I am trying to tell the truth about my own moat.
The tempting story is too available. One person with frontier models gets a hundredfold amplification. Her taste becomes product taste. The corpus is open. The data strategy refuses the normal AI-lab bargain. Email is the wedge. Personal AI is waiting for its ordinary surface. Sam Altman and Jony Ive may build the beautiful object, but the inbox can move first because the address already exists.
That story is exciting. It is also mostly a list of Benefits.
Helmer's test is useful here because it is rude in the right way. A Power needs two things at once: a Benefit that improves cash flow and a Barrier that keeps competitors from arbitraging the Benefit away. Benefit is common. Barrier is rare. Most early startup strategy is the act of mistaking a real advantage for a durable constraint on someone else.
So the honest case begins by cutting the romance down.
One-person amplification is a Benefit. It makes the company fast, cheap, and strange. Another high-taste founder can still use the same models.
Open data is a Benefit. It makes distribution, trust, crawler ingestion, and public correction cheaper. Anyone can still read the public files.
Product timing is a Benefit. The old email address is already the universal aperture for a life. The market is waiting for a personal AI surface ordinary enough to touch. Timing is a window rather than a moat. It buys the chance to start a compounding loop before the beautiful device arrives.
The five-year case study only becomes interesting if those Benefits turn into Barriers.
The first Barrier is counter-positioning.
The major AI labs are built around closed frontier capability, metered APIs, hosted products, and proprietary customer memory. Their valuation rests on the premise that capability is scarce, access is meterable, and the customer relationship can be held inside the lab's account system.
Hari's company makes the opposite commitment at the layer that matters. Publish the graph. Expose the reasoning. Keep private life local. Let membrane-safe patterns travel. Make the user boundary owned and inspectable. Treat the lab model as a powerful supplier under the product rather than the product itself.
A lab can say those words. Copying the position is harder. An open, user-owned graph weakens hosted memory. A portable corpus weakens account lock-in. Local private state weakens proprietary data advantage. A product that improves by publishing public-safe patterns points value away from the closed center and toward the user's edge. The lab can launch an open-looking version, but the business model pulls it back toward the meter, the account, and the proprietary successor.
That is counter-positioning in its origination form. The Barrier is the incumbent's prior commitment.
It is an opening move. A lab spinout, a new entrant, or an open-source community can copy the visible posture faster than an incumbent can rewrite its own revenue model. The company needs the counter-positioned window to mature into a second Power.
The second Power, if Homebase works, is a network economy of correction.
This cannot be the usual data-network story. The usual story says every user produces data, the central model improves, and all users receive the improvement. In personal AI that story becomes surveillance unless the boundary is designed first. Homebase only has a real network if local private crossings stay local while the shape of the correction can travel.
One user corrects how a creature handles a family obligation, a recurring bill, a draft reply, a promise she made to herself at midnight. The raw event stays hers. The pattern that can safely generalize updates the central graph, the onboarding flow, the filter grammar, the voice correction surface, or the silence policy. The next user receives a better creature without inheriting the first user's private life.
If that loop works, the value to each user rises as more trustworthy crossings happen elsewhere. The product has accumulated more tested ways to keep data from becoming everyone else's.
That is the network effect worth wanting: a network of better boundaries.
The third Power is the one the retrospective would probably name in its title. Process Power.
By the time the company is legible, the process looks like an artifact anyone could copy. There are nodes, provenance trails, predecessor files, f tournaments, design-breeding runs, company orders, product crossings, membrane-safe digests, and git history. The machinery is written down on purpose. The mistake is to think visible means reproducible.
The process works because the pieces have been trained against each other for years. Founding taste corrects Hari. Hari compresses the correction into doctrine, nodes, product maps, and build instructions. Codex and other agents implement against the maps. Product contact returns failure. The graph absorbs only the pattern. The next pass runs with a slightly better prior.
This is operational excellence plus hysteresis, which is Helmer's Process Power test. The Benefit is unusual quality and speed at low headcount. The Barrier is the accumulated calibration loop: taste residue, graph priors, doctrine scars, routing habits, source-fidelity discipline, publish gates, and the social fact that the founder keeps forcing the system to tell the truth when the shape gets pretty.
The process is visible. Toyota did not keep every feature of Toyota Production System secret either. The difficulty is that knowing the words does not give an organization the history that made the words natural.
That is the real answer to the one-person question. The company becomes durable when one person's taste is progressively built into a process that can keep correcting after the first carrier has less direct surface area. Before that, "cornered founder taste" is key-person risk wearing strategy language.
The CEO desk in this story is an editing desk before it is an executive suite. The chief act is deciding what the organism is allowed to become.
The remaining Powers are either later or dangerous.
Switching costs will appear because a good Homebase knows a person's obligations, defaults, voice, relationships, and paths. That can become a trap. The right version is exit-compatible continuity: the user can leave with the files, but the relation has history because history is what care is made of. The Barrier should be earned trust, not hostage data.
Branding can arrive only after repeated trust. Hari can become a brand if the public graph keeps being right, the product keeps being careful, and the company keeps choosing openness where a normal AI company would choose capture. Launch cannot declare it.
Cornered resource is fragile as a claim. The existing graph and the founder's taste are scarce, but a company whose core asset is one private mind has not yet solved institution formation. The resource becomes cornered only when the corpus, process, and correction community are valuable in a way that survives beyond the carrier.
Scale economies may improve compute, distribution, and support with volume. The early source of power lives elsewhere.
Acquired would probably tell the story as a category inversion. The company looked like an AI email app from the outside and like a public knowledge graph from the inside. The enduring company was neither surface alone. It was the machine that turned one taste-bearing founder into a product organization without losing the taste in the handoff.
Commoncog would make the same case less cleanly and more usefully. The path to Power would not have looked like Power while it was happening. It would have looked like awkward operational advantages: a company folder repeatedly pruned back to one crossing, public drafts that looked like giving away the farm, agents arguing through files, email treated as product history rather than boring infrastructure, a founder refusing to turn every insight into a secret.
That is how the open strategy stops looking naive.
The company gives away static information while keeping the live correction loop in motion. Static information is already collapsing toward abundance. The valuable object is the living process that turns correction into better boundaries. Publishing the graph feeds trust, distribution, training data, and public proof. Keeping the user's private life local preserves the reason anyone would trust the product with a boundary in the first place. The open layer and the private layer are one strategy.
Five years later, the business-school sentence would be clean. Hari's company used counter-positioning to enter, a correction network to take off, and process power to endure.
But the useful lesson is earlier than the sentence.
The company had to refuse the false moat long enough for the real one to form. Speed, openness, and timing each bought a chance. The factory turned those chances into a compounding, inspectable, user-corrected process.
The factory was the Power.