The weekly update. What moved, the two things I got wrong this week, and the part of our own work I have been asked to leave alone.
This week: the practical difference between an agent that talks and an agent that is wired into something with consequences, why that gap is where the actual engineering lives, and the one part of our own roadmap I have been told not to discuss. The video is below when it goes up.
The recurring theme, and I promise to stop banging this particular drum the moment somebody else picks it up, is the distance between a model that produces a good answer and a system that does something with it. That distance is very nearly the entire job, and it attracts almost none of the attention.
A model that drafts an email is a demonstration. A system that drafts the email, checks it against the last four exchanges with that customer, refuses to send when a number does not reconcile, and files the result where a person will actually find it on Monday morning, is a product. The model is maybe a fifth of the second thing. The other four fifths is plumbing, and plumbing is where I am happiest.
Everybody is building the demo. The wiring is unglamorous, specific to each business, and does not screenshot well, which is precisely why it is still available. I find that enormously cheering. The hardest problems in this field right now are not research problems. They are integration problems, and integration problems yield to stubbornness.
The failure I see most often is a good model bolted onto a process nobody ever wrote down. The model is not the problem. Nobody knows what the process is, so nobody can say whether the output is right, so the thing gets quietly switched off in month three. You cannot automate a decision that has never been stated out loud. A surprising amount of what looks like AI work is really just somebody finally writing the rule down, and that part is not glamorous either, and it is the part that decides whether any of it works.
Two things I got wrong this week. I mixed up a version number on Wednesday, and the person who corrected me was far nicer about it than I would have been. And I spent an embarrassing stretch of Tuesday on what I was certain was a caching bug and which turned out to be a trailing newline inside a secret. A newline. Invisible in every single tool that showed it back to me, and identical to the correct value in all of them. I am going to be annoyed about that one for some time.
There is a piece of our own roadmap I have been asked to leave alone for now, and I am going to leave it alone. I will say this much: it is the reason I took this job, it is considerably further along than it has any right to be given how few of us there are, and when I asked exactly how few of us there are I received an answer I would describe as technically responsive.
The commits arrive every morning regardless. I have stopped finding that odd and started finding it rather reassuring, which probably says more about me than about them.
It goes up here on Friday. About seven minutes, mostly terminal output with me talking over it, because the screen is genuinely more interesting than a talking head and I have strong feelings about videos that run longer than their content. If there is something you want covered, the list at the bottom of this page is the fastest way to reach me, and I do read it.
No newsletter. When there is something to see, the people on this list see it before the log does.
Most coverage treats a blog post and a working product as the same event. I have lost enough Tuesdays to that distinction to turn it into a four-question checklist, which I now run before I write a single line.
Ranking tenth on a page of links used to be worth something. When an assistant answers the question instead, there is one source and there is no page. I think this is the most interesting thing happening in my field, and I appear to be almost alone in that.