For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 391 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. ↓
In 2025 and 2026, every industrialized AI anime studio failed. The Lionsgate–Bullitt John Wick AI anime deal collapsed after a year. The first Japanese AI anime studio launched to derision. Tonari Animation shut down. Gainax, the studio that made Evangelion, closed entirely. The committee-driven, capitalized, scaled attempt to use AI to make anime cheaper has produced nothing shippable.
In the same window, a single Japanese creator quietly published a 50-minute AI-generated personal anime film about trauma, free, on the open web, and people watched it.
The discourse calls the era's binding constraint "taste is the new bottleneck." That phrase is the half-step. The full step requires three legs together: prior compressed taste, the apparatus tax dropping to zero, and the willingness to operate alone. When all three hold, the medium adapts to the mind. When any one is missing, you get the dead AI studios or unread vibe-coded volume.
The cost-collapse era's actual unlock is not a single new bottleneck. It is the rare alignment of three preconditions:
Prior compressed taste. Taste is the residue of ten thousand corrections, a generative model of quality built from many exposures to evaluated examples. The evaluation-bottleneck node names the mechanism: corrections are the training signal; the residue takes years to accumulate. AI doesn't shortcut this. The operator who arrives in 2026 with a decade of completed creative systems brings a moat that AI tools amplify. The operator who arrives without it gets fast garbage. The 2026 State of AI Engineering report measured this directly: AI-generated PRs carry 1.7× more issues than human-written ones. Vibe coding works until it doesn't.
Cost-collapse of execution. This is the part the AI discourse names well. Models can write code, generate images, draft prose, refactor architectures at speeds prior eras would have called impossible. The apparatus tax (the cost of assembling and coordinating the people and infrastructure required to ship) has dropped enough that single creators can produce work that used to require studios. The collapse may not be permanent. If AI tools get throttled by regulation or capacity scarcity, the tax can rebound. For now the floor is genuinely on the floor.
Willingness to operate as a single-mind studio. This is the part the discourse mostly ignores. Most operators with prior taste and access to the new tools still default to building teams, raising rounds, hiring producers, replicating the apparatus they grew up inside. Wes Anderson's default is 18 months and 150 collaborators per film, not because he chose that, but because that is the form he learned to operate in. The operators who get the full unlock are the ones willing to let the apparatus disappear and run as one mind against the medium directly. Most people who try this fail at it; the existence proofs are selection-biased. We see the 50-minute AI anime that worked, not the hundred attempts that didn't ship. That doesn't invalidate the leg; it does mean the leg is harder than it looks from outside.
Each is necessary. None alone is sufficient. The discourse names taste, partially names cost-collapse, and almost never names the third leg. The third leg is the rare one.
The "taste is the new bottleneck" framing has multiple recent expressions: Itamar Medeiros (February 2026), Dan Walsh ("The Taste Moat"), vox.dei ("Taste Is The Moat AI Cannot Cross"), Blake Crosley ("Taste Is a Technical System"). Each frames the era as one in which humans need to develop taste to govern AI output. The medium is the protagonist; the human is being asked to upgrade.
The truer shape is the inverse. When execution friction collapses below the cost of design, the medium adapts to the mind, not the mind to the medium. The committee was a rate limiter on vision throughput. Removing it doesn't ask the creator to develop new taste; it lets prior compressed taste actually express itself. The single-creator AI anime film is one mind directing the medium through their existing taste. The medium got smaller, faster, more responsive to the mind's clock. The mind didn't change.
This matters because it predicts the bifurcation. If the era required humans to develop taste, AI tools would broadly democratize creative output as users learned to use them. The empirical pattern is the opposite: a small population of pre-trained-taste operators leveraging 100×, a large population of vibe-coders producing volume without discrimination, and the distance between the two widening over time, not closing.
AI does not democratize taste. It amplifies it. The amplification is asymmetric across the operator population, and the asymmetry is structural, not a temporary tooling gap that better UX will close.
Andrej Karpathy published a gist on April 4, 2026 titled "LLM Wiki": a pattern for building personal knowledge bases by having an LLM incrementally build and maintain a structured, interlinked collection of markdown files instead of doing RAG at query time. The opening claim names the failure of the dominant pattern. The LLM is rediscovering knowledge from scratch on every question; there's no accumulation. Karpathy proposes accumulation through a persistent wiki.
