Summary: For most of this year my AI lab has run entirely on hardware I own: a local speech pipeline, local models on one GPU, a retrieval layer over my own notes, and a set of assistants that talk...

Summary: Every architect carries a coin. One face reads agent of change, the other reads agent of chaos, and which one people see often has less to do with the quality of your work than with how mu...

Summary: My desktop copy of AnythingLLM now has fifteen custom agent skills and five flows. They answer the questions I actually ask about the lab: how it's doing, what's on the GPU, whether the ba...

Summary: I ran a usage tool across my AI coding assistants and it told me four weeks of work was worth $2,941.30. It wasn't a bill. Working out why it wasn't a bill meant actually mapping how four ...

Summary: Helen, the default voice of my lab assistant, picked up three new abilities this week: she can run a small set of real operations by voice, she can answer general questions instead of defl...

Summary: This is the last post in the series, and it's about the thing all the rest of it was for. At seven in the morning and seven in the evening, Helen tells me how the lab is. The interesting ...

Summary: Helen can restart services on my machines. That made me uneasy enough to spend more time on the boundaries than on the capability itself. This post is about where the limits live, and the...

Summary: For most of this series the assistant has been an it. A pipeline. Somewhere along the way it stopped being one, and this post is about that transition - how the lab ended up with four dist...

Summary: The lab has a service that plans training sessions. It's the one people assume is most obviously a job for a language model, and it's the one where I most deliberately didn't use one. As ...

Summary: Everything so far has been about the lab looking inward - its own health, my own notes. This post is the one place it looks outward: a small data lake of public sources, collected on a sch...