notes, mostly on making AI real
I work on the gap between what models can do and what deployed systems actually do. Most of what I write here comes from running that experiment on myself: frontier-quality inference on hardware I own, agents that survive daily use, and the behavioral limits of models in sensitive and corporate settings. Less about me, more about what I found.
Roughly nine out of ten enterprise AI pilots never reach production, and almost none of them die because the model was not smart enough. After a decade of watching institutions meet frontier technology, first in digital assets and now in AI, I think the failure lives somewhere specific, and it is not in the weights.
Something has disappeared from the process of building things. Not effort. Not difficulty. The queue.