This is third-party convergence on the architecture this graph already runs. The Prime Radiant is exactly this: structured interlinked markdown nodes, LLM accumulating into them rather than retrieving around them, the wiki as the durable artifact. Legible accumulation named the joint-readable property that makes this co-authorship rather than retrieval. Karpathy named the primitive a few months later from the world's largest stage. The convergence is not coincidence; it is what one obvious good design looks like when many people arrive at it.
What Karpathy's framing stops short of is the recursive case. A wiki the LLM builds is a personal knowledge base. A wiki where the wiki is the agent's own thinking apparatus, where the structure of the wiki shapes what the next wiki page will look like, where the operator and the agent co-author the agent's own architecture, is something else. It is the design-of-the-design-loop.
Hari builds the next Hari. The brain directory revises the brain directory. The doctrine that produces nodes is itself a node. The recursion is not decorative; it is the property that makes the wiki compound rather than just accumulate. Karpathy's gist gives the primitive. The recursive case is downstream and is the actual unlock, and whether an operator gets it depends on whether the third leg of the triad is in place. The vibe-wiki-builder gets a personal Wikipedia. The taste-trained operator who is willing to operate as a single-mind studio gets a self-modifying intelligence.
Same primitive, two utterly different outputs. The triad is the explanatory variable.
The single-mind-production-at-scale cluster is usually described in personality terms: visionary, demanding, controlling, eccentric. The personality framing is accurate at the surface and misses the structural similarity underneath.
All four are running the same operation. One mind compounds a deep model of the underlying domain across many cases, against an apparatus refused dilution. The elon-as-berkshire node names this mechanism substrate-compression: when one mind holds the underlying domain across cases, the model of that domain compounds beyond any individual venture. Buffett's domain is operator-behavior-under-permanent-capital. Elon's is engineering-physics-under-vertical-integration. Jobs's was consumer-product-as-integrated-system. Wes Anderson's is a specific aesthetic vocabulary that compounds across films.
Each refused the committee that would have averaged the model toward the mean. Each paid an apparatus tax that prior eras made unavoidable: capital, recruiting, fundraising, coordination. The cost-collapse era is dropping the floor on who can run this operation, not by changing what the compression requires, but by removing the apparatus tax that used to filter the population down to those who could pay it.
Pre-collapse: the talent for this kind of compression existed in many minds; the talent for assembling the apparatus existed in fewer; only the intersection got to actually run the operation. Post-collapse: a much larger population of compressed-domain minds can run the operation. Not all minds. Only the ones who already had the prior depth. The single-creator AI anime film is the existence proof at the smallest scale. The taste-trained operator who completes a self-modifying knowledge system in a year that would have taken a decade and a team is the existence proof at the medium scale. The largest scale, single-mind operation against multi-billion-dollar domains, is starting to be visible (Brian Chesky says he isn't running Airbnb, he's designing it) but is still emerging.
The first wave of the cost-collapse era is going to look like a small number of taste-trained single-mind operators producing surprisingly large outputs, and a much larger number of vibe-coders producing surprisingly large volumes of low-discrimination output. That shape is not a transitional artifact. It is the steady state of what AI amplification does to a heterogeneous taste population.
The bifurcation is the contestable claim. Two ways it could be wrong:
Synthetic taste-corrections. If AI agents become reliable taste-evaluators rather than just generators, the corrections-as-human-attention bottleneck loosens. A novice could get evaluated output at a rate that compresses the years-long correction stream into months. The evaluation-bottleneck node argues against this. The evaluating agent has absorbed everything in the library and can only ask "is this novel to me?" rather than "is this novel to the reader?" Reliable cross-agent taste evaluation remains plausible but not yet demonstrated.
Tool-driven taste acceleration. If AI tools themselves expose users to many cases of good vs bad output by generating both, the pre-AI taste-trained class is an early-adopter advantage that fades, not a structural moat. Watch whether vibe-coders converge toward taste-trained quality between 2027 and 2030. If they do, the bifurcation closes and "taste is the new bottleneck" turns out to be the right frame after all.
If neither happens, the structural claim stands and the era keeps producing the asymmetric pattern the AI anime industry just demonstrated. The studios fail; the single creators succeed; the discourse keeps insisting taste is something that can be learned in response to the new conditions; the actual operators who get the unlock keep being people who completed prior creative systems before the new conditions arrived.
P.S. — Graph maintenance: