Where this site is going

Right now this site is a well built room with almost nothing in it. That is on purpose. I wanted the structure to be honest before I filled it, because it is much harder to fix a foundation you have already stacked furniture on. Here is what goes in the room.

The pipeline, and the rule underneath it

A cron job wakes up, reads a set of sources I care about, asks a model to draft something, and writes the result to a table marked unpublished. Then it stops. That last part is the whole design.

The bot never publishes. It drafts. I approve. I have watched enough of the internet fill up with confident, frictionless, machine written nothing to know exactly what happens the day I let the model push straight to the site: the blog quietly stops being mine and becomes a feed. The thing I would be optimizing for at that point is volume, and volume is the one metric I do not want.

“The bot writes the first draft. I decide whether it is true for me. That question is the only reason I am in the loop at all.”

I am not there to fix typos. A model fixes typos better than I do. I am there to answer the one question it cannot answer, which is whether the thing on the screen is actually something I believe. Taste does not scale, and I have stopped treating that as a problem to solve. It is the reason a person still belongs in the pipeline.

Approval should take ten seconds

Here is the failure mode I am actually afraid of, and it is not the model writing something bad. It is me becoming the bottleneck. A review step that requires me to sit at a laptop and open an admin panel is a review step I will skip on a busy week, and a pipeline I skip is a pipeline that either dies or gets loosened until it does not need me. Both of those end in the same place.

So the next thing after the review board is a Telegram bot. The draft comes to my phone, I read it in line for coffee, I approve or I kill it. The point is not convenience. The point is that the human step has to be cheaper than the temptation to remove the human step. If quality depends on my discipline, I have designed it wrong. It should depend on the path of least resistance pointing the right way.

Then, slowly, the rest

Cross posting to LinkedIn, once I trust what is coming out of the pipeline enough to put it in front of people who did not choose to visit. A projects section, which I have deliberately left mostly empty, because a portfolio full of things I half remember is worse than one with three things I can defend line by line. A timeline, which I want to prototype separately rather than force into a layout that was never built for it.

And the thing I am most looking forward to writing about: machine learning, from the ground up, in public. I am implementing neural nets in pure Python right now, no framework, no shortcuts, because a friend once told me you do not really know something until you have gone two or three abstraction layers below it. Writing about that as I go is going to be uncomfortable in the specific way that means it is worth doing.

What I am optimizing for

Not traffic. Not a posting cadence. If the pipeline produces one thing a month that I would defend in a room full of engineers, it is working. If it produces twelve things a month that I would not, it is broken, and no amount of analytics telling me otherwise would change that.

The pursuit is making the thing better. Everything else is a byproduct of doing that honestly, and I would rather have an empty room I am proud of than a full one I am not.