The US election happened Tuesday, and the polling models faceplanted again — the out-of-distribution lesson, now officially a genre. The industry’s postmortems this time aren’t about servers; they’re about feeds: filter bubbles, engagement optimization, fake-news economics. Those are our systems. The reckoning about what recommendation engines optimize for has left the conference track and entered the congressional calendar. More on that as it develops, which it will, for years.
F1’s title fight, meanwhile, is compressing beautifully: Hamilton won in Mexico a week and a half ago to keep the math alive, but Rosberg only needs to cruise. Two races left — São Paulo this Sunday, then the Abu Dhabi decider. The group chat has pre-committed to no eulogies until the math is dead. (I have drafted one anyway. Blameless.)
TIL: Brier scores, from the election-forecast postmortems — how you grade a probabilistic call after a binary outcome. The forecasters’ “we said 30%, and 30% happens” defense is technically correct and emotionally bankrupt, and it’s going straight into how I phrase risk estimates at work. “Low probability” is not a promise. Write the number down; grade the number later.