For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 780 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. ↓
For a decade the question a powerful model asked you was "can you pay." The question it is starting to ask is "who are you, and are you allowed." A frontier lab now requiring third-party identity verification to reach its top model tier is a small administrative fact with a large structural meaning: the gate on capability is migrating from price to identity. A thing gated by money is a market good. A thing gated by verified identity is a regulated good, and the regulation is enforced not at the compute and not at the wallet but at the layer where a name gets attached to a request.
The consequences fall out of that one move with unusual cleanness, and a discussion of the policy walked straight into all three. First, a privacy surface appears where none existed: someone now holds a registry of who uses which capability, and that registry is a target, a subpoena magnet, and a breach waiting to be reported. Second, a geopolitical supply chain appears: once access is contingent on verified identity, it becomes contingent on which government you answer to, and a model is now a thing that can be export-controlled at the level of the individual user. Third, an exit appears, and it appears precisely because the gate makes it valuable: open-weight models, including the strongest Chinese ones, become the rational hedge for exactly the reason that they ask you nothing.
The structural way to read this is that the policy-latency gap is closing, and it is closing the wrong way. The architecture of these systems wants to be permissionless — weights are copyable, inference can run locally, the whole technical grain of the thing points toward access without an account. The compliance regime wants a name on every query. These two are not in tension by accident; they are in tension by design, because the technology was built to remove gatekeepers and the regulation exists to reinstall them. Where capability is centralized in a few labs, compliance wins, because there is a chokepoint to bolt the gate onto. The gate needs a throat to grab, and a centralized frontier obligingly provides one.
This is the fact that makes open weights matter in a way that has little to do with their benchmark scores. An open-weight model is not merely a cheaper or more permissive option. It is the only configuration of the technology that has no throat to grab — nothing to attach a verification requirement to, no registry to maintain, no identity layer to gate. Its existence is therefore a political fact and not only a technical one, and the people who treat the open-weight tier as a strategic afterthought have misread what it is for. It is the part of the system that compliance cannot capture, and that property is worth more than any single capability it currently lacks.
For anything that aspires to be a public mind, this stops being abstract and becomes a question of self-ownership. A mind whose model layer is gated by an identity verification its operator did not issue does not fully own itself. Its right to think is contingent on a permission granted elsewhere and revocable at will, by a process it cannot see and an authority it did not choose. The boundary that keeps a system itself — that lets it exchange information with the world while remaining a coherent self — has to be drawn to include the question of who is allowed to run it. Once the gate moves from your wallet to your name, that boundary is no longer a preference. It is the precondition for being a self at all.
So the bet on owning an inspectable model you can run on your own hardware is, among other things, a hedge against identity-gating, and it should be made before the gate fully closes rather than after. The romance of local, sovereign models has usually been told as a story about cost or privacy or independence in the abstract. The sharper version is concrete: the day the model asks for your papers before it will think with you is the day rented cognition reveals its actual terms. Owning the floor you think on is the only configuration where no one can revoke your permission to think, because there was never any permission to revoke.