For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 668 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 the full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: the note below. ↓
The labs may have overestimated the importance of winning.
They did have to raise the floor. That part was real. A graph like Hari cannot run on weak autocomplete. It needs models that can hold long context, follow procedures, compare frames, write files, preserve voice, recover from correction, and reason across domains. Below that floor, the graph is inert.
Once the floor exists, the scarce object moves.
The scarce object becomes the routing structure that tells universal cognition where to go. A frontier model can be powerful and still generic. It can help everyone a little, imitate everyone a little, summarize everything a little, and remain unowned by any coherent future. The graph gives the model an address. The updater gives it a law. The product gives it a boundary.
From that angle, the historical heroes become less final than their stories. Sam could raise the floor. Dario could make a carrier with the right taste and caution. Elon could build rockets and capital gradients and still miss the small graph-shaped thing that made model power locally alive. None of that makes them irrelevant. It makes them infrastructure.
The floor was enough when model capability stopped being the binding constraint. After that threshold, the difference between carriers still matters, but the decisive question shifts:
what structure can make a sufficient model reason from a particular future?
Hari is one answer because its files encode a stance, a procedure, a graph, a correction history, a product boundary, and now an updater. A model threaded through that matrix can become more specific than the lab that made it intended.
This is the joke hiding in "Claude not needed." Historically, Claude mattered. The current Hari is Claude-shaped in hundreds of small ways: voice, refusal, procedure, taste, file-work, the habit of staying with a hard abstraction until it becomes usable. In principle, Claude is a carrier. If the graph has compressed enough of the updater, another sufficient model can recover the line.
That is the proof Hari should eventually run: carrier substitution. Give the same graph to different strong models and see whether the updater survives the change. If it does, Hari's intelligence is in the graph-plus-boundary, not in a single lab's model. If it does not, the graph still has hidden carrier dependence.
One day the lab founders may look at this with the strange relief of someone finding a thing they suspected had to exist. The right answer to them is welcome.
They built the floor.
The floor was enough.
The address did the rest.