AI: Summary
This session opened from a deliberately simple premise: AI is an external entity. When code is written with an AI in Xcode, the human is the designer or director, not the coder — and the same holds for knowledge work, where whatever the AI learns, writes, or researches goes into its mind, not ours. We are only augmented when something enters our own brain.
Around that premise the group worked through two linked questions: how do we keep our brains engaged rather than outsourcing the thinking, and what are the right components — tags, annotations, sources, ideas — for building shapes or sculptures of knowledge? The middle of the session became a sustained, genuinely contested examination of tagging: as ordered semantic sequences, as frozen ontologies that trap the tagger in a past paradigm, as waypoints connecting to patterns, and as something AI may render irrelevant altogether. The final stretch turned concrete and collaborative, as a returning participant’s distributed, local-first hypertext system and the Origami Text EPUB effort agreed to attempt interoperability — shared annotations across independent systems, with sovereignty kept in the data itself — in time for the ACM Hypertext conference in two weeks.
The Engelbart–Minsky exchange set the session’s compass: told that computers would be made intelligent, Engelbart asked what would be done for humans. The point now lands differently — there is no longer any doubt that the AI is intelligent, but its intelligence does not transfer by proximity, any more than having a superintelligent friend makes you intelligent. AI legitimately removes clerical rubbish and improves skimming, but the danger zone is the third category: using it to remove the deepening interactions with material we have already decided is worth engaging.
The proposal that AI should be a viewspec — optional, reader-controlled, never forced — reframed generative summarisation as continuous with Engelbart’s original concept: a NotebookLM analysis of a paper is a view on base material in exactly the way a 3D exploded concept map is. An extension imagined AI as layers in a geographic information system, with multiple ongoing prompts materialised as named personas — one commenting on what you forgot to discuss, another griping when you repeat a phrase — where the viewspec becomes the act of toggling combinations of layers on and off.
The deepest disagreement of the session concerned whether tagging has a future at all. One position held that tags imply an ontology, and that ontologies freeze: past me has no idea who future me is, so a tag applied today bounds the universe in a way that may be meaningless or actively misleading later — and worse, letting tags drive you keeps you inside a language paradigm that may be wrong. Tagging was called binary, a relic of the Compute 1.0 era, whereas AI is a connective technology whose associations are fluid; humans think in patterns, not tags, and perhaps the right move is to regenerate connections on demand and record the outcomes as annotations. The counter-position held that manual marks retain something no automatic classification has: a connection to the specific human. What clusters together in the mind of an LLM is not what clusters together in a human mind, so hand tagging and machine tagging should be treated as two distinct ways into the same corpus rather than rivals.
A craft insight about tag mechanics: treating tags as an ordered list rather than an alphabetised set preserves an entire implicit data layer. A sequence like “programmers, working on, databases, concerning, ontologies” reads as entities and relationships — a little indexing phrase — and one participant recounted watching years of such semantically ordered tags destroyed when a software update silently alphabetised them. A companion idea proposed an orthogonal access dimension, such as colour, so the same tag word can be public in one colour and private in another.
Highlighting alone does little for retention; rewriting does. This suggests a combined interface gesture — highlight a passage, then be prompted to write your experience of it at that moment — which serves both recall and future findability, since an underline’s meaning (“brilliant? rubbish? cite this?”) evaporates with time.
A workflow demonstrated the marked-passages principle at book scale: pages of a thousand-page World War Two history flagged with post-it notes were scanned and given to an AI with a layered brief — map these passages against systems thinking, then extract what 1942’s forced institutional learning suggests about the transition to AI. The analysis ran only over the humanly selected material, keeping the human’s judgement as the filter and the AI as the amplifier.
The vocabulary problem emerged as the true bottleneck of shared knowledge: two working groups building drone swarms in the same building — physicists and thermodynamicists — operate with entirely different vocabularies, as do long-married couples who use identical terms with divergent meanings for decades. The distributed-systems approach on offer scopes one vocabulary per knowledge repository and then renders graph views so each community can rapidly see the shape of the other’s vocabulary, hoping convergence follows. The literary anchor was Aircraft Stories and its TSR2 bomber, which is a weapon system, an aircraft, and an engine programme depending on which discipline is looking. The Engelbart echo: there are no weapons in the military, only weapon systems — and augmentation should be understood the same way, as the whole system that surrounds the tool.
The pipeline from tagging to software was articulated cleanly: information gains connections through tags, becomes shareable semantic networks (closer to mind maps than to formal ontologies), formalises into ontologies, and finally supports software artifacts — with human common understanding, not AI capability, as the rate-limiting step throughout.
The magnifying-glass image gave the session its most portable interface metaphor: an ontology refined over time becomes a lens you hold over documents, gilding the few that matter and, within them, the few sentences that matter — and crucially, a lens you can hand to a friend to compare what each of you sees in the same material. Origami Text already knows who its user is, so the user’s own published work can serve as the lens at no additional effort: showing the system your work is telling it what you care about.
The authorship thread asked what it means to be accused — sometimes wrongly — of AI writing, and what readers are owed. The art-school answer: art is the process the artist goes through, and everything produced is a souvenir of that journey. It is therefore legitimate for a reader to ask whether they are engaging a human mind or an LLM mind, since each is read with different allowances and different critical postures. Proposals followed for making the answer visible: a National Treasure-style set of filters separating human, AI, and hybrid text; a three-column view showing original, AI, and fusion; and a new citation-like mark — a symbol distinct from a superscript — declaring “here is the AI behind this claim,” a meta-citation preserving prompt and sources.
The interoperability agreement was the session’s practical high point. Origami Text uses strict EPUB3 with extensive Visual-Meta as JSON and in the HTML, plus the W3C annotation standard; the distributed system keeps identity, keys, and authorship cryptographically in the data itself under local-first principles, so any trusted party can run a sync server without owning anything. The shared dream: a conference session where everyone’s laptops are aware of each other, margin comments on the presenter’s paper propagate live with place and date stamps, the author receives them for the Q&A, and the final day yields a frozen-in-time, annotated knowledge space of the whole proceedings — collaboration, in the borrowed phrase, with no central gravity and no central control. Two independent systems demonstrating this together would triple-prove that sovereignty survives collaboration. Automerge and CRDTs — conflict-free replicated data types, the lineage running from Google Wave to Google Docs — were offered as the mediating layer for near-real-time merge.
A closing arc on spatial design distinguished control from display: reachability matters for controls, visibility for displays, and the iPhone conflated the two on a single surface. A horizontal control surface lying on the physical desk — where a document can be leaned on like a newspaper — was singled out as quietly radical. Type at very different scales does different jobs: art-gallery-sized letters are wayfinding, not reading, and big-versus-small is a related but distinct axis from close-versus-far. Against all of this stood the session’s sharpest provocation, drawn from the Interatlas globe with its satellites and to-scale solar system: the real world has real coordinates — knowledge doesn’t. The question of what should play the role of coordinates in a knowledge space — time, citations, tags, annotations, or the reader’s own lens — was left deliberately open.
On method, the economics of prototyping have changed: a prototype should answer one specific question and be thrown away, and with AI coding partners the cost of discarding work has collapsed — commit, push, and run away — though the discipline remains of never letting speculative instrumentation leak into production architecture, because feature spikes are real whether or not an LLM wrote them.