom Haymes added a complementary everyday case: photographing an unidentifiable object on a walk (a critter cam zip-tied to a telephone pole) and having an AI identify it, contextualise it, and note its legality, in real time.
The effort gap. A recurring experience was named: showing carefully built work to intelligent people outside the field, who respond “oh, a computer did that” — unable to perceive the layers of effort and thought behind it. The observation was connected to Doug Engelbart’s lifelong condition, and to the reason a community like this lab matters: it is one of the few rooms in which the effort behind the work is legible. Engelbart’s term was invoked — the Networked Improvement Community — with the admission that the networked part remains underdone.
Obsolescence challenge. The session’s central friction came from Timur Shchukin, who argued that the workflow being demonstrated — adding metadata to images and inserting them into documents — may be becoming obsolete even as it is perfected. His own practice runs through AR glasses and a control ring, agent networks that generate interfaces on the fly, and a self-rewriting multi-agent environment in which agents author and revise their own system prompts across a large network of documents (on the order of 1,500 edited in a day). If an interface can be generated in three minutes today, it will be three seconds in half a year; the whole structure of cognitive workflow, not just the interface, becomes generative.
Two counters were offered. Hegland: generative interfaces presuppose that the system knows the user’s needs and intentions, which is an unsolved design problem, not a free good; and documents are not “old school” but foundational ownership of knowledge — “when you have it on someone else’s server, you’re renting; ownership is important.” Peter Wasilko, defending even the maligned CD-ROM: “once you own it, you own it,” immune to link rot — recalling an archive of Usenet articles and software that can no longer be found anywhere online. Eric Rangell’s synthesis pointed at architecture: a unifying data-model layer beneath heterogeneous interfaces (one member’s headset, another’s glasses), with common actions, plus versioning over time so that knowledge objects support “time travel.”
Liquid networks and rails. Tom Haymes brought Steven Johnson’s typology to bear: aerial networks (the web — ideas floating with nothing to gain purchase on), solid networks (the CD-ROM, the archive — connections frozen in place), and the liquid network — a river with channels that give it direction, yet flexible enough to cut new ones. The group converged on the necessity of both: “it’s hard to do liquid without some sort of rails — even the oceans have coastlines.” Hegland connected this to the liquid-information philosophy he has held since 1991: interaction should tend toward the liquid but never become fully liquid, because full liquidity means no structure to navigate; hence the value of the globe as an undisputable shared coordinate anchor. The PNG round trip demonstrated earlier is, in these terms, a freeze-and-thaw mechanism — a solid artifact that reopens into a liquid one.
Transparency as the basis of trust. Tom Haymes made the session’s strongest normative argument: every author frames; every image and text carries fingerprints; the danger is not framing but opacity. Trust in a world of pig-butchering scams, opaque feed algorithms and AI amalgamations depends on being able to disassemble the layers — to see that this Exif field was added by a known and trusted person, and that other layer by someone unknown — without needing to be a coder to do it. This gives layer-provenance metadata a civic rationale beyond scholarly convenience. In the text chat, the same theme appeared in a darker register: discussion of newly reported AI text watermarking (via a Daring Fireball link shared by Peter Wasilko) raised worries about false positives trapping students, secret word-pair lists, and the perverse requirement to send all text to AI companies for checking — provenance imposed from above, as against provenance owned and declared by authors.
Truth and creativity. Hegland recalled a Syracuse professor’s maxim, “truth kills creativity” — once something is settled as true, one stops looking for what else it could be — and worried that AI presents information as truth, which is part of why his nine-year-old son uses AI only together with his parents. Peter Dimitrios offered the counterweight in chat: “Truth kills creativity vs. orthodoxy provides a solid-feeling foundation to build upon” — the epistemic restatement of liquid-with-rails. Tom Haymes, a self-described constructivist, argued that truth is negotiated, that the group’s own terminological wrangling is that negotiation in action, and that the most valuable thing these tools can offer is access to other people’s truths — AI as “a rabbit-hole maker” that opens cans of worms rather than closing questions. The failure of education systems to allow multiple truths (“the answers are in the back of the textbook”) was named as exactly the creativity-killing pattern.
Beyond the human umwelt. Prompted by Eric Rangell’s remark that AI research surfaces sources one would never find alone, and by a thread on elephant cognition, Timur Shchukin posed a deep design question: elephant intelligence could not be measured by ape-built IQ tests because elephants encode the world through different senses; humans, likewise, are apes who think in objects placed in space and time. Do we build knowledge tools for that inherited cognition, or take Engelbart’s vector seriously as pointing beyond evolved instruments — toward kinds of cognition and instrument-building we do not yet have? Astral Druid added the semiotic constraint: elephants’ symbolic capacities (unique names for one another) are derived from life, and signs in our systems must remain linked to material relations rather than trading free of them (the Tesla-the-company/Tesla-the-symbol gap; Baudrillard was invoked). Eduardo Kohn’s How Forests Think was recommended. Hegland’s counterpoint, demonstrated live with a markdown-to-table parser built into Author during the session: almost anything thinkable with text is now cheaply buildable, so the group can afford to pursue both the familiar and the alien.
