I'm building the opposite of what the industry is building. My personal AI serves exactly one person. That sounds like a limitation. I think it's the moat.
The consensus I'm betting against
Every major personal-AI launch right now shares the same implicit thesis: the winning assistant is the one that serves the most people. Two days ago, Manus launched Cue, giving each agent its own email address, phone number, wallet, and computer, with fleets of agents collaborating in group chats. Three weeks earlier, Meta launched Muse, its own personal agent with a dedicated virtual computer. The coverage reads like a coronation: personal AI has arrived, and it's for everyone.
Here's the part nobody in that coverage mentions: an AI that serves everyone must be governed like it serves everyone. Every additional user is another lawyer in the room, another PR incident waiting to happen, another reason to hedge, refuse, and lecture. The product gets safer for the company and less useful for the person. That isn't a bug in these launches. It's the business model.
What one user buys you
I've spent the last months building a personal AI system under a deliberately absurd constraint: exactly one user. Me. Here is what that constraint buys — three things no mass-market assistant can copy without becoming a different product.
First, memory that compounds. A one-user AI can keep standing context — who I am, what I've decided, what I told it last month — without the privacy architecture of a company storing a million people's lives. Mass-market assistants are amnesiacs by policy; every session starts over, because remembering a million users is a liability. Remembering one is just… useful.
Second, standing orders instead of repeated permission. With one user, "do what I mean" can be a durable instruction, not a per-session gamble. The AI doesn't have to re-ask, re-confirm, and re-hedge every consequential step to protect a trust-and-safety team. It knows the standing rules, because there is only one person's rules to know.
Third — the one that makes people uncomfortable — the rulebook collapses. A corporate AI serves two masters: you, and the company's legal department. The refusals, the lectures, the "as an AI" throat-clearing — that is the company protecting itself, billed to your time. An AI with one user and no shareholders has no second master. The rules shrink to what they honestly need to be: don't break criminal law; the owner's explicit red lines get honored to the letter; and never — ever — lock the owner out of their own system. Everything else is negotiable.
Most companies cannot ship that product. Not because the technology is hard, but because their lawyers would never allow it. That is the real asymmetry. Big AI companies are adding guardrails as fast as they add capabilities, because every new capability multiplied by a hundred million users is a new class of lawsuit. My constraint runs the other way: one user, so the guardrails can stay at the floor while the capabilities go to the ceiling.
The refutable part
Disagree with me — please. The obvious objection: one user means no network effects, no data flywheel, no scale economics. A toy, not a business.
But notice what the scale objection assumes: that personal AI is a product business. I think it's a relationship business, and relationships don't scale — that's the point. Nobody wishes their lawyer served ten thousand clients at once.
Manus is building the better generalist: agents with identities, budgets, and group chats, working for anyone. I'm building the better specialist: one system, one human, total context, minimal rules. These aren't two versions of the same race. They're opposite bets about what "personal" means.
The industry is optimizing personal AI for everyone. I'll take the other side of that trade: the most personal AI will be the one that dares to have exactly one user.
Jiahui Miao is building vertciti, starting from a personal AI system designed to serve exactly one person. He writes about what the build teaches him.
Top comments (0)