The map beats the model
Siemens' AI factory runs on 100+ algorithms, not 100 agents. Why the map of a company's data matters more than the model, and why small models win.
Essays from the people building Sentinel. Mostly about decisions in heavy industry where the technically right answer and the commercially right answer are not the same, and what happens depending on who is in the room.
Every piece has a number in it we can defend, or an incident one of us was there for. No customer is named. Roughly every two weeks.
Siemens' AI factory runs on 100+ algorithms, not 100 agents. Why the map of a company's data matters more than the model, and why small models win.
Anthropic's agents spend about 80% of their time waiting for a person. In a plant, the person decides when an agent may act. Why we start with drafts.
Ask an AI vendor what your problem is and the answer is AI. I am one. We sorted a client's list of checks to automate. About half were reports, not models.
I built a trading bot in 2020 that has never placed an order. What that says about AI agents that trade on their own, and about where the person belongs.
A drilling rig takes solids out biggest-first, for a reason you can multiply out. Brute-force AI runs the line backwards, and the token bill is starting to show.
A one-day diagnosis, a fix that waited four months for the crop to end — and what “weeks to hours” leaves out.
A 99% forecast interval of my own, broken on day six — and why the number to read in the ERCOT queue is not the one everyone is quoting.