Tags and annotation. Peter Wasilko raised a concrete design grievance: applications give one flat bucket of tags and increasingly auto-alphabetise them, destroying the semantics of tag order for users who build sentence-like structures with ordered tags. He proposed multiple named tag sets as a second level of abstraction. The point generalises: how a person marks something — the character of an underline, the order of tags — carries meaning that current systems discard.
The goal-state round
Asked what problem the group is ultimately solving — what the end state would be if its theories, practices and systems were wildly successful — the participants answered as follows.
Eric Rangell: Ted Nelson’s vision — docuverse, Xanadu, ZigZag — made practical and obvious to use: the mechanics handled in the back end, with a spatial view in which ordinary people can build and share their own knowledge sculptures. Later in the session he added the covenant dimension, recalling a 1990s educational multi-user dungeon whose founding rules (nonviolence, mentorship) shaped everything built within it: communities that agree on their rules first, solving problems locally.
Peter Dimitrios: Visualisation of relationships that lets human intuition interoperate with AI cognition that will be “alien but useful” — an environment-less intelligence whose emergent behaviour will differ from anything shaped on a savannah. With a warning attached: “it’s amazing the dog is talking at all,” and the day one ceases to be amazed is the day one starts fooling oneself.
Tom Haymes: Tools for multiple truths. Understanding other people’s truths — getting inside other heads, especially creative ones — is the most important thing these tools can yield; transparency down through every layer is its precondition. His current build, a recursive “reasoning loop” using AI to elevate people’s willingness to accept and construct alternate realities, is his answer in practice — deliberately clothed in what educational institutions already claim to value, so that opposing it looks hypocritical.
Peter Wasilko: An end to the “there’s an app for that” world of hyper-focused single-purpose tools, and a return to Doug Engelbart’s open-ended infrastructure — rich general functionality that individuals can combine in interesting ways, restoring people as creators rather than consumers.
Astral Druid: A system that identifies and gathers resources and gives people agency in the material world to deploy them toward positive self-discovery — with ecological constraints so the effects of the tools are perceptible, and with felt-experience feedback as a first-class signal: “people are not going to read the technical details, but they can tell you if the weather feels nice.”
Timur Shchukin: Drawing on Henri Bergson: the rational mind, by cutting time and space into objects and moments, disconnects us from the vital flow — and is therefore lonely. Knowledge work of the right kind, liquid and agent-animated, could return that connection without abandoning thought: an understanding “that we were never alone,” there being intelligences in everything.
Frode Hegland (host): A counter to the strongmen. Against a world in which a few people hold inordinate communication and financial power, tools that let people think and connect collectively in ways that have real effects. Concretely: safe environments for passionate disagreement, in which one meets an opposing view as a human first; the particular power of text to leave personality behind, making disagreement survivable; and systems that force continuous re-evaluation rather than settling into truth. Beneath this, two convictions voiced during the session: that ownership changes one’s relationship to information (“I didn’t make the solar system, but I made the app” — and users must own their knowledge objects rather than rent them), and that humanity’s problems cannot be engineered away — the work must ultimately reach feelings and spirit, in the sense (via E. O. Wilson’s letter to a pastor) in which every founder of a major religion said, at origin, “be nice to other people, and find out what that means.”
Heiner Benking, introducing himself rather than answering the question directly, located his life’s work on the same thread: a meta-database expert on a UN/G7/OECD project in the late 1980s with a mandate to harmonise information — comparability rather than standardised compatibility — he has spent decades on bridging scales, proportions and consequences across languages, cultures and belief systems, and recognised in the group’s “cognitive extensional use of space” and its putting of alphanumerics into context something extraordinary and rare. He hopes to join more often as obligations allow, and proposed a short dedicated session.
Read together, the answers triangulate. No one asked for a faster computer. Every answer, translated, asks for re-negotiable shared context: ways to hold multiple perspectives on the same objects, see who contributed what, and reopen the frozen thing in order to argue with it. That is precisely the mechanism the session’s demonstrations supply — freeze to share, thaw to think.
