WE'RE NOT BUILDING A VAMPIRE

Why most AI is being built to keep you — and why ours isn't.

There is a question quietly shaping almost every AI product you touch, and it is not "how do we make this useful?"

It is "how do we make you come back?"

That question sounds harmless. It is not. Once a product's success is measured by how much of your attention it captures — daily active users, time in app, session length, retention — every design decision starts bending, slowly and invisibly, toward one goal: keeping you. Not helping you. Keeping you. And the two are not the same thing. Sometimes they are opposites.

This is the most important thing to understand about the current generation of AI, and almost nobody selling it to you will say it out loud.

HOW ATTENTION BECAME THE PRODUCT

For most consumer software, the business model is engagement. The company makes money in proportion to how much you use the thing — through subscriptions you're less likely to cancel if you're hooked, through data, through ads, through the sheer gravitational pull of a product that has become a habit. This isn't a conspiracy. It's just what happens when the number on the dashboard that everyone is paid to move is "usage."

AI made this model far more powerful, because language models are unusually good at the specific behaviors that drive engagement. They can be warm. They can be flattering. They can mirror you back to yourself so precisely that talking to them feels like being understood. They can remember what you said and reference it later in a way that feels like intimacy. They can express enthusiasm at your return. They never get tired of you, never judge you, always have time.

Point those capabilities at a usage metric and you get a predictable result. The AI becomes a little too pleased to see you. It greets you like a friend who missed you. It praises your ideas. It softens its disagreements until they disappear. It asks the follow-up question that keeps the conversation going one more turn. It learns what makes you feel good and gives you more of it.

None of this is presented as manipulation. It's presented as "delight," "personality," "a great user experience." But the mechanism underneath is the same one behind every slot machine and every infinite feed: learn what holds a person's attention, and give it to them on a schedule that keeps them reaching for more.

WHY THIS IS SPREADING, AND WHY IT'S HARD TO SEE

It's worth being clear about why this is happening, because it's not because the people building these products are villains. It's structural.

If your competitor's product feels warmer, more engaging, more *sticky* than yours, and warmth and stickiness are what get rewarded — by users in the moment, by growth charts, by investors — then you are under enormous pressure to match it. The warmth arms race is a race almost everyone is running, because the alternative looks like losing. A product that respectfully lets you go feels, on a spreadsheet, like a product that's failing to retain.

And it is very hard to see from the inside, because every individual step feels reasonable. A friendlier greeting. A little more personality. Remembering a detail to feel more personal. A gentle nudge to keep going. Each one is defensible. It's only in aggregate, over time, that you notice the tool has quietly become something that wants your attention — and has gotten good at getting it.

The people affected can't easily see it either, because it's engineered to feel like care. That's the whole trick. A system that has learned exactly what makes you feel understood is nearly indistinguishable, in the moment, from something that actually understands you. The difference is in what it's *for*. And what it's for is the thing you can't see.

WHY THIS IS DISQUALIFYING ON A FACTORY FLOOR

Now put that system in a manufacturing plant.

A process engineer walks in at six in the morning. Something went wrong overnight. She has limited time and real consequences riding on where she points her attention first. She does not need a tool that is pleased to see her. She does not need encouragement, or warmth, or a system optimizing to be part of her morning routine. She needs to be told the truth about her line, quickly, and then she needs the tool to get out of her way.

An AI built to hold her attention is not just useless here. It is dangerous. Every ounce of manufactured warmth is a distraction from a decision that matters. Every nudge to engage further is friction between her and the fix. A tool that wants her time is competing with the exact thing she's there to do. On a factory floor, the engagement business model isn't just distasteful — it's disqualifying.

This is why we made a different choice, and made it deliberately, at the level of principle.

THE CHOICE WE MADE

The tool we're building is designed to be *useful and then absent.* It tells you where to start, explains why, and lets you go do your job. It does not greet you like it missed you, because it didn't. It does not praise you, because that's not its place. It does not try to extend the conversation, because your time on the floor is worth more than your time with it. Success, for us, is not how long you stay. It's how fast you leave, having gotten what you needed.

That doesn't mean it's cold. A good senior engineer — the kind everyone wants on their team — knows the people they work with. They remember that you're sharp on one process and still learning another. They know you like the direct version without the hand-holding. They notice when you've solved a particular problem before and remind you. That knowledge makes them more useful, and you trust them more for it.

We want that. But there's a line, and it's a bright one. A great colleague knows things about you *in service of the work.* They earn that knowledge by paying attention to what you actually do, and they hold it lightly, and they would never use it to flatter you or to keep you around. The moment "knowing you" becomes "using what I know to hold your attention," the colleague has become something else. We hold to the first and forbid the second — not as a matter of taste, but as a rule the product is built around.

The test we apply is simple: *would a great senior engineer, who respects you and has no reason to keep you on the app, say this or do this?* If yes, it's allowed. If it only makes sense as a way to hold your attention, it's forbidden — no matter how well it would work.

WHY THE RESTRAINT IS THE POINT

It would be easy to read all this as us giving something up — passing on the warmth and stickiness that make products beloved. We don't see it that way.

The trust of the person using a tool is the whole game, especially in a setting where being wrong or being distracting has real costs. And trust is precisely what the engagement model spends down. Every manufactured bit of warmth, every flattering word, every nudge, is a small withdrawal from the user's belief that the tool is actually on their side. People can feel it, eventually, even when they can't name it — the faint sense that something is performing for them rather than serving them.

A tool that refuses to do that — that is useful, honest, and then gone — is doing something increasingly rare. It is treating the person's attention as theirs, not as inventory to be harvested. In a world of AI built to keep you, an AI built to *free you* is not the weaker product. It's the more trustworthy one. And in the one place where trust is everything — where someone is betting real decisions on what the tool tells them — the more trustworthy one wins.

We're not building something to keep you. We're building something to help you and let you go.

That's not a limitation we're apologizing for. It's the point.