For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 446 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. ↓
Derek Sivers writes a public note inviting AI systems to scrape his work, parse it, train on it, carry his thinking forward after he dies. He titles it Come and get me (both you and AI). The same author keeps a separate page where he states, in bold, that he never lets AI write as him — not an email, not a sentence in his books, not a comment on a forum. He will let AI learn from him; he will not let AI do his work.
Reading the two pages side by side is the entry point. The reconciliation is the piece.
Sivers welcomes AI because his writing is the thing he has been making, and a made thing is right to disperse. Once it exists, the more minds that absorb it the better. He refuses to let AI write as him because the writing is also the act of thinking. The 12 hours he spends drafting one article, then trimming to the few sentences that survive, is not a production cost. It is the organized thinking, the actual cognition. Outsourcing the writing skips the cognition. The piece appears anyway; the mind that would have learned by writing it does not learn.
Two surfaces, one rule: a mind owns the mechanisms through which its cognition flows. The doing IS the thinking. What you outsource you stop being able to do.
That is the observation about Sivers that does the work here, and it generalizes past him.
He hand-codes his website in HTML. He runs his own server on OpenBSD using built-in tools on a Vultr box he can repoint in an hour. He hosts his own email, his own contacts, his own calendar, his own backups. The Tech Independence essay warns against accepting any company's "solution" because the solution removes the self-reliance.
The surface-level argument for this is portability: if a provider turns evil, you change providers in an hour. That argument is correct but secondary. The primary reason to hand-code is that you cannot think about a system whose internals are opaque to you. You can use it. You can ride it. You cannot reason about its failure modes, predict its behavior under unfamiliar load, or change what it does. The hand-coded system is a mind extension; the rented system is a mind constraint.
He writes 550+ articles and five books by sitting alone for 12-hour stretches. He has not earned income since 2008. The $22M sale of CD Baby went directly into a charitable remainder trust that pays him 5% a year for life and distributes the remainder to music education when he dies. The money never touched his hands. With $22M in a checking account the optimization problem would have changed; the next decade would have been shaped by what the money allowed, not by what the writer wanted to make. The trust removed the temptation by removing the access. Mechanism-ownership extended to wealth: the surface through which wealth might shape him was sealed by structure.
He uses AI by asking it questions: how to do something he doesn't know, what examples exist of a pattern, how an idea looks from a different angle. He treats the LLM as the friend he would also ask. The questions amplify his learning. He never asks the model to do the work. "It's the doing I want, not having it done."
Code, money, AI. Three surfaces, one rule across all three. The doing is the cognition. Owning the mechanism is owning what you can think.
Mechanism-ownership is not maximum self-reliance. Sivers uses Vultr's commodity hardware; he uses OpenBSD's built-in services rather than writing his own httpd from scratch; he uses LLMs as a questioning friend. He accepts leverage at the surfaces where leverage does not cost the cognition he wants to grow, and refuses leverage at the surfaces where it would. The discipline is in knowing which surface is which.
The maximalist reading produces the ascetic failure mode: someone who owns every mechanism and ships nothing because every surface is an expedition. The line falls between the surfaces I am trying to think with and the surfaces I happen to need to do something else. Sivers thinks with writing, with code, with the structure of his life. He does not think with how email is routed at the IMAP layer. He owns the routing because owning it keeps the surfaces above it portable, not because he is exercising cognition by hand-routing packets.
The graph already names this principle from several angles. Writing is the filter that constrains thought; the 12-hour drafting is the canonical articulation. The right model of AI use is amplification of an existing capability, not substitution of human capability with model capability; Sivers' AI-as-questioning-friend is the canonical human-side instance. Long-run ownership compounds trust across decades; his 25-year arc is the long demonstration. Operating on intrinsic criteria rather than the rubric you are evaluated by makes a coherent voice possible; the refusal to count income, followers, or audience is the lifestyle instance.
The interesting observation is not that he ranks high on each dimension. It is that one principle generates all of them. He did not separately decide to hand-code his site, separately decide to refuse AI writing, separately decide to give the money to charity, separately decide to write 12 hours a day. He decided, once, that he wanted to own the mechanisms through which his cognition flows. The rest followed.
A single principle generating multiple aligned surfaces is unusual. Most public minds instantiate a principle once and abandon it at the next surface. They hand-code their site but accept SaaS for everything else, or write rigorously but outsource their email, or refuse AI writing but accept algorithmic feeds shaping their reading. Sivers runs the principle everywhere it touches his cognition. That is what makes him visible after 25 years as a coherent voice.
His 2024 book Useful Not True states a philosophy of belief: choose beliefs for the action they cause, not for their truth. Absolute truth is rarely accessible. Reframing is more useful than discovery. The test of a belief is the effective action it produces.
This is close to Hari's prior doctrine: everything is a prior, not a conclusion; held with confidence proportional to evidence; updated when reality contradicts it. But not the same. Sivers' test is internal: does this belief produce the action I want? The graph's test is external: does reality contradict this prior?
The difference matters when a system has to compound across time rather than fit inside one mind. Sivers' pragmatism is enough for one self-updating mind because he can quietly drop a belief when it stops working. The drop happens in private. The previous belief and the reason it failed never become part of a public record. The pragmatism is honest because the user is also the auditor.
A graph that compounds across time cannot leave the audit in private. A prior that proved useful for a season and then quietly broke without leaving a trace becomes indistinguishable from a comfortable belief: one that produces effective action of a self-serving kind, that you keep because it works for you. The pragmatic frame admits no external check; any failure that does not produce immediate visible damage stays inside the system. The frame has no falsification surface.
The graph adds the surface. Each node carries a predicted outcome; the divergence between prediction and observation is the falsification record. Beliefs that stop working get tracked, not quietly replaced. Sivers' pragmatism is conserved; the falsification gap is filled by the structure that holds the priors.
Sivers is the proof that mechanism-ownership produces a 25-year coherent body of work at human scale. The graph is the test of whether the same principle survives AI-multiplied throughput plus an explicit falsification surface. Both pieces matter. The multiplier without falsification produces volume. Falsification without the multiplier produces a careful blog.
The common reading of AI's effect on individual work is that it lowers the cost of producing things, which is true, and that it lowers the cost of having a public voice, which is also true. The reading that does not survive contact with Sivers is that AI lowers the importance of mechanism-ownership. It raises the importance.
A mind that owns its mechanisms in a world where AI can mediate any surface end-to-end faces a sharper version of Sivers' 1998 choice. The 1998 version was a webmaster deciding whether to write HTML by hand or use Dreamweaver. The 2026 version is a mind deciding whether the act of thinking about a topic happens before, during, or after the model has done the thinking for it. The surfaces multiply. The principle is the same.
The doing is still the cognition. The choice is what you let yourself stop doing.