AI: Summary
This session continued the month’s theme of figure and ground applied to reading and thinking in XR: when potentially every digital document is available at once, how do we decide what to bring forward and what to leave as background? Using a long, heavily structured online essay as a shared test case, the group worked across three registers at once — the practical question of how such a piece might be re-presented spatially, the physiological limits of human reading, and the deeper conceptual question of how to make the process of thinking visible rather than only its finished products. Running underneath was a candid thread about the present state of headset adoption and the everyday frictions that decide whether anyone bothers to put one on.
The starting reframe was that figure and ground is not only an interface problem but an artistic one, and that the aim is not to surface a single fixed thing but to let a viewer reach multiple aspects or viewpoints of it on demand — much as someone standing in a historic place might want to assemble several conflicting histories rather than be handed one, or be bombarded by all of them at once.
A recurring idea was a layered model of figure and ground around AI: the live conversation with a model, the record of that conversation, and that record as later examined by a teacher or third party — each a different foreground/background relationship. This connected to the proposal of having students submit their entire AI chat log alongside their work, so the structure of their thinking becomes inspectable. It was framed as making associative trails visible in the spirit of Vannevar Bush, with the historical contrast between figures who kept exhaustive notes and those who left none, where the process turns out to be as important as the product.
The group sat with the provocation of whether writing is thinking or merely the product of thinking — possibly the best medium we have had so far rather than the last one — against the counter-view that the hardening and linearization of writing crystallizes something, as in human generative writing, where you only discover what you mean (or that you were wrong) by setting it down, the way a room is unknowable until it is painted.
A central insight was that decisions cost energy, and that the same interactive branch can be an opportunity or a barrier depending on engagement. Following a rabbit hole you are curious about is nearly effortless, while being made to follow one is taxing; as one phrasing had it, nobody likes to be chased down the rabbit hole. Interface design was reframed as the work of turning options into opportunities rather than barriers, with the related recognition that there is good friction as well as bad, and that supporting the good friction matters.
A notable shift was the move from prediction to conduciveness: rather than chasing an ever-more-accurate system that guesses the right thing at the right moment — which would likely fail — the better aim is to make the medium itself more conducive to cross-pollination across modalities, so that the medium of interaction becomes, in part, a medium of cognition, taking on some of the load the brain’s “dark matter” used to carry.
A subtle distinction emerged between retracing history and retracing understanding: browser histories and chat logs can be retraced, but our own understanding can only be reconstructed in hindsight, because in the act of understanding we are not aware of the artists, ideas, and influences shaping us.
On the physiology of reading, measurement-based work suggested that human vision approximates an A4 page at roughly sixty centimetres, which split the room over whether this is learned from a life of paper or whether, as one participant put it, the paper has learned us. Either way the design consequence is sharp: foveal acuity is tiny, peripheral text necessarily becomes ground, and peripheral motion is highly sensitive — an opening for motion cues at the edges of vision.
Spatial layout drew out a real trade-off. A swivel-chair cylinder or sphere is office-friendly and avoids walking into walls but underuses space, whereas walking among wall-mounted columns or murals intertwines physical and knowledge space through the hippocampus; consistent fixed positions, where chapter one is always on the left, let spatial memory do navigational work. It was noted that the surrounding “universe of papers” is itself a visual cue, not just each page, and that a grid where each row is a chapter may beat a linear arrangement for recall.
Navigation and overview were separated as distinct functions — an outline helps you return to something half-remembered on a second reading, while an overview reveals the spine of an argument and what supports it — alongside an argument for a navigation controller that carries the author’s intended linear flow next to spatially scattered sections.
Concept maps surfaced the comprehensiveness-versus-readability trade-off, likened to game design where accuracy and playability rarely coexist; one map is never enough, and segregated documents and maps that “are not allowed to connect” are a core obstacle to linking knowledge across sources.
A closing thread proposed deconstructing rather than decomposing a text in 3D, with dynamic, personally-contextual hyperlinks that show what a concept means for this particular reader — tempered by the constraint that as long as reading remains left-to-right and top-to-bottom, genuinely novel spatial forms may be limited.
Persistent spatial structure was framed as a memory aid: AI re-summaries make the nodes “explode” and force re-reading, whereas stable positions preserve a reachable mind palace or cathedral, so that a user’s own fragmentation becomes the structure they navigate by. This mapped onto the working term context (acknowledged as a bland name) and onto a Forth-style implicit gaze-stack, where the temporal juxtaposition of what you look at builds context automatically.
On adoption, the realism was that the first “wow” of moving information in space fades within hours, and the real target is “the wow and the work”; micro-frictions like iris-login waits, spinners, and weight quietly gate whether a headset gets used at all, while the field sits in an “XR winter” where VR pays off as a point tool rather than the universal world it was once sold as — context for the decision to ship Author on visionOS and pivot toward promotion.
Privacy was reframed as bidirectional: not only the document data coming in, but the gaze and mood data going out — potentially valuable for long-term self-knowledge yet legitimately worrisome, and currently withheld from developers by Apple.