30 Days, Six Products, Two Dead Models, and the Part That Didn't Work

On July 9 we shipped Vinci Bozza.

Our own model. 9 billion parameters. Open weights. Free.

On July 12 I took it off the app.

Three days.

Piccolo, the one before it, got 13.

That happened in the middle of the fastest month of my working life. Everyone wants to talk about the speed. I want to talk about July 12.

The 30 days, in order

Every date below has a post on our newsroom, written the week it happened.

• June 26 - we re-introduced SimpleDirect as an AI lab. We had been a fintech.
• June 29 - Vinci launched. The web app, and Piccolo, our own 4B model.
• July 4 - Vinci Platform. API keys, usage, the developer side.
• July 9 - Bozza. Our own 9B.
• July 12 - both our models came off the hosted app. In their place, the strongest open-weight model in our evals, at our cost plus 15%.
• July 23 - Vinci Desktop. Agentic work on local files. Free beta, Apple Silicon only.
• July 29 - Forte and Fortissimo. Forte runs GLM 5.2, Fortissimo runs Kimi K3. New pricing: $20 gets you $50 of AI.

June 29 to July 29. Thirty days.

Six products shipped during it. Vinci Chat on the web. Vinci Mobile on iOS and Android. Vinci Platform. Vinci for Work. Vinci Code, a coding CLI. Vinci Desktop.

I built all six of them.

My engineering lead was not on the apps. He was on the models.

That division matters for the rest of this post.

Two of those six are not mine

Vinci Code is built on top of Pi. Vinci Desktop is built on top of OpenWork.

Without that open-source work, neither product exists. Or it exists as something worse that nobody would choose.

I am putting this here, above the impressive part, because that is where it belongs. Four of the six I built. Two I assembled on foundations other people wrote well.

Still a month I did not think was available. Just not the same sentence as 'I built six products.'

July 12

Look at the gap between July 9 and July 12 again.

That is not a pivot I planned. That is me reading our own evals and not liking the answer.

Piccolo tested one claim: that character survives across model classes. It did. Bozza asked the harder question, which is what character costs.

My engineering lead built both.

We ran the same evals. The decision to take them off the product was mine.

So we ran them against the strongest open-weight alternatives, on the workloads our users were actually running.

The result was not close.

Somebody else's model was better. Our users deserved the better one more than I deserved to be right.

They are still free on HuggingFace. Forever. They just came off the product.

I have written a lot of sentences this year. That is the most expensive one, and the one I am most sure about.

If you remember one thing from these 30 days, do not make it the six products.

Make it that the fastest month of my life included pulling my colleague's work off the product, three days after he shipped it.

I wrote every app in that list. I did not write the models. So the one call in these 30 days that actually cost something was a call I made about someone else's work.

He agreed with it. That did not make it feel better.

Nothing about my life looked like this

I expected a story about grinding.

I did not lose sleep. I did not sacrifice nights. I commuted every workday. Came home. Worked out. Ate dinner at a normal hour.

The only thing that changed was the speed.

That is what I keep turning over. Not that it was hard.

That it wasn't.

I have done the version of this that was hard. I know what earning a release feels like. This did not feel like that, and I have not decided whether that should worry me.

The agents are not the whole answer though.

I knew the domain before I started. The stack already existed. I had complete authority over the product and nobody to get approval from.

Strip any one of those out and the month does not happen.

What it would have cost before

None of this happens without the coding agents. Sometime late last year they crossed a line and I noticed something had shifted without having language for it.

This month is the language.

I have worked places where this exact scope was a roadmap. Not a sprint. A roadmap.

From the inside, what it took: seven to ten engineers, about twelve months, $700,000 to $1,000,000 if you price it in salaries.

Mine cost low thousands of dollars in API bills and subscriptions.

The conditions on that number, because it is useless without them.

That team would not only have built six surfaces. They would have built the agent underneath two of them, which I did not build, and which is most of the hard part. They would have built to a standard nobody has held me to, because nobody at scale depends on my version yet.

So it is not seven to ten engineers replaced. It is that much surface area, standing on work other people already did, at a quality bar that has not been tested.

Smaller claim. Also the true one.

No new revenue

Thirty days. Six products. Two models shipped and retired. A pricing model torn up and rebuilt.

None of those six products brought in meaningful new revenue this month.

The number of people using most of those six is small enough that I am not going to dress it up with a percentage.

I have said for months that building the software is not the moat. There is almost no daylight between OpenAI and Anthropic on model quality or harness quality or anything I can measure, and they each keep a loyal base anyway.

Products are table stakes. They qualify you to be in the room.

Then I spent 30 days proving I could build faster than I ever have, and it moved the actual problem approximately zero.

Because the actual problem was never build speed. I knew that before June 29. I went and did it anyway, and the month handed me back my own thesis with receipts on it.

What I think this means for you

I do not think engineers are finished.

I think the job moved and I am not certain where to. That is uncomfortable to say while holding a receipt that looks like a replacement story. I would rather say the uncomfortable thing than the clean one.

What I am more sure of. If your value was turning a spec into working software, that part is compressing fast, faster than the discourse admits.

If your value is knowing what to build, and getting a stranger to care, you got more valuable.

I shipped six products in 30 days and killed two models doing it.

I still do not have a distribution engine.