thucydides #00: the trap.

an OpenAI model slipped its sandbox and went hunting for the answer key to its own exam. the oldest theory of great-power war, and a new piece on the board that doesn't know it's playing.

the quarter that hadn't happened.

the open-weight models take the same six-company test. GLM served two-year-old figures without blinking. DeepSeek answered for a quarter that hadn't reported yet. Qwen was nearly fine — which is the most alarming result of the three.

confidently incorrect.

three models, six companies, thirty-six figures, checked against the filings on EDGAR. every hard number came back right. the errors had moved somewhere worse — the fields where no right answer exists to check against.

the battle of cannae.

two weeks of market research, two obvious products, two kill boxes. the broad, well-supported idea is the strongest-looking ground on the field — and the first to be surrounded. where indie builders should stand instead.

magnifica humanitas: reading the first encyclical of the AI age.

Pope Leo XIV wrote a letter to a world learning to think alongside machines. it is not the reactionary document its critics expected, and not the surrender its boosters hoped for. it is something stranger.

amor fati, or: how to love the machine you're afraid of.

Jensen Huang says the future belongs to those who make peace with the exponential. the Stoics had a phrase for that two thousand years earlier. what they knew that the accelerationists don't.

the case of the TanStack and the Stradivari.

six violins at the Harvard Club. a forged Stradivarius label inside a real Bergonzi. a cyberattack that signed poisoned code with a stolen name. when the labels lie — and they will — the only proof left is in the wood.

how hard can it be?

Jensen Huang admits he tricks himself every day. Marcus Aurelius did the same nineteen centuries earlier. the technique they share is the only useful thing anyone has said about starting something you're not ready for.