Despite significant progress in applying artificial intelligence to materials discovery, the field has yet to produce a breakthrough equivalent to AlphaFold's impact on protein structure prediction. Prof. Heather Kulik, a pioneer in combining computational tools with AI for materials science, argues that success requires deep integration of domain expertise with machine learning techniques rather than relying on hype-driven models alone. Kulik's research demonstrates both the promise and limitations of AI in this space. Her team recently used AI to design polymers that proved four times tougher than expected, with the algorithm discovering novel quantum mechanical effects that surprised even the synthesizing scientists. However, current large language models still struggle with fundamental tasks that materials scientists find trivial—such as designing ligands with exactly 22 heavy atoms. Tests conducted months after recording revealed that both Claude and ChatGPT failed consistently on this constraint when designing metal-organic framework ligands, despite succeeding on related protein design tasks, suggesting fundamental gaps in how AI systems reason about materials versus biological domains. The takeaway for AI practitioners is clear: success in materials science depends on rigorous laboratory validation and maintaining healthy skepticism about model capabilities, not on chasing flashy benchmarks or waiting for a single transformative breakthrough.