On publishing in design

We are publishing the company before the product. The four pillars are on the site. The cores are on the site. The pipeline is on the site. None of them have shipped.

This is deliberate.

The default move for an infra company at our stage is stealth. Pick a code name. Talk to a few design partners under NDA. Pre-announce six weeks before the first release. Ride the surprise.

We are doing the opposite. The roadmap is public. The voice is fixed. The principles are stated in the present tense. The eval plan is named. And we have given ourselves a date by which the first eval has to be signed cold.

Anyone can call us on any of it.

The stealth default and why we are skipping it

Stealth optimizes for surprise. Surprise is a one-shot resource. You spend it at launch, the press cycle runs, and a quarter later the surprise is gone and what is left is the product.

We are not selling surprise. We are selling a substrate teams build on for years. The half-life of the surprise is shorter than the half-life of a cognitive core in production. The math does not work.

The other thing stealth optimizes for is option value. If we never say what we are doing, we can pivot without leaving fingerprints. But option value bought that way is expensive. It buys flexibility for us at the cost of legibility for everyone else. The people we want to talk to — researchers, engineers shipping the frontier on a budget, teams whose inference bill is a problem — they need to know the shape of the thing before they can decide whether to engage. Hiding the shape costs us those conversations.

So we are spending the option value down on legibility, while we still have something to be legible about.

What “in design” actually commits to

“In design” is not vapor. We do not get to say “in design” about anything we have not actually committed to.

What it commits to, concretely:

• A public roadmap. Three core sizes named. Edge, workstation, server. Sizes stated as targets, not specs.
• A frozen voice. Mantras, terms, forbidden phrases written down in a brand bible that anyone can read.
• A stated eval plan. Frozen evals on every commit. Drift, regression, and refusal tracked per release. Reports reproducible from a single hash.
• A date by which the first eval is signed cold. Not a launch date. The date the trust language earns the right to be on the site.

If a competitor reads any one of those lines and finds it vague enough to mean anything, we have written it wrong and we go back to the file.

The rule is simple. The site is the working roadmap, not the marketing roadmap. Future tense where the work is future. Present tense only where the work has happened.

What it costs

We pay for this. Several ways.

Competitors see the shape of the cognitive core line a year before the first weights ship. They can plan against us. They can ship something with our shape and call it theirs. The trade we made is that we would rather be copied with attribution than not copied at all. A line that nobody copies is a line that nobody noticed.

Customers expecting a quarterly roadmap update get the unedited working roadmap instead. Some of them want a cleaner story than that. A neat product page, a price, a button. We do not have those yet. The people who want the neat story will bounce, and that is fine, because the neat story would be a lie at this stage and they would bounce harder later when the lie surfaced.

Investors expecting a hockey-stick narrative get principles before product. Some of them will pass on us for that reason. That is also fine. The investors we want are the ones who can read a principle, recognize the work it implies, and price the company off that. The other kind we cannot serve anyway.

Some readers will bounce. That is the trade.

What we get back

What we get in exchange is the part nobody talks about when they recommend stealth.

Compounding feedback. Every time someone reads the site and writes back disagreeing with a principle, that is information we would not have had in stealth. It compounds. By the time the first eval is signed, we will have argued the four pillars with twenty engineers we did not have to pay. The product gets sharper, in public, before it exists.

Honest pull. The people who reach out after reading the site already know what we are. They are not asking us to be something else. The conversation starts in the middle. Sales cycle shrinks. Onboarding shrinks. The first ten users are people who actually want the substrate.

A team that cannot quietly retreat from a stated direction. This one is for us. Stealth makes it easy to drift. When the roadmap is private, you can revise it on a Tuesday and nobody notices. When the roadmap is public, drift is a public act. It forces discipline. It makes the work harder to fake.

A brand that, on day one of shipping, already looks like the thing we claim to be. The site has been there for a year. The voice has been there for a year. The principles have been there for a year. When the first eval lands, it lands into a context that has been built for it. No rebrand. No repositioning. The story does not start at launch; the story has been running, and the launch is just a point on it.

Compression infrastructure especially

For a compression infra company, publishing in design is not optional. It is the only way the work gets shaped right.

Compression is downstream of frontier labs and upstream of every team trying to ship the frontier on a budget. We do not get to pick what frontier teams release. We do not get to pick what target hardware our users will deploy on. The only way to be useful is to be in the conversation while the conversation is still moving.

If we publish in stealth, we miss that conversation. We ship something we designed in isolation, in a year, against a frontier that has moved, for hardware that has changed. The work is the conversation. The blog is part of the product. The roadmap is part of the product.

Compression is cognition. We have to be in the room while people are deciding what cognition they need.

The rule we keep

There is one rule under all of this. We restate it here so it can be quoted back at us.

Nothing on the site is a forecast we cannot defend cold.

Sizes are targets, not promises. When we say “targeting 1.5 – 2 B params” we mean the design budget, not the spec. The shipped model might be 1.8 or 2.1. The number on the site is the target the design was built against, not a number we will hit no matter what.

Dates are commitments only when the eval is signed. We do not have launch dates on the site. We have a single internal date for when the first eval is signed cold. That date will become public the moment it passes. Until then it is a commitment to ourselves, not to readers.

The roadmap is the working roadmap, not the marketing roadmap. If we change our minds, the page changes. If we kill a pillar, the page changes. The page is the artifact, not a static document. Read it with that in mind.

That is the contract. The site is the working draft of the company, published in real time. Anyone who reads it knows what we are doing, what we are not doing, and what we have not yet earned the right to claim. The thing we are building is downstream of that contract. We think the contract is more interesting than the launch.

We will know in a year whether we were right.