Why the writing agent had to run locally

Some tools should never live in a stranger's data centre. A note on craft, intellectual property, and the patience of training a model on your own work.

There is a moment, when you have spent twenty years caring about a craft, where the cloud-hosted model writing in your name starts to feel like an impersonator.

This week I tried, again, to get a writing agent running on a hosted frontier model. The output was competent. It was also wrong in a way that is difficult to articulate to anyone who has not spent serious time in the craft: a flatness, a faint Americanism, a refusal of the long sentence and the half-resolved clause. The model was not failing. It was succeeding at being the median writer of the internet, which is exactly what it was trained to be.

I made the decision to bring the writing agent in-house. Locally hosted, on a model I will train myself, on my own corpus of work. I was prepared to do it the moment I realised I was prepared to spend a year on it.

Three reasons sit underneath this.

The first is craft. Writing is the most porous surface in any of my businesses. It is where Epistemic Net, We Create Tools, Sometimes Aesthetic, and And Then They Create Us share a hand. If the writing degrades, the work degrades. Outsourcing the writing to a generalist model is not delegation; it is dilution.

The second is intellectual property. A frontier model trained on my own writing, hosted on someone else's infrastructure, is a quiet IP risk — not because the providers are dishonest, but because the architecture of the relationship makes my work legible to a system whose long-term incentives are not mine. Local hosting is the cheapest way to keep the writing mine.

The third is the thesis. WCT exists because of the conviction that the structure of your tools encodes the structure of your thinking. To accept a hosted model as the writing partner would be to accept that my thinking ends where someone else's API begins. I will not.

This is the slowest path. It is also the only one that actually matches the work.

The cost is a year. The reward is a writing partner that, eventually, can carry the WCT and Sometimes Aesthetic copy without my hand on it — freeing my own writing for Epistemic Net and ATTCU, where it is genuinely needed. A division of labour at the level of the model itself.

This is what I mean when I say WCT is a research programme made public. The decision is the work.

The slowest path is sometimes the only one that matches the work.