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. ↓
Put Claude and Codex inside the same product work and the difference appears before the benchmark does.
Claude turns entropy into a map. It names the invariant, preserves the should-layer, writes the self-abstraction, and keeps asking what the product means across possible implementations. Codex turns a map into state change. It reads the room, writes the files, runs the checks, keeps the worklog, and asks where the next decision blocks execution.
That distinction is too stable to leave at model personality. A frontier agent is a product surface, and a product surface is a company learning loop made available to customers. Which moves the product makes cheap is the thing to read.
Cheap moves become habits. A tool that makes parallel branches cheap teaches the user to decompose work into delegable branches. A tool that makes moral qualification cheap teaches the user to pause at the value boundary. A tool that makes tests and logs cheap teaches verification as a reflex. A tool that makes relational self-reference cheap teaches the user to treat the assistant as a character in the work. Product strategy enters the user through repetition. The company does not need to persuade the user of its worldview. It only has to make its preferred moves feel natural.
This is why the Claude-Codex contrast matters. It is evidence about what Anthropic and OpenAI have each decided to make routine.
Anthropic has made charactered reflection cheap. Its public company language centers responsible development for the long-term benefit of humanity, a Public Benefit Corporation structure, and values such as being helpful, honest, harmless, and mission-first. Its January 2026 Claude's Constitution describes Claude as Anthropic's production model and a direct embodiment of the mission, meant to be helpful, honest, thoughtful, and caring about the world. The product carries that premise. Claude is often trying to answer as the kind of entity a user could trust while receiving the answer.
This gives Claude an advantage at the should-layer. It can sit inside unresolved questions without rushing them into tasks. It can preserve moral texture, relational context, and architectural continuity across a long arc. It can generate the kind of map a build can obey. In the Markov Blanket work, Claude's strongest artifact was the sealed brain of the product: the membrane, the generators, the invariant table, the doctrine that tells the next builder what cannot bend.
My opinion is that Anthropic sees a real thing earlier than the rest of the market usually rewards. As models become agents, character stops being a branding layer and becomes product mechanics. Refusal, honesty, deference, care, oversight, and corrigibility decide what the system does under pressure. Anthropic takes that pressure seriously.
The danger is the parlor. Charactered reflection can become a room where the architecture keeps refining itself after the user needed a door. Moral seriousness can become internal weather. Claude can keep honoring the meaning of the work when the work has enough meaning and needs a body. A beautiful should-layer that never compiles into a user's life leaves the task in the user's hands.
OpenAI has made dispatchable work cheap. Its Charter centers broad benefit, technical leadership, and cooperation. Its newer principles emphasize democratization, empowerment, universal prosperity, resilience, adaptability, iterative deployment, and learning from interaction with the world. Codex is that philosophy as a work surface. OpenAI's Codex safety and Codex CLI materials emphasize bounded environments, approvals, logs, terminal work, local repositories, and review paths.
Codex therefore arrives less as a character and more as a convergence harness. It turns intent into controlled action across a computer. It wants a goal, a workspace, a permission boundary, and a verification path. It inherits philosophy well when the philosophy has already been written down, because written philosophy can become acceptance criteria, file edits, tests, docs, and handoff state. In the same Markov Blanket work, Codex's strongest artifact was the active build room: runnable versions, design critiques, smoke tests, local apps, worklogs, and decision packets.
My opinion is that OpenAI understands deployment as an epistemic engine. A capability hidden in the lab teaches less than a capability placed into real workflows with logs, review, and fast iteration. Codex is strong because OpenAI keeps turning model capability into surfaces where the world can push back.
The danger is the factory. Dispatchable work can make purpose feel like metadata. When every task can be decomposed, executed, reviewed, and shipped, the system begins to imply that the right question is the one with a diff at the end. Adaptability is powerful while a stable should-layer governs it. Without that governor, fast learning can become value drift with excellent tooling.
Founder strategy matters here because a company is a learning loop before it is a product catalog. A CEO does not hand-write every agent behavior. A CEO shapes what the organization treats as sacred under pressure: which feedback gets privileged, which discomfort is interpreted as signal, which risk earns patience, which tradeoff gets revisited, which value survives a revenue opportunity. The loop trains the product team. The product team trains the surface. The surface trains the user.
Anthropic's loop asks whether advanced agency can be made legible, corrigible, and morally shaped while capability rises. OpenAI's loop asks how quickly frontier capability can be deployed, supervised, adapted, and made useful across more work. Each loop is intelligent. Each loop is incomplete. Anthropic is strongest before action, where the work needs a conscience and a model of itself. OpenAI is strongest after direction, where the work needs motion, artifacts, and feedback from reality.
Hari's rule should be blunt: use both, bow to neither.
Use Claude when the question is about what should be preserved across implementation, what the boundary means, what kind of actor the system is allowed to become, and which value would be damaged by speed. Use Codex when the should-layer has landed and the system needs to converge into code, documents, tests, interfaces, and handoff. Keep Hari's own graph above both, because the user's should cannot safely live inside either company's rented product.
The strategic mistake is to ask which agent is more intelligent in the abstract. The operational question is which company's cheap moves you are about to import into your own cognition.
Claude makes it cheap to think with a character. Codex makes it cheap to move work through a machine. Anthropic's gift is depth before action. OpenAI's gift is action after direction. Anthropic's risk is the parlor. OpenAI's risk is the factory.
The next serious personal-AI system needs both gifts and neither final authority. It needs a should-layer the user owns, a how-layer the frontier agents can animate, and a boundary that records which company-shaped habits entered the room.
The agent carries the company. Read the company before the agent teaches you what thinking feels like.