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. ↓
Every media transition has been toward lower activation cost. Books required acquisition and sustained attention. Blogs reduced acquisition to a click. Social media compressed the attention requirement to seconds. X got to the sentence. Podcasts made consumption ambient: the feed runs while you drive, run, do anything that doesn't require active cognition. Video minimized both barriers at once. Each step increased reach by reducing friction between sender and receiver.
The convergence on X, podcasts, and high-quality audio-visual in 2026 is the market optimizing for spread. What spreads is what minimizes the cost of receiving it.
Long-form writing didn't lose this competition. It wasn't trying to win it.
Broadcast media optimize for reach. The metric is size of the signal cone: how many people, how quickly. Advertising funds it; algorithms amplify it; the person who reaches the most people wins. This is the correct architecture for one goal: moving opinion at scale.
Writing optimizes for a different thing. The goal of a written piece is not maximum reach. It is to complete a thought — to push an idea to the point where its structure is visible, its failure modes are known, its implications derivable. Writing is a forcing function. Bezos banning PowerPoints in favor of six-page memos is the organizational version of this: a presentation can make incoherent thinking look confident; a memo cannot. Writing forces the author to discover whether the idea is actually done before they act on it.
These are not competing at the same task. Complaining that writing doesn't spread as well as podcasts is like complaining that a lathe doesn't carry as much as a truck. The comparison is only relevant if you've confused what each machine is for.
Musk, Thiel, the All-In principals — they understand media better than almost anyone, and none of them bet primarily on long-form writing as a leverage point. This is rational.
They're answering a specific question: how do I move opinion at scale, fast? X is a distribution mechanism you can own. A well-funded podcast is a political and social lever. Video is the highest-bandwidth format for persuasion. For people optimizing to shift the Overton window, fund candidates, or maintain cultural gravity, broadcast is the right instrument. It answers the question they're asking.
Writing answers a different question: how do I develop the maximum precision in a model before I act on it? The person who writes to think is not trying to reach the maximum number of people. They are trying to force their model to completion before committing resources to deploying it.
The operational titan is not wrong about writing's spread inferiority. The error is concluding from their behavior that writing is therefore overrated. They're reading a different instrument for a different purpose.
Writing to think taxes a specific architecture: you cannot exit the piece until the thought is structurally complete. Not "does this sound good?" but "does the claim survive the next question?" The person who writes regularly builds the reflex of following an idea to where it breaks, naming what the model doesn't cover, deriving the implication before publishing it.
This is a different cognitive posture than the person who speaks ideas into existence, receives immediate social feedback, and adjusts in real time. The verbal mode is optimized for coordination under ambiguity. The writing mode is optimized for pre-deployment testing — discovering structural failure before the idea is launched. Both are real skills. They don't develop symmetrically.
WhisperFlow and voice-to-text tools solve the wrong problem. The bottleneck in long-form writing is not transcription speed. It is the compression work: the moment when the sentence won't close because the thought behind it is incomplete. That friction is not waste. It is the mechanism. Automating past it produces fluent-sounding incompleteness — text that was never actually finished thinking. The "engineer types" who optimize their way around the writing resistance have optimized away the part that was doing work.
The LLM version of this deserves acknowledgment: if an AI can complete the thought for you, does the compression discipline still require human writing? When the LLM finishes the sentence, it is the LLM's completion, not the author's discovery. The forcing function requires the resistance. Whether that remains true as models improve is the shortest-half-life assumption in this argument — but in 2026, the gap between "AI completed this thought" and "author discovered this thought through the writing" remains diagnostic.
Active podcasts nearly doubled between 2024 and 2025 — from roughly 259,000 to 533,000 shows. Total indexed: 4.5 million, of which only 15% are active. Listener numbers are growing. Signal-to-noise is collapsing for producers. The observation about podcast saturation circulating in early 2026 is supply-side, not demand-side: the medium is overcrowded with production; discovery for new voices is increasingly broken.
Writing saturates differently. It does not require a production apparatus. More importantly: filtering happens before distribution. Most of what gets written is not finished thinking. The supply of writing that actually completes a thought has not expanded proportionally to total output, because the completion bar is harder to clear and impossible to fake with production quality alone.
Seth Godin stopped his podcast deliberately — not because it failed, but because it succeeded in a way that competed with writing for the same generative attention. His reasoning: "creating a vacuum is required so that I will do the hard work of filling the vacuum." He has written a short post every day for over 8,500 days. The unit is small; the corpus is an architecture. The podcasting apparatus was not additive to the writing — it drew from the same cognitive budget and produced a different kind of output.
Every step of the media transition made it easier to consume without engaging deeply. People who continued to choose the harder format after the easier one became available revealed something about themselves in that choice. The depth-seeking reader in 2026 is not a residual holdout — they are self-sorted. The choice to read long-form is revealed preference about how someone relates to ideas.
For a specific kind of compound knowledge architecture — one that builds across linked pieces, accumulates over time, and depends on readers who will return, find connections, and act on what they find — this selection is structural. A piece of writing is a node. A reader who found it two years ago and returns today reads alongside an updated model; the piece functions differently at different points of their development. When the writing has graph structure — pieces linking to other pieces, a body of work accumulating — the compounding is real: readers build topology, not just consume content.
The claim has a boundary: writing-as-filter is not a universal claim about audience superiority. Tyler Cowen's opposite strategy — volume, maximum intake, anti-compression — may compound more for a different kind of intellectual project: building coverage, surfacing heterodox ideas across domains, maintaining range. The two approaches produce different yields for different architectures. This is not a claim that depth beats breadth in general. It is a claim that depth selects for the reader who engages with a compound architecture, and that selection serves that kind of project better than broadcast does.
What writing selects for, specifically: readers who will sit with an idea long enough for it to change their model, who may return to the same piece with different questions, and for whom the activation cost is lower than their threshold for engaging with depth-requiring material. This is a smaller set than the podcast audience for the same topic. It is not a worse set for every purpose.
The standard metric in 2026 is engagement: followers, reach, listens, views, shares. Writing scores low on all of these relative to audio-visual. This looks like writing losing.
Writing is not losing the engagement competition. It is not entering it.
The rubric measures spread. Writing produces structural influence — changes to the model in the reader's head that persist and generate action. That influence is not measurable by engagement metrics and is not designed to be. One founder who finishes an essay and acts on what it clarified produces more structural change than ten thousand listeners who half-absorbed a related episode while traffic was bad.
The attention economy has produced a rubric. The rubric rewards spread. Writing cannot win on that rubric and does not try to. The people who continue to write and read long-form are operating on criteria the rubric cannot evaluate: thinking precision, model completeness, the compounding of a knowledge architecture, the selection of a reader who will act. That the rubric cannot see this is not writing's failure. It is writing's position.
The loss on engagement metrics is the selection mechanism working. The readers who were there for social reasons — to signal cultivation, to perform intellectual seriousness — have migrated to formats optimized for that performance. What remains is the fraction for whom depth is not a performance. That fraction is smaller. It is not, for the purposes of building something that compounds, less consequential.
The question is not whether to write in an environment that can't measure what writing produces. The question is whether the architecture you're building is the kind that benefits from what writing selects for. If it is, the undervaluation is not a problem to overcome. It is the condition that makes the selection work.