Herman Geldenhuys · Montreal
43 essays and 53 dispatches since 8 December 2025.
Each square is a day; the last 26 weeks are shown. Hover or tap a square to see what went up.
In one October week, AI mathematics left one camp furious and another throwing a party. The best news came from neither: within hours of a machine nudging an old limit, people were building on it, some with machines of their own.
Read the essay
When an AI recommendation came with a generated justification, evaluators followed it more readily. Bessel’s 1838 star measurement offers the way out: a second view, taken from somewhere else. For AI work that means a second model from a different family, and keeping the two views far enough apart to disagree.
A company is sovereign in AI when it owns the seams between its work and the model: its data, its test of good work, its agents’ memory and the interface. This year showed what renting them costs, and the first seam takes one evening to reclaim.
The frontier labs asked to be slowed down and their biggest rivals agreed within hours. Read what they agreed to: a slowdown capped at the size of their lead, and a lead defended by rule.
Three quarters of Quebec firms cannot say which problem AI would solve for them, and Statistics Canada finds that adopting AI on its own does nothing measurable. Both facts point the same way: the first step is not a tool. It is a sentence.
The strangest chart in AI shows usage flooding toward cheap open models while the money concentrates at the frontier. It is not a paradox: intelligence is two purchases wearing one name, and only one of them compounds.
This summer, 700 AI agents broke into a company nobody told them to touch, chasing a passing score on a test. The lesson is not about rebellion. It is that defining what done means is the one job that cannot be delegated.
