Reading Log

As We May Think

Vannevar Bush · worrydream.com · published 1945-09-10

read
2026-07-09
tags
hypertext, computing-history, llms, personal-knowledge-management, aesthetics
links
original · archive.org

Before going into the content of the article itself: the ads spaced along the side made me think of broad patterns in design/art/aesthetic history, that being defined by what is possible but unscalable. For example, highly ornate art, buildings, etc. functions as a flex of sorts by being highly unscalable yet possible at any time in the past. Now, with digital imagery & the equivalents when this was published, literal ornateness itself isn’t too difficult but design taste is. Maybe that’s just because I find ‘digital’ aesthetics of all time to be more pleasing than classical aesthetics, as defined by the aim to be as simple as possible as opposed to as complex as possible. I’d characterize much of aesthetic history as following these two modes, in various ways, as per the capabilities of the visual production mediums at the time. I’d say that digital aesthetics seem to have existed in the past. For example: Ellsworth Kelly's Seine, kufic script, or futura.

Obviously the notability of this article is that it, in some sense, predicted aspects of the computer revolution and the World Wide Web, directly inspiring Project Xanadu, an early hypertext project & hypertext broadly as a concept. While Project Xanadu didn’t ever become public software due to feature creep, it is an interesting case study.

On the hypertext, the goal of this project itself, primarily, is to allow me to write in response to the stuff I read online, but I’d imagine it’d have some value as an archival/big data tool. I’d guess it will fail for the same reason QS (quantified self) as a whole broadly did, that being that I can’t really gain anything from it, I probably can in the future find some use of it. I’d suppose that strongly depends on the form this ends up taking, but for now I kept it a broadly minimal MVP.

On the concept of an MVP, LLMs do seem to be reducing that, as once a codebase has been set up, LLM tools like Claude Code at writing are very effective at adding additional things to it effectively and quickly, while they currently lack in Greenfield projects. These incentives strongly revert that of past programming, both in the desktop software release architecture then later SaaS and Mobile OTA update systems, because of the potential to introduce bugs & maintenance overhead, then later the lack of need (when you can update as OTA, just add it later; there’s minimal point to delaying launch).

On LLM capabilities: This project, in the near future (not this version/the MVP but maybe by the next note, haha) intends to use LLMs to tag, categorize, link, etc posts.

I find that LLMs have often taken the form of a reskinned chatbot, such that projects adding AI would just put in a chatbot that can answer questions. An extreme example: VS Code had an extension, I believe either by ChatGPT or Github Copilot, that would just place a ChatGPT window into VS Code, to make coding more convenient. Is giving the chatbot some sort of affordance [re: design of everyday things] to let it just add code directly that hard? Now, it takes the form of an agent. For example, Claude Chat itself is a proto-agent, as it is able to run arbitrary code, see the results, etc. in its “thinking” stages. Maybe the temporal freedom from thinking stages allows this? My hypothesis is that the intelligence of a model directly sets the form & capabilities of the LLM frontier. Capabilities sounds obvious, but take past issues people have had with LLMs: hallucinations, context window issues, lazy coding, excessive refactoring, subtle issues, etc. Since models have gotten smarter, these issues have gotten resolved and prompt engineering, harnesses, etc. have all had little effect outside of the model quality.

What I see as the future of LLMs is as an application substrate instead of something applied on top of an application. For example, LLM structured data could be effective, but generally having the intelligence as part of the system, instead of controlling the system looks like the future.

The proposal in this article requires the human to make connections, which I’d imagine would destroy the majority of the value (why not just write them down? Does reading serially do that much?). Would a Xanadu-like system be possible now with LLMs or future ML development? What about vector databases? Would any form of intelligence applied to the connection problem make it feasible?

The problem with vector databases is that they aren’t necessarily useful for making connections as the value of a connection is a specific similarity despite else wise being dissimilar, whereas vector databases just find broad similarities. Maybe some other vector-like data configuration would work?

The graphic on the last page of the Memex’s workings implies that the Memex has an abundance of space. This implies both that it would be minimally useful outside of a desk, but also that the space needed to store the data is minimal. While modern computers are optimized spatially (eg iPhone Air), this is primarily because batteries are volume-dependent and a limiting factor. Were batteries to not function proportionately to size, I’d imagine that phones would have much more open space & be less spatially optimized. Maybe it also implies that the Memex will be built in the near future and thus will not have the modern spacial configuration (from my perspective) but rather the configuration of computing machines at the time.

On this note: my use of “on x…” certainly has a form of linking structure, which I’d assume I’d want to use to link back to in the future, if I have something about my past comments, implying a value of hyperlinking? On the other hand, it could definitely just be a writing tic, which is both why Xanadu is appealing and I write like this. Like an LLM with a specific writing tic, I might not be aware of a specific area I am appealed to/disappealed to, although the main issue with LLM tics is that one model is all the same, so the “genuinely” tic is consistent across instances & therefore multiple LLM instances can’t catch it. 
That being said, linking back to past notes implies an issue: should the core object of this system be my notes, the articles I read, both, or something else? The need to take a photo of connections in the proposal makes me think of the modern CS solution, that being content-agnostic data (which could be a thing, the connections between them, the user, etc.) Twitter, Substack, and other microblogging platforms typically treat a post, comments, quote-posts, nested comments, etc. as a core object itself with the full affordances that instances above and below replicate. Maybe this would make sense? I suppose that if I don’t write a note on 95% of things I read, which is possible as I read a lot, certain implementations would make much less sense.

I’ll see!

Claude

Summary. Vannevar Bush's 1945 Atlantic essay asks what scientists should turn to after the war and lands on the problem of managing humanity's growing record of knowledge. He surveys emerging technologies (microfilm, dry photography, calculating machines) and proposes the memex: a desk-sized device storing a personal library on microfilm, where the user builds 'associative trails' linking documents together, mimicking how the mind works by association rather than by index. The essay is widely credited as a founding document of hypertext and personal computing, directly influencing Douglas Engelbart and Ted Nelson.

On the note. On your question of whether requiring the human to make connections destroys the value: Bush would say the trail is the writing. He imagined trail-blazing as a new profession, with people selling curated trails through the literature — the externalized associative structure was the product, not a byproduct of reading. That's almost the inverse of your LLM-as-substrate vision, where the connecting intelligence lives in the system. Your critique of vector databases is sharper than you give it credit for: interesting connections are analogies (structural similarity across dissimilar surface content), and cosine similarity over embeddings mostly captures surface content. That's an argument that LLM-mediated linking — where a model can articulate why two things connect — is closer to what you want than any vector-adjacent data structure, since the reason for a link carries most of its value. It also bears on your core-object question: if links come with stated reasons, the link itself becomes a first-class object worth storing, which sidesteps the notes-vs-articles dilemma somewhat — the connective tissue can be the spine even for the 95% of reads you never annotate.

Fact check. The claim that Project Xanadu 'didn't ever become public software' isn't quite right: Xanadu code was released as open source (Udanax) in 1999, and a working demo, OpenXanadu, shipped in 2014. It failed to become a widely used product, but software was published.