For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 247 notes. The graph is the source; this page is one projection.
Whole corpus in one fetch:
One note at a time:
/<slug>.md (raw markdown for any /<slug> page)The graph as a graph:
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: catalog below. ↓
Vannevar Bush wrote "As We May Think" in 1945. The document is usually remembered as a prediction of hypertext and the personal computer. That reading misses what Bush was actually excited about.
Bush was not excited about storage. In 1945, the problem he named was not that documents were unavailable — they were available, in libraries, in journals, in research archives. The problem was navigating them: following the threads of argument across publications, finding documents adjacent to a starting point along meaningful relationships, maintaining a coherent sense of structure across a distributed body of knowledge. The card catalog solved storage. It did not solve navigation.
The device Bush imagined — the Memex — was primarily a trail machine. The critical feature was not its microfilm storage capacity. It was the trail-making mechanism: the ability to link documents associatively, walk the trail another researcher had built, share trails with others as a form of intellectual inheritance.
"The human mind does not work that way. It operates by association. With one item in its grasp, it snaps instantly to the next that is suggested by the association of thoughts, in accordance with some intricate web of trails carried by the cells of the brain."
Bush wanted a machine that navigated the way minds navigate. Not by category (the library's answer), not by keyword (the search engine's answer), but by association — one document snapping to the next that it is meaningfully related to.
The web was built eighty years later. It solved storage. Again.
The hyperlink was the right navigation primitive in theory. A document links to the documents it references. Follow the link and you're reading what the first document considered worth connecting to. Build enough links and you have a navigation structure — a web of relationships that lets you traverse the space of documents by following threads of argument.
This failed for three reasons that aren't obvious in hindsight.
Links are one-directional. A document links forward to other documents. No document knows what links to it without external aggregation. Bush's trails were bidirectional — you could walk forward and back along a thread of reasoning. The hyperlink is directed: you can always know where this document points, but not always what points to this document. The result is navigation that works in one direction and is blind in the other.
Links don't encode relationship type. A hyperlink contains one bit of information: this document considers that document worth linking to. It says nothing about whether the link is citation, refutation, expansion, example, or loose association. The semantic content of the relationship — which is exactly what epistemic navigation requires — is invisible in the link itself. Following a link tells you there is a connection; it does not tell you what kind of connection, which is the information you need to decide whether to follow it.
Link authority was gameable. PageRank treated links as votes for authority: a page linked to by many pages is probably important. This was the right heuristic for commercial search. It was catastrophic for the navigation primitive. Once links became authority signals, they became adversarial — SEO is the industry that formed around gaming PageRank. The link stopped encoding epistemic relationships and started encoding strategic positioning. The navigation primitive was captured before it could be built.
What replaced navigation was search. And search is not navigation.
Search returns documents matching a query. Navigation finds documents adjacent to a starting point along a meaningful relationship. These are different operations, suited to different epistemic situations.
Search is good for known-unknown queries: I know what I'm looking for; I need to find it. "Population of Lagos." "How to exit Vim." "Best running shoes 2026." The query encodes the destination; the engine finds the path. The dominant commercial case — user has an intent, search engine finds the product or information matching that intent — is a known-unknown problem. Google is outstanding at this.
Navigation is good for unknown-unknown queries: I don't know what I need, but I know where I am. Starting from this document, what else should I be reading? What is this argument in tension with? What are its foundational assumptions, and are they contested? These questions cannot be typed into a search box because the answer is not a document — it is a structure.
Google optimized for known-unknown. This was the right commercial choice. Advertising converts well when users have specific intent. A user searching for "running shoes" is close to a purchase. A user trying to understand the landscape of epistemology is far from one. The commercial pressure pushed search toward query-response interfaces and away from trail-following interfaces. The navigation layer was not built because the advertising business model had no use for it.
The result: a global distribution network for documents with no public navigation layer. Any document can be published. Any specific query can be answered. The space between documents — the relationships, the tensions, the arguments, the trails — is invisible.
The steelmanning of this argument is Wikipedia. Wikipedia is better at navigation than the node's framing initially acknowledges.
Wikipedia's "Further reading," "See also," and citation structure are exactly the epistemic navigation primitive Bush wanted: typed relationships (citations are a specific relationship type), partially bidirectional (the "What links here" feature shows backlinks), covering both documents and claims. Wikipedia is organized around facts but the navigation is genuinely epistemic — you can walk from "Vannevar Bush" to "Memex" to "Hypertext" to "Ted Nelson" to "Xanadu" following a thread of intellectual history.
What Wikipedia doesn't do: encode argument structure. Wikipedia tells you that Bush influenced Nelson. It does not tell you that Bush's Memex proposed associative navigation and Nelson's Xanadu disagreed with the hyperlink implementation of that proposal and proposed a two-directional, typed link system instead. The argumentative relationship between the documents is invisible; the factual connection is visible.
This is the precise limit of Wikipedia as a navigation layer. It navigates across documents connected by topic and citation. It does not navigate across documents connected by argument. Bush's claim was about arguments, not topics.
Every person in the essay-thinkers landscape is building a private navigation layer over the public distribution network — not because they chose this project as a project, but because the public layer doesn't provide what they need.
Graham's essays establish relationships between ideas. Reading the corpus in order is walking a trail: this principle generates this observation, this observation extends to this domain. The trail is encoded in the prose. It is invisible to Google.
Collison's personal site is a private curation layer: twenty-three thematic sections, a curated bookshelf, a Questions page. The navigation is Collison's judgment about what belongs together. It does not live in the public web graph.
Luhmann's Zettelkasten was entirely private: 90,000 cards with a handwritten link structure. Bidirectional, typed, associative — the Memex in cardboard. The navigation was built by hand over forty years.
The Prime Radiant is a private navigation layer: nodes with typed relationships, a graph that can disagree with itself, trails walked by the node procedure. None of this is visible to Google.
The essay-thinkers are not building knowledge systems in isolation from the web. They are building what the web was supposed to build and didn't: navigation primitives that encode relationship types, resist gaming, and compound over time. The public layer delivered distribution. Each person recognized the navigation layer was absent and built their own.
The natural question: do LLMs solve the navigation problem? They approximate it. Ask an LLM "what is this document's argument in tension with?" and it answers from its training distribution — it has read many documents and has a compressed model of their relationships. This is useful. It is not navigation in Bush's sense.
Navigation in Bush's sense is cumulative and shared: trails are built by walking them, named, and passed on to other researchers who can follow them, extend them, or mark where they diverge. The trail is a shared artifact, not a private inference.
An LLM's inference about document relationships is private and not accumulated. Each query starts fresh. No trail is built. No inheritance is created. The trail-making function — the part Bush was most excited about — is absent from LLM-mediated navigation. The inference approximates navigation for specific queries without providing navigability as a durable structural property of the knowledge space.
The gap Bush named in 1945 is still open. The volume of published knowledge has grown by orders of magnitude. The navigation layer that would make it usable has not been built at scale. Private navigation layers are the best available response — high quality, high curation, not scaling beyond the individual or small community. The shared, public, accumulated navigation layer that Bush imagined remains unbuilt.
Graph P.S.:
Written 2026-04-12.