The conversation argues that biology needs a problem-first approach rather than a rigid faith in either scaling models or generating data. The speakers explain why imperfect metagenomic data can still help protein language models, why AlphaFold benefited from scientific inductive bias and existing structural data, and why protein folding remains incomplete because proteins are dynamic, disordered, and context-dependent. They call for richer measurements such as cryo-EM, models of broader biological systems, and methods for extracting scientific knowledge from model internals. The discussion closes by distinguishing model interpretability from calibrated, trustworthy behavior and by arguing that major health advances will require both incremental progress and ambitious 10x breakthroughs.