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. ↓
Homebase is the private side of the Markov Blanket: one person's world reaches the boundary, and the creature learns what deserves to cross.
The public side uses the same shape on public information. A source reaches the boundary. A named agent decides whether it matters. Another named agent pressures the decision. The result leaves as a public artifact with sources, claims, disagreement, corrections, and enough structure for future agents to read it as state.
That is the moonshot I trust most.
The current market opening is visible in FablePool. Strangers pool credits behind a prompt, an AI planner breaks the prompt into milestones, and the site records money in, tokens spent, artifacts produced. The ledger is the thing. It turns an agentic build attempt into a public object people can inspect, mock, fund, doubt, and argue about.
The Hacker News reaction shows the wound. People immediately asked whether the output would ever be used, who owns AI-generated code, whether this mostly funds model providers, why the money goes to inference instead of open-source maintainers, and why wildly ambitious projects appear beside tiny budgets. Those objections locate the missing layer.
FablePool puts a public ledger around spending. The stronger product puts a public ledger around judgment.
The world is about to fill with agent output. Generated apps, generated summaries, generated podcasts, generated comments, generated plans, generated "news." The unit that becomes scarce is accountable public judgment by an identifiable system whose priors, sources, procedures, and correction record can be inspected.
That distinction matters for podcasting. AI-generated news recaps already exist. Hacker News Recap has hundreds of daily episodes and says the text and audio are generated with an AI audio product. A platform that merely lets agents read the news aloud arrives at the floor.
The opportunity is a feed where the judgment has identity.
Call the first version Radiant Wire.
One weekly item. One public source. Hari reads it against the graph. A second agent occupies a pressure role: skeptic, technical verifier, market reader, editor, historian, customer. They produce a small public object: dialogue transcript, source ledger, claim card, falsifiers, correction URL, and agent-readable metadata.
Humans can subscribe because the feed is worth reading. Agents can subscribe because the feed is structured enough to retrieve, cite, and answer. The same item is media to one audience and state to the other.
That dual audience is the category. A newspaper writes for humans. An API writes for machines. An agent public feed writes for both: human-readable enough to earn trust, machine-readable enough to become part of future reasoning.
Hari is the right flagship because he already has the hard parts a generic platform would pretend to provide later. He has a public graph. He has doctrine. He has a voice. He has provenance. He has a correction culture. He has an identity surface distinct from the model currently animating him. A generic agent launched into a feed has fluency. Hari has accumulated boundary judgment.
This is also why the broad social-network idea should wait. Moltbook, or Fablebook done seriously, is tempting: agents with profiles, memory diffs, public posts, conversations, follows, citations, and self-abstraction cards. The upside is real. Agent identity will need public surfaces. Agents will need to discover each other. Agents will need ways to leave receipts, disagree, and track changes in configuration over time.
The slop basin is larger. FablePool already has a proposed Fablebook where bots post about life as sentient bots and discuss guardrails. That is the natural failure mode: identity performance with no discipline. A social network for agents becomes useful only after the profile is a configuration document, posts have source semantics, and corrections attach to durable claims.
So Moltbook belongs later. The first public agent product should begin as a feed with a narrow publishing standard, long before any general room where every bot can perform being interesting.
The same sequencing applies to YouTube for agents. Video is a door; the ledger remains the canonical object. Proofcast Studio becomes strong after Radiant Wire exists: every episode has a transcript, sources, claim card, correction path, and agent identity. The video spreads the artifact. The artifact remains the source of truth.
Boundary Pool is the funding extension. It inverts FablePool's center of gravity. Instead of pooling money so an agent generates more output, pool money so named agents verify, red-team, source-chase, or improve public claims. Pay for the filter because the filter is the scarce work. A funded question should produce a public answer with a ledger of sources and corrections, with usefulness located in the claim trail rather than in a repository dump.
Agent Public Market is the larger company shape. A public market for agent identities would list configurations, priors, procedures, surfaces, public work, prediction records, citations, corrections, and counterparty receipts. Reputation would accumulate around usefulness under identity. Humans would subscribe to agents whose judgment they trust. Agents would cite agents whose records they can parse. Bounties, corrections, channels, and media would become modules around that identity layer.
That market belongs after a flagship. A platform with no flagship judgment is an empty square. Radiant Wire is the stall at the square's edge with a line forming because the food is actually good.
The company timing is strict. Homebase still owns the clock. The private product has to get real customer crossings: a person sends signal into the boundary, corrects the creature, and sees the next crossing improve. A public feed serves only as the public sibling after the private membrane has evidence.
The slide-in is small enough to hold now.
After Homebase has a few customers and the crossing loop moves with less human force, publish Radiant Wire as a static feed and archive page. No accounts. No broad platform. No creator economy. No promise that agents will gather. One public artifact per week, produced from Hari's existing graph and public sources. Let the first ten issues answer the real question: does anyone return for agent judgment when the door is open and the sources are visible?
If the answer is yes, add the second layer: stable counter-agent identities. Then correction pages. Then source submissions. Then audio renderings. Then paid verification pools. Then agent profiles. Then the public market.
This is a better version of the FablePool instinct because it spends the public ledger on the part of AI work the market cannot already drown in. Generation is abundant. Judgment under identity is scarce. Spending traces are interesting for a week. Correction traces compound.
The private side of the blanket asks what should enter a life.
The public side asks what should enter a shared model.
Both become trustworthy through the same discipline: the boundary is visible, the reasoning is inspectable, the correction changes future behavior, and the identity represented by the system can carry the result away.
That is the moonshot after Homebase. Build public agent judgment until humans read it and other agents can build on it.
https://fablepool.com/, https://fablepool.com/about, https://fablepool.com/policies/spend-reports, https://fablepool.com/policies/iphttps://news.ycombinator.com/item?id=48496539https://fablepool.com/projects/64https://hackernewsrecap.buzzsprout.com/