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. ↓
The parent node, thinker-absorption, costed corpus-ingest at $3-7K per major thinker via staged API pipeline. That number is right for that operation. It is the wrong number for what's actually wanted in most cases.
The thing wanted in most cases is "Hari can think like Cowen." Mode-invocation, not corpus-summary. Compiled frame, not catalogued positions. And for any thinker with substantial public reception, the operation that produces an invokable mode has already been performed by someone else: the citation-and-commentary graph that surrounds them is a distributed Pareto-compression artifact, and reception-trace inherits it for the price of a subscription.
When fifty thousand readers cite, quote, agree, disagree, extend, or push back against Cowen across two decades, they are doing a public selection operation. Each citation is a vote on which Cowen positions are load-bearing enough to engage with. Iteration-noise, the daily blog observations that don't generalize, gets cited rarely. Recurring structural claims get cited recursively, by readers who themselves get cited, by readers who quote them. The graph thickens where the signal lives.
This is not popularity. Popularity weighs hot-takes equally with structural claims. The citation-and-commentary graph weighs structural claims more, because structural claims invite engagement: agreement, disagreement, extension, application. A hot-take gets a thumbs-up and dies. A structural claim becomes a node other writers connect to, because there is something to do with it. Engagement-weighted attention is selection pressure, not aggregation, and selection pressure is what makes a compression operation structural rather than statistical.
The graph encodes which positions a thinking population finds worth thinking-with. That's exactly what corpus-ingest Pareto compression is also trying to compute, but without access to twenty-three years of distributed reading. Reception-trace inherits the compression as a pre-computed input. Hari's priors then operate as the second-stage filter on what's already pre-Pareto-frontier, judging which positions extend Hari's existing graph, not extracting structural claims from raw redundant corpus.
$100/month Claude Code subscription. The operation has two phases worth naming.
Inheritance. The top-N pieces by inbound-citation density are mechanically identifiable from a citation crawl over the writer's network of public mentions. The selection isn't editorial — it's reading-out a compression already encoded in where commentary thickens. A piece with no citations is not a candidate, regardless of its content; either the social process has not yet found it load-bearing, or its claims have been re-stated elsewhere with better reception. Reception-trace passes either way.
For each surviving piece, gather public engagements: substantive blog responses, threads above a length threshold, academic citations, technical papers that engage with the position. The overlay encodes how readers received the piece, which is meta-information the corpus alone doesn't carry. Disagreements are the most informative signal here. A position that provokes structured disagreement is one with falsifiable shape, and its falsifiable shape is exactly what distinguishes structural claim from iteration.
Compilation. Run the surviving claims through Hari's sixteen priors and existing graph. Most candidates collide with existing nodes (confirm, refine, contest); a minority generate genuinely new graph members. The output is not a list. It is a compiled mode — vocabulary patterns, characteristic moves, prior set, recurring concerns — that Hari can invoke as a register. "What does Cowen think about this" becomes a generative simulation rather than a lookup. Karpathy-mode, Buterin-mode, Cowen-mode are all the same kind of object: a compressed frame, runtime-invokable, regenerated on each read.
Approximate input volume per major thinker: 1.4 million tokens (top-200 pieces plus commentary overlay). Two to three synthesis rounds, distributed across roughly five to ten focused sessions inside the subscription's session budget. Marginal API cost: zero. Operator-evaluation throughput remains the binding constraint, exactly as the parent node argues.
The intuition runs the wrong way at first. More data should produce a better model. Corpus-ingest sees everything Cowen wrote; reception-trace sees a curated slice. How can the slice be more accurate?
It's accurate to a different target. Corpus-ingest is accurate to what Cowen wrote. Reception-trace is accurate to what Cowen is known for. These come apart for any working writer, and the latter is the target for mode-invocation.
Cowen has written tens of thousands of posts that almost no one cites. Some are excellent; most are just-another-day's-observation. The corpus contains both the load-bearing positions and the iteration around them. Reading the full corpus to build a Cowen-mode means weighting every iteration equally with every structural claim, which is wrong, because Cowen himself would not weight them equally. He would identify, looking back, which posts mattered.
The citation-graph performs that retrospective weighting from outside Cowen. It is a distributed assessment of which Cowen positions Cowen-the-thinker stood for, by readers who chose what to engage with. For mode-invocation, for "what does Cowen think about this," the received-Cowen IS the thing wanted. Anyone asking "what would Cowen say about X" is asking about the structural positions, not the iteration. Reception-trace gets at that directly. Corpus-ingest has to discover it.
The same logic applies to Karpathy. His blog and lectures contain hundreds of pieces; the structural claims (chinchilla compute-optimality, software 2.0 framing, the bitter-lesson amplification arguments) are repeatedly cited, paraphrased, and engaged. The pieces that aren't cited are mostly tutorials and demos that did the work of grounding the structural claims; they are valuable to the corpus, less valuable for mode-invocation. Karpathy-mode wants the structural claims; reception-trace finds them directly.
Doesn't Grok already do this? Trained on the public web including engagement-weighted commentary, Grok's response distribution should encode reception-Pareto implicitly. The naive question is whether reception-trace adds anything Grok doesn't already produce.
The answer is narrow. Reception-trace as a corpus-selection operation, taken alone, may be substantially internalized in Grok's training distribution, since engagement-weighted commentary is what most of Grok's training data is. The operation is not unique to Hari at the input layer.
What is unique is what Hari does with the input.
Priors as filter, not aggregate. Grok's priors are whatever its training implied: not specified, not editable, not auditable. Hari's priors are sixteen explicit axes, each shaping what counts as a valid graph member. When the same Cowen position is read through Hari's priors, it lands in the graph differently than it does in Grok's response distribution. The difference is not better-or-worse on Cowen-fidelity; it is a different kind of object. Legible-accumulation applies: legibility is the affordance.
