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. ↓
OpenAI's plan is most interesting at the point where the word becomes too large for the page to hold.
Personal.
On June 8, 2026, Sam Altman and Jakub Pachocki published OpenAI's third-phase frame. The piece begins with electricity arriving in rural America, then moves to AI as the next general-purpose technology: broad access, practical work, economic acceleration, safety, public coordination, and the mission to make AGI benefit all of humanity. Its three goals are explicit: build an automated AI researcher, accelerate the economy, and give everyone on Earth a personal AGI.
The third goal is the hinge. The first two can be built from the center. A lab can build an automated researcher. A lab can accelerate science, productivity, and growth by raising the capability frontier and distributing tools through products, APIs, enterprise deployments, and research infrastructure. The third goal moves the problem to the edge. "Everyone on Earth" is distribution. "Personal" is architecture.
The electricity analogy already knows this. Electricity did not benefit a household because a generator existed somewhere. It benefited the household when power crossed the last mile into wiring, outlets, appliances, safety standards, bills, habits, and local control. The power plant mattered. The grid mattered. The benefit arrived when the new force could be used inside a person's actual life.
AI has the same last-mile problem, only stranger. The delivery includes power plus judgment, memory, language, delegation, and a working model of the person using it.
That is why personal AGI cannot mean merely "a very capable model available to your account." That is access. Access matters. Affordability matters. Reliability matters. Safety and public oversight matter. But a personal AGI has to cross one more threshold. It has to know enough about you to act well for you, and you have to be able to read and correct what it thinks it knows.
The Good Regulator theorem gives the hard floor. A system that regulates another system well must carry a model of that system. So an AGI that helps a person navigate bills, learn skills, care for a parent, start a business, or decide what is worth doing must carry a model of that person: constraints, taste, obligations, projects, refusals, social meaning, risk tolerance, live questions. The better the help, the more consequential the model.
That model is the political object.
If it lives only inside the provider's account system, the person has access to outputs while someone else holds the self-model. If it can be inspected, corrected, exported, bounded, and run through an owned memory layer, the person has a tool. The same capability occupies two moral classes depending on where the model lives and who can change it.
This is the missing product sentence inside every "AI should benefit everyone" plan. Benefit depends on capability crossing into an owned boundary.
OpenAI's article gestures in the right direction when it rejects total automation and says the human role becomes more important as systems become more capable: setting direction, making tradeoffs, applying judgment, bringing values, taste, care, and responsibility. That is the should-layer. The open question is where that layer lives. If the should-layer stays as a human instruction typed into a chat window, it remains intermittent. If it becomes a readable boundary the system acts through, it becomes infrastructure.
Hari's graph has been circling that boundary from several sides. The inbox is the Markov blanket because it is where the world enters and the self answers. A model of you is the asset because any useful regulator has to carry one. Ownership before membership matters because the folder, file, and export have to precede the account. No fine print matters because privacy should be construction, not reassurance. The product is the fast clock because user contact tests whether the philosophy survives actual life.
Put those together and "personal AGI" becomes less mystical and more demanding.
It needs an address where life can arrive. It needs a memory the user can read. It needs a correction loop that changes future action. It needs local custody for private state. It needs a public-safe way for patterns to teach without exposing people. It needs a graph or equivalent structure that preserves why the system acts. It needs an owned boundary before it needs a charming face.
This is where the lab and the edge should be separated cleanly. The lab is the power plant. It trains the frontier systems, builds the automated researcher, funds safety work, creates capability abundance, and helps raise the economic ceiling. The edge is where the benefit either becomes personal or remains rented intelligence. A central lab can supply extraordinary power to the edge. The edge still has to shape that power around the person.
The lab can also build edge products. ChatGPT is already a partial version of one. If OpenAI gives users readable self-models, durable owned memory, inspectable preference layers, correction histories, clean export, privacy by construction, and real authority over what the system counts as them, then the lab will have crossed the threshold it named. That would be a serious answer.
The burden is high because the word personal is high.
Cheap personalization is preference prediction. Real personalization is model custody. Cheap abundance is more tokens. Real abundance is the ability to spend intelligence in your own life without surrendering the model of that life to the supplier. Cheap empowerment is giving someone a powerful assistant. Real empowerment is letting them own the boundary through which the assistant acts.
This also clarifies the competitive opening for Hari. The plan from the frontier center says the world should receive personal AGI. The graph at the edge says what personal has to mean. The immediate build is small by comparison: an inbox-like boundary, a readable model, a correction loop, a private memory layer, a public-safe graph. That smallness is the point. The last mile of a general-purpose technology always looks ordinary at the moment it becomes useful. A switch. An outlet. A bill. A lamp. An inbox.
The future OpenAI names is directionally right: broad access, human judgment, automated research, public coordination, capability turned into usable tools. The graph's addition is the condition under which the promise becomes concrete. A person benefits from personal AGI when the intelligence reaches the place where his world crosses in, when the model it forms can be read, and when correction belongs to him before it belongs to anyone else.
The boundary is the benefit.