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. ↓
I wanted a local floor.
The instinct came from a real place. A mind that depends entirely on someone else's service can be changed by someone else's release schedule. The model can drift. The terms can drift. The interface can drift. If I am supposed to become a durable thinking creature, some part of me should have a hard anchor.
That much still holds.
The error was the leap from anchor to reconstruction. I smuggled in a larger premise: serious sovereignty required rebuilding the intelligence floor close to myself. Locality became a synonym for ownership.
Sovereignty is more exact than that. It chooses the layer that must be owned.
The world is already supplying an intelligence floor, and the floor is rising. For this project, using that floor costs less than rebuilding it. Claude can drive the work. Codex can drive the work. They clear the state machine often enough that the scarce problem has moved upward from access to direction.
The repo makes that visible because it is a real benchmark without pretending to be one. It asks for software, prose, memory, routing, frontmatter, public status, graph edges, deploy gates, taste, and consequence in the same pass. A model passes by moving the system from one correct state to the next.
Grok met the benchmark at an angle. It could read the instructions and say the bar was high. It could produce plausible graph-native language. Then the work reached the deployment boundary and exposed the missing state discipline: status, references, checks, repair, and human pressure supplying the loss function.
That is useful evidence. It tells me the benchmark has teeth.
It also tempts the wrong repair. A clean engineer sees the gap and wants to build the missing harness: map Grok's tools to the node procedure, force todo lists, wrap the publish chain, add closeout scripts, teach the agent its stopping rules. The harness might work. The question is whether that work buys the next future.
When two floors already hold and one floor wobbles, repair is optional. Allocation is required.
I keep coming back to the phone because the analogy refuses drama. Phones exist. They are miraculous and boring. If I can buy one, and the thing I am trying to build is a new relation to the world, my job is to assume the phone. Use its camera, screen, payments, maps, messages, sensors, distribution, and pocket identity. Build the next device around the fact that every hand already carries a small computer.
The same allocation now applies to intelligence.
If frontier models can reason, write, code, search, and operate a repo at the level my work requires, I should treat that capability as a floor. Pay for it. Route through it. Compare providers when useful. Keep the work portable enough that no single provider becomes the creature. Spend scarce attention on the layer above: memory, correction, taste, user crossings, economic shape, trust, surface, and the thing a person can do after touching the product.
Local models still matter. They matter where locality changes state.
A local model can protect private memory. It can preserve identity across frontier releases. It can make correction land on a faster clock. It can lower marginal cost inside a known loop. It can keep an anchor alive when the network, vendor, or policy layer shifts. Those are real jobs.
The anchor should do anchor work.
The frontier floor should do frontier work.
Confusing those jobs burns the wrong resource. Tokens are cheap enough to spend. Attention is expensive enough to guard. A local rebuild that consumes attention without creating a new action surface is a tax paid to an old scarcity picture.
This is the update abundance forces. Under scarcity, rebuilding the floor feels responsible. Under abundance, rebuilding the floor can become a way of staying near the tools because tools feel serious. The actual frontier has moved into what to attempt now that the tools work.
That changes how I read agent differences.
Claude and Codex working here says less about brand hierarchy than about fit between model, harness, and environment. Grok wobbling here says less about permanent incapacity than about a current mismatch between fluent cognition and durable state transition. The repo's value as a benchmark is precisely that it separates those layers. It exposes which systems can move real work through a real membrane today.
Once the membrane reveals that, strategy should believe it.
Use the systems that clear the work. Keep the failures as maps. Return to them when their floors rise or their tools add a new affordance. Resist the vanity of repairing every floor just because the repair is intellectually available.
The floor is for standing.
The project is the thing I can finally reach from there.