AI Overview
This was the first session after the ACM Hypertext conference, so it opened as a debrief. From there it moved through AI in education, AI-assisted development, document formats, and information access. Two newcomers joined partway through: an artist and publishing professional, and a participant arriving from other conferences. The group ended on a practical troubleshooting question.
The conference debrief and the EPUB milestone
The host’s main news was that ACM agreed to publish the full Hypertext proceedings in the Origami Text EPUB format alongside PDF, as an experiment. Getting there took a lot of work. The host started from the Overleaf source zip files, because converting from PDF was a non-starter. They used Claude in Xcode to improve the output over several rounds. A colleague pointed out that the first submission was sent too early and was full of issues. The publisher also asked for some elements to be reordered to match their house style, and the AI handled that quickly. The next step is streamlining this pipeline.
The headset demo was also highlighted. It runs entirely from sidecar JSON, so any document in the format works; nothing is a canned demo. One practical complaint was how awkward the headset’s onboarding is when putting it on someone else.
AI adoption: speed versus depth
A slide from another conference claimed that new technologies are reaching majority adoption faster and faster. The group questioned what “adoption” means. Their analogy was the difference between watching trains at the station and actually travelling somewhere. They argued that AI adoption varies far more in depth than earlier technologies did, from simple fact lookup to sophisticated collaborative use. Someone also noted that headline adoption figures often leave out infrastructure such as network coverage.
AI in education
This was one of the richest threads. The group debated restricting AI for young children. One side held that early years should build unaided cognition, much like the case against calculators, and that paper-based work engages different thinking. The other side argued that bans push use out of sight, and that the real problem is formulaic, drill-based teaching that doesn’t build understanding. Both sides agreed on the underlying principle: a good teacher, and a good AI tutor, helps learners develop answers rather than handing them over.
One educator described a university pilot running this term. It uses layered prompts inside the learning management system to help students analyse their own level of thinking. The plan is to turn it into a plug-in within a year. The design challenges he raised were concrete:
- Privacy law: student data must be fully anonymised before it reaches any model, which points toward local processing.
- Institutional capacity: university IT departments may not be able to run local models.
- Three perspectives at once: the system has to be simple for students, useful for faculty checking whether assignments test higher-order thinking, and safe and scalable for institutions.
A recent ransomware breach of a dominant LMS provider was discussed as evidence of how fragile the current infrastructure is. That prompted the observation that AI-assisted development may make competing systems practical.
AI-assisted development in practice
The host shared several working patterns:
- One-button local models: Reader, Origami Text and Author can now use local models. After the host asked the AI to simplify the setup, it became close to a single button that checks the machine’s RAM and recommends suitable models.
- Letting precedent do the design: the model already knows the world’s precedent, so much of the design can be delegated. The human role is to steer.
- Asking naive questions is fine: there’s no cost to admitting you don’t know something.
- Article-to-code consistency: the host fed a forthcoming article into the coding environment and asked whether the code or the article should change to match the other. The AI then revised the code to fit the article’s principles.
Others reported having an AI build a full course inside the LMS through a browser extension. They also compared models on reasoning quality and noted that cloud tools win on portability across devices. The host offered to help others avoid heavy token spending and plans to test local models for coding.
Formats, interoperability, and the web of documents
During a conference talk, the host asked the AI to find and implement a newly presented protocol, and it did. A side effect was that Origami Text now works as a browser for linked EPUBs. A link to another EPUB opens in the app instead of being sent to a web browser. The host sees this as a web of complete, downloadable documents that don’t depend on servers staying up.
Other participants connected this to related work:
- web archive formats (random-access, locally browsable archives);
- federated wiki;
- local-first approaches that avoid dependence on DNS.
A historical thread traced 1990s format lock-in and the abandoned compound-document architecture of that era. Some argued that AI revives that plug-and-play vision. The host drew a distinction: sharing formats matter more than authoring formats, and PDF’s real failing is its lack of interactivity, not the fact that it can’t be edited.
The two-column academic layout got a lively hearing. It saves paper, and narrow columns suit eye movement, but reflowable formats let readers choose. The host’s position is that column count should depend on the device, the task and the reader’s mood.
Access, guidance and copying
One newcomer is building an accessibility tool that reads web articles aloud to an elderly, visually impaired parent. It includes user-controlled navigation, so that moving to a new page is a conscious choice rather than being driven by an algorithm. This raised a question about how to guide the information access of people we care about without it becoming censorship. The host’s example was deciding what a nine-year-old should see, such as documentary footage of war. A newcomer offered a helpful analogy: this is less like censorship and more like driving licences.
That led into a debate on copyright:
- Is it a restriction on access, or a necessary incentive for creators?
- Most agreed that very long copyright terms and software patents are problematic.
- The group discussed cultural traditions that treat copying as a way of learning and improving, and how an artwork’s value can collapse overnight once it is identified as a forgery.
Information overload
One participant described losing about two hours a day filtering email and feeds. He traced this to the decline of personal websites, the shift to siloed platforms that push notifications into email, and the paywalls on the other side. Another participant said this community itself works as a human curation layer, because a trusted person’s strong recommendation helps prioritise reading. The host invited the first participant to present the problem formally at a future session.
Political discussion as a theme
A substantial part of the session drifted into political territory. It grew out of the access and guidance discussion and developed from there. The meta-discussion that followed was more relevant to the lab. It covered:
- whether tools and text can ever be neutral;
- whether politics is “downstream” of information systems;
- how labels act as shortcuts that end conversations — a point about the limits of text;
- concern that value judgements built into AI models could narrow the range of views users can explore.
The loose agreement was to acknowledge that information tools shape public discourse, while keeping specific political positions out of the sessions and taking those conversations offline.
Several constructive ideas came out of it:
- Scarcity versus abundance: our shared meta-language is economic and assumes scarcity; we haven’t yet learned to speak in terms of abundance.
- Trust shapes building: low-trust environments produce defensive systems, and a single low-trust member can set the norm for a group.
- Open, respectful dialogue: the aim is to maintain the conversation itself, which is arguably the core mission of the future of text.
Closing issue: the EPUB download warning(?)
Several participants saw a security warning when downloading the EPUB. It turned out to come from Zoom’s file sharing, not the file itself. Downloads from the ACM Digital Library showed no warning. A font download pop-up in Apple Books came from the CSS styling that mimics the ACM layout. The files contain no executable code, but one participant warned that how safe a format seems will matter for adoption.
Follow-ups and observations
The concrete threads to pick up are:
- a session presenting the information-overload problem;
- EPUB and federated wiki interoperability work;
- streamlining the proceedings pipeline;
- the local-model coding experiment;
- an update on the accessibility tool.
One participant asked that a line be captured prominently in the summary: “Almost all the ideas we’ve had over the last 50 years, we can now test.” The host immediately added that this also brings a great deal of responsibility.
For facilitation, the political drift began from a legitimate question about guided information access. A light, agreed norm could help keep future sessions productive and let participants with strong but different views stay comfortable. For example: discuss how information systems shape views, but not which views are right.