Building Wren in the open

Let me start with the sentence most AI companies won't say plainly: we did not build the model. Wren runs on hosted frontier models — the same kind of foundation any product company can license — and everything that makes Wren Wren is what we built on top: the streaming interface, the Frames panel, Projects, the model ladder, the voice. We're a product company. We're not a lab, we don't do frontier research, and we're not going to borrow the authority of people who do.

That admission isn't modesty. It's the foundation of how we've chosen to build, and this post is about what follows from it.

What we are, and what we're not

There's a certain grammar to AI marketing right now — "our AI," "our model," a lot of first-person possessive over capabilities the company licensed rather than created. We understand the temptation. Sounding like a lab is good for fundraising and better for press. But it's a small dishonesty, and small dishonesties compound. Once you've implied you built the intelligence, you have to keep implying it, and pretty soon your whole public voice is load-bearing on a thing that isn't true.

So we drew a hard line early: Wren describes itself accurately or not at all. The model ladder is named honestly — Seed, Field, Grove, tuned tiers on hosted models, and we say exactly that on the page. We don't claim safety-lab authority we haven't earned. We don't invent enterprise scale or uptime we don't have. The product has to be good enough to stand on what it actually is.

Building to be watched

If you can't out-lab the labs, what's left to compete on is craft and trust. Craft we've written about elsewhere — the manuscript philosophy, the seam between thinking and making. Trust is harder, because it's not a feature you ship; it's a thing you accumulate by being catchable when you're wrong. So we designed the company to be watched.

Two habits carry most of the weight:

• Field Notes. Dated, plain-spoken posts about what we're learning as we build — what worked, what didn't, the ideas we tried and dropped. Not a launch feed; a lab notebook, kept in public. You can read them at /notes.
• Public evals. A weekly self-eval, run the same way each time, published whether the numbers flatter us or not. We wrote up how the harness works; the live results are at /evals. Our worst row sits in the same font as our best one.

Both of these cost us something. Field Notes occasionally document a mistake we'd rather not advertise. The evals page has a number we're not proud of, in public, every week. That's the point. A company that only shows you its wins is asking for a kind of trust it hasn't done anything to deserve.

No hype, as a discipline

You'll notice what's missing from our writing: no "revolutionary," no "game-changing," no breathless launch theater. Partly that's taste. Mostly it's discipline. Hype is a loan against future disappointment — you borrow excitement now and repay it when the product turns out to be a normal, useful thing rather than a miracle. We'd rather never take the loan. When we ship, we tell you what it does, what it's good for, and where it falls short, in a normal voice. If that's less thrilling than a launch video, good. Thrilling wasn't the goal.

Why this is the better bet

Building in the open is slower and it removes some easy moves from the table. But it compounds in the right direction. Every honest Field Note makes the next one more credible. Every published eval — including the bad one — makes the good numbers believable. Over time you build something rarer than a growth curve: a reputation for saying true things, which is the only durable moat a product company on someone else's models can actually dig.

We think the future of AI has room for companies that aren't labs — that take extraordinary underlying models and turn them into tools people trust enough to keep open all day. That's the job we picked. We intend to do it in daylight. If you want to watch, the Field Notes and the evals are the window, and they're always open.