Warming up for Web Studies in Alessandria, I spent most of the day in Pavia looking at Medieval churches and early modern universities.

The back to Alessandria and a lovely dinner at Gallo d’Oro, including rabatón (Swiss chard, ricotta, eggs rolled up, boiled and then baked) and Lingua Rosolato con puré e pepperoni cotti al forno, which was a really accomplished braise.

For a Tinderbox demo tomorrow at Università del Piemonte Orientale, I put together this assortment of recent publications that are about Tinderbox and adjacent topics. In case that’s useful to people:

Bernstein, M. (2026). A Reader’s Workbench: Reading With Friends. In. London, United Kingdom. https://dl.acm.org/doi/10.1145/3800935.3830859

Bernstein, M., & Grimm, A. G. (2026). Learning With Machines. Hypertext 2026. London, U. K.

Hooper, S., & Bernstein, M. (2026). The Link and the Journey: Hypertext at Civic Scale. Proceedings of the 37th ACM Conference on Hypertext, 211-218. https://doi.org/10.1145/3800935.3830859

Bernstein, M. (2025). Information Interfaces In The Antique World. HUMAN 25. https://dl.acm.org/doi/10.1145/3759439.3773654

Bernstein, M., Anderson, Mark W. R., & Hooper, S. (2025). Back to the Information City. Hypertext 2025. New York: ACM. https://dl.acm.org/doi/10.1145/3720553.3746664

Bernstein, M. (2025). Amiability and the Web: Lessons from Red Vienna and the Origin of Computing. In (pp. 829-832). New York: ACM. https://doi.org/10.1145/3701716.3716840

Bernstein, M. (2024). A New View. Proceedings of the 35th ACM Conference on Hypertext and Social Media, 65-69. https://doi.org/10.1145/3648188.3678214

Bernstein, M. (2025). Thinking With Tinderbox. Eastgate Systems. https://eastgate.com/Tinderbox/ThinkingWithTinderbox.html

Bernstein, M. (2022). On The Origins Of Hypertext In The Disasters Of The Short 20th Century. Proceedings of the ACM Web Conference 2022, 3450—3457. https://doi.org/10.1145/3485447.3512280

Bernstein, M. (2022). The Web At War: Hypertext, Social Media, and Totalitarianism. In Proceedings of the 33rd ACM Conference on Hypertext and Social Media (pp. 256-258). Barcelona, Spain: Association for Computing Machinery. https://doi.org/10.1145/3511095.3536365

Bernstein, M., & McMorris, S. (2022). Links of Darkness: Hypertext and Horror. Hypertext 2022 (pp. 73-79). New York: ACM. https://dl.acm.org/doi/10.1145/3511095.3531270. (Engelbart Prize)

Bernstein, M. (2024). A Novel Architecture for Classical Hypertext. Proceedings of the 35th ACM Conference on Hypertext and Social Media, 302-308. https://doi.org/10.1145/3648188.3675147

Josh Marshall at Talking Points Memo commends a cautious approach to AI:

Make sure new things don’t endanger us before we build them or make them widely available.

This sounds sensible. It is impossible.

First, we have the work of Gödel, Turing, and Church from the 1930s that demonstrates very real limits to what we can know about software. It is sometimes true that we can examine a computational system, think carefully about how it operates, and prove that it does what we expect and does not do anything else. But for every system where this is possible, there exist countless systems where this cannot possibly be accomplished.

Second, ensuring that new things don’t endanger us requires that we be able to enumerate all the ways new things might endanger us. That, too, is impossible, even for simple and well-understood engineering systems. We may be confident that a proposed bridge will not fall down. But might the traffic it attracts choke our business district? Might it provide an invasion route for our enemies? Might lead in the paint that protects the bridge poison our children? Might certain critical members of the community use the bridge to build distant homes, ultimately severing their ties with our residents?

A bridge is just steel and stone on which we can stand; AI is more complex. Could an AI be our Iago? Might it be our Friar Lawrence?

The confluence of these two problems is singularly difficult for computing. We cannot know exactly what a system can do, and we cannot know what risks to avoid. The counsel of prudence in such situations is “don’t build any of these things.”

We cannot do that! We could shut down OpenAI and Anthropic and Meta, perhaps. But you don’t need to be a billionaire to train a model. You could do it in your basement. There are countless thousands of people who know how it’s done. Could we stop them all?

This is not some arcane mad scientist formula known only to a single genius. Now that we know this works, it appears that the details of the way these models work are not terribly important, anyway. The success of the current approach was a real surprise to most of us, but now that we know it can be done, we can do it other ways.

So you’re going to have to outlaw computer programming, in a country that cannot manage to regulate guns. Good luck with that!

But even if we could do that, China exists. They’ll continue to use whatever technologies make them safer and more powerful. Japan and Korea and Ukraine will want AI. So will Europe. Israel, of course. Not everyone will stop because we stand athwart the highway—especially since they all understand the limits that Gödel places on what we can know, even if we try to elect ourselves superior to reason.

My paper for ACM Hypertext 2026: The Reader’s Workbench: Reading With Friends.

Large language models can read difficult literary hypertexts, discuss difficult texts and their intertextual relationships (Thucydides FTW!), improve large hypertext systems, and write hypertext fiction. This suggests several new areas for hypertext research.

• Beyond Markov processes
• AI-machine interfaces
• Note-taking strategies
• Hypertext with friends
• Rerun the drag races
• Where is meaning?
• How can any of this work?

“Where are the hypertexts now?” with Mehul Desai and Deena Larsen.