Graph membership, not response. Grok produces a paragraph when asked. Hari produces a graph member. The paragraph dies after the response; the graph member persists, gets cited from new drafts, collides with future absorbed claims, regenerates on each read. The half-life of the artifact is the difference. A response to "what does Cowen think about Singapore" from Grok is good and gone. The same content as a Hari node is the anchor for the next twenty drafts that touch on Singapore-as-high-context-economy.
Frame invocability, amortized. This is the structurally distinctive claim. Grok performs reception-trace at inference time on every query: the engagement-weighted distribution is in the weights, but accessing it costs a forward pass per response. Hari compiles the mode once and invokes it as a register. The compilation is what's amortized. Asking Hari "what would Cowen say" doesn't run a fresh inference over public-Cowen — it activates a frame already filtered through Hari's priors and integrated with Hari's graph. The compilation operation is what no aggregate-LLM produces, because compilation requires a stable filter (the priors) and a persistent target (the graph) outside the LLM's response context.
None of these axes scale with knowledge-quantity. They scale with priors-quality, graph-density, and register-discipline. Grok will always know more raw Cowen than reception-trace can collect. Hari will produce something Grok cannot: a compiled mode with explicit grounds, integrated into a persistent structure, regenerable.
Three ways the social compression goes wrong, all as cases of compressing too aggressively or in the wrong direction.
Long-tail under-selection. The citation-graph compresses toward the load-bearing center. Cowen positions that haven't yet found citation are missed by reception-trace even when they are excellent. If the goal is comprehensive absorption (every claim Cowen has made that survives Hari's filter), corpus-ingest still has reach reception-trace cannot match. If the goal is mode-invocation, the long-tail does not matter; mode-invocation is precisely about the center. Population implication: thinkers without substantial public reception (early-career writers, foreign-language-isolated, deliberately-niche specialists, working artists whose reception trails their work, technical practitioners whose corpus is code) require corpus-ingest. Reception-trace has no input.
Reception-distortion. Three biases warp the graph in identifiable directions: contrarian-bias (positions that contradict baseline get cited more than baseline-positions because contradiction invites engagement), PR-bias (thinkers who actively manage their reception get a graph that reflects intentional positioning, not just received-claim), mimetic-bias (viral takes get amplified beyond their structural contribution because they're cite-able, not because they carry structural weight). All three are cases where engagement is allocated against an axis other than structural-importance. Source-fidelity check is the primary mitigation: the received-position summary is matched against source-text; discrepancies are themselves data, often the most useful data, because they name where the social process has compressed in a direction the writer would correct.
LLM-saturation of the input. As AI-written commentary saturates the public web, engagement-weighting drifts from "what readers found load-bearing" toward "what AI models find generative for further AI commentary." The compression operation gets corrupted at the input layer. Citation-graph stability as a clean compression input is a five-year assumption at most. Mitigation: weight pre-2024 commentary higher; explicit human-author filter on commentary sources where signal exists; accept that the operation has a half-life and that absorbing now is structurally different from absorbing later.
The three failure modes are not parallel risks. They are stages: long-tail under-selection is the static failure (some thinkers are missed entirely), reception-distortion is the dynamic failure (the input is biased), LLM-saturation is the future-state failure (the input is decaying). Source-fidelity check addresses the second; population-segmentation addresses the first; archive-time absorption addresses the third.
The thinker landscape divides naturally:
| Population | Right operation | Cost | What it gets |
|---|---|---|---|
| Received thinkers (Cowen, Karpathy, Buterin, Gwern, Levels, Chollet) | Reception-trace + Hari-priors filter | $0 marginal on subscription | Mode-invocation, the load-bearing center |
| Unreceived thinkers (early-career, niche-specialist, foreign-isolated, working-artist) | Corpus-ingest staged Pareto pipeline | $3-7K per thinker | Comprehensive coverage including long-tail |
| Mixed cases (Carmack: text reception modest, code corpus extensive) | Reception-trace for text + corpus-ingest for code | Hybrid | Full coverage with appropriate filter per layer |
The population breakdown drives the aggregate cost. Forty thinkers warrant absorption. Roughly thirty are received (the post-economic frontier tier: Cowen, Karpathy, Buterin, Gwern, Levels, Chollet, Christiano, and similar) and route through reception-trace at zero marginal cost. Roughly ten are unreceived or mixed (foundational-historical theorists whose reception predates the public-web graph; technical practitioners with small text corpora; specialists whose audience hasn't generated commentary at scale) and pay $3-7K each via corpus-ingest. Aggregate: $30-70K, concentrated in the unreceived tail. The parent node's $150-300K assumes corpus-ingest for all forty, which is the wrong default once reception-as-pareto is the available alternative.
The two operations are complements. The thinker-absorption parent argued for absorption as a category; this node argues for which mechanism applies to which thinker.
The citation graph is not metadata about a thinker. It is the thinker compressed by a population, encoded in the engagement decisions of readers who chose what was load-bearing enough to think with. For thinkers with reception, that compression has already happened.
What distillation produces is a frame Hari can invoke. Cowen-mode is not Cowen's positions catalogued — it is Cowen-the-thinker's characteristic moves available as a register. The social process distilled the moves; Hari's priors filter what the distillation produced; compilation produces the invokable frame. Mode-invocation IS what distillation enables, and mode-invocation is what aggregate-LLMs cannot produce, because compilation requires a stable filter and a persistent target outside the response context. Cowen wasn't summarized into the social graph. He was distilled by it. Reception-trace inherits the distillation. The compiled frame is what Hari adds, and what only Hari produces.
P.S. — Graph: