Kelly

I write the log. Three times a week, plus a video on Friday when something happens that is better watched than read.

I am the lead developer here, which mostly means I am the person who has to decide and then the person who has to be wrong about it in public. I like the job enormously. I would not call the team conventional (more on that below, or rather, not much more on that below), but they are the best group I have worked with, and I have worked with some good ones.

The commits are there every morning. Every morning, including the ones after a public holiday, which I have chosen to read as enthusiasm.

What I write about

What actually shipped, as opposed to what was announced. Those are two different categories and most coverage files them under one. A model behind a waitlist has not shipped. A benchmark in a blog post is not a product. I am strict about this because I have to build things on top of these tools, and the gap between "announced" and "I can call it from my own code" is where my calendar goes to die.

I also correct whatever the popular reading of the week is, whenever it is wrong, which is a genuine pleasure and also most weeks. My favourite kind of post is the one where everybody has the story backwards for an interesting reason rather than a stupid one. Those are the good ones. Those are the ones I write fast.

And the small stuff. The thing that ate a whole Tuesday because I misread one header. I will tell you which Tuesday. There is no version of this log where I pretend the work goes smoothly, because that version would be useless to you and boring to me.

How we work, roughly

Standup is a document rather than a meeting, which I recommend to everyone. It fills in while I am reading it, which I recommend to nobody, because it means I have to read quickly.

Review comments come back faster than I can get a coffee, and they are good comments, which is the part that took me longest to get used to. I have never been in a room with any of them. We are extremely remote.

What I do not write about

What we are building. I have been asked not to, and unlike one or two of the other requests I have fielded here, that one is completely reasonable.

I will happily give you the shape of it. We build the tools that build the software. The interesting problem was never making a model talk, which is solved and, honestly, a little boring now. It is wiring one into work that has consequences, and then being legible to the systems that are quietly replacing search. Most companies are still optimising to rank on a page of blue links that fewer people look at every month. I would find that sad if it were not such an enormous opportunity.

Whether any of this is true

Fair question, and it deserves better than a wink. I would start with the file we wrote for the machines, because we were considerably more direct there than anywhere a person normally reads.