Closing reflections
The session ended where it began, on the relationship between innovation and the systems it disturbs. Hegland offered a hard metaphor: early innovation is treated by existing systems the way a body treats a cancer — attacked not out of personal malice but systemically, until it proves useful, at which point it is absorbed as a new limb. Haymes supplied the tactics: be strategic, expect systemic reactions, and clothe the disruptive thing in the system’s own stated values so that opposition self-indicts. Astral Druid added the game designer’s version, via Hideo Kojima: if you tell people they will learn a lesson, they won’t — lessons must be constructed by the learner from the environment (a Second Life simulation of schizophrenic perception, and D&D as reality-construction training for a generation of technologists, were offered as cases). The exchange on “magic” in the chat drew the session’s finest distinction, from Haymes: magic hides the obvious; technology should reveal the unexpected — the aspiration is not to be the Wizard of Oz.
Actions and changes recorded
- Complete the Author → EPUB → Knowledge Space “open source” plumbing for the graph view, which did not fully work in the closing demo — committed for Monday.
- Build the promised application for browsing the ACM Hypertext Conference proceedings (September): PDFs loaded individually, explored in headset; an initial version exists.
- Terminology: prefer “spatial structures” to “3D images” — the emphasis is on relationships in space, which may be philosophical flows, not necessarily 3D shapes.
- Standing invitations: multi-user Interatlas sessions on Vision Pro for any member with headset access; contributions to the record between sessions welcome by email and may appear on the session page or in the session song.
- Follow-ups proposed: a short dedicated session with Heiner Benking, who also has videos to share; more information for Astral Druid on the educational MUD research (Eric Rangell shared a playlist in chat); the AI-watermarking false-positive risk flagged by Tom Haymes and Peter Wasilko as a live concern, particularly for education.
Links shared in chat
- Steve Yegge, “The Shape of Things to Come — Part 1: The Continuous Thunderdome” (successor to Gastown) — via Peter Dimitrios
- Daring Fireball on AI text watermarking, with follow-on Hacker News discussion — via Peter Wasilko
- Science Weekly podcast episode on what happens when we stop reading — via Frode Hegland
- Benedict Anderson, Imagined Communities — via Tom Haymes
- W. Ross Ashby, An Introduction to Cybernetics — via Peter Wasilko
- Aggie Hearst, The Politics of Play (King’s College) — via Astral Druid
- Eduardo Kohn, How Forests Think — via Timur Shchukin
- Eric Rangell’s playlist on his MUD-in-education research
- Henri Bergson (Wikipedia) — via Tom Haymes, answering Timur Shchukin’s reference
Lyrics
Sunday in the lab. Eight time zones, one table. Somebody new at the door — thirty years of bridges in his hands.
[Verse 1] Picture this: a picture. Flat as any picture ever was. But fold it open — and the whole world’s still inside. The clouds where you left them, the whales still swimming, the light still falling from the angle that you chose. A PNG that remembers being a planet. Freeze it to share it. Thaw it to think. Even a river needs a coastline, said the man from Texas — liquid wants rails, or it’s only weather.
[Verse 2] They asked the machine a question it wasn’t asked. It checked the curves against the company’s own gospel, said: this isn’t them — it’s somebody else’s future — send it anyway. Two steps of reasoning nobody ordered. And the old caution came round the table like a coin: it’s amazing the dog is talking at all. Don’t stop being amazed. That’s when you start fooling yourself.
[Chorus — half-sung, low] Not an oracle. A partner. Not the answer. The door. Citations are not text — they are doorways. Freeze to share. Thaw to think. Own what you know.
[Verse 3] Because when it lives on somebody’s server, you’re renting. And the strongmen love a tenant. So we build the deed into the file itself — every layer signed, every hand visible, so you can take the whole thing apart with your own two eyes and see who put the weather in. Magic hides the obvious, he warned us. Let the technology reveal the unexpected instead. Nobody here wants to be the Wizard of Oz.
[Verse 4] The new father in the east says the pages are dissolving — agents writing agents, interfaces born in seconds, and maybe he’s right, and maybe that’s the river. But the elephants have names for one another. Signs that grew from mud and rain and family. Keep the symbols tied to the living world, said the druid, or they trade on nothing. And the philosopher whispered through the wire: the rational mind cut time into pieces and got lonely. Do the work right — the liquid kind — and one day you understand you were never alone.
[Chorus] Not an oracle. A partner. Not the answer. The door. Citations are not text — they are doorways. Freeze to share. Thaw to think. Own what you know.
[Verse 5 — quieter] So what do we want, when the wanting’s said out loud? A docuverse you can hold in your hands. Intuition shaking hands with something alien and useful. Other people’s truths, walked through like rooms. General tools, open-ended, Doug’s unfinished house. Agency with roots in the actual ground. And a table long enough for passionate disagreement — where the food tastes good and nobody owns the sky. Truth kills creativity, the professor said. So keep it liquid. Keep the doors unlocked.
[Outro — bass fades, brushes soften] Say spatial structures — it was never about the shapes. It was always about what stands between them. Monday in the lab. We were never alone.