The notebook
Things I've tried, things I've changed my mind about, and things I'm still working through.
A carefully maintained AI memory helps one person work with an account. Colleagues need that context too. My session archive grew into shared support investigations and account playbooks that people can use from their own AI sessions.
Frontier-adjacent inference is close to free right now, and the free version is not yours. I ran the same support cases through a cloud frontier model and a box in my house. The cost gap was 36x to 96x. The bigger gap is who controls the endpoint.
I split my inference by latency sensitivity rather than model size. Once you own the hardware, slow stops being a cost, and every job with no deadline inside the hour can move to the hours you are asleep.
My task board runs models at three points in its pipeline, and none of them decide anything. The rule that fell out of building it: put the model where being wrong is cheap, and put deterministic code everywhere being wrong is permanent.
An instruction to an agent is not a control. Before I let anything act on my mail, I built the layer that does not depend on the agent behaving: a read-only, append-only, encrypted copy the agent cannot reach, and a gate it has to pass before it touches a message.
I run shadow jobs to grade my local inference stack against the cloud models. The first thing the test caught was a cloud model fabricating a result from nothing. Reading output is not verification. A diff is.
AI is the fastest way to discover the structure of the tool you need. It is the wrong thing to call every time you use that tool. Once the structure is clear, graduate it off the AI runtime.
Someone, somewhere, has already solved your exact problem or close enough to it. The skill is learning to abstract your query until you can find them.
The operations reflex is to spot the friction you have stopped noticing and smooth it out. That reflex used to stop at the edge of software. AI moved the edge, and now it reaches the physical things in your house.
I left hardware for a decade. When I came back, the toolchain was unrecognizable but the architecture in my head still worked. AI makes that surviving skill useful again.
AI tools commoditized the building. What survives is the skill that was always the hard part: keeping the architecture of a system in your head and describing it well enough to get something useful out.
XKCD 1205 told us when automation was worth the time. AI collapsed that calculation. But the chart measured the wrong cost all along.
AI skills are the fastest way to prove an idea works. When other people need what you built, the skill has taken you as far as it goes.
The Google Sheet is the most popular application platform in enterprise. AI is finally changing that.