From a developer’s perspective, it’s currently not feasible to build production-grade applications of significant scale (e.g., 50k+ lines of code) entirely with AI. The limitations are fundamental: automated testing is not reliable across frontend or backend, tooling for validation is immature, and there’s no true modularity. In practice, AI cannot consistently generate isolated, testable components that integrate seamlessly into larger projects. These challenges are architectural rather than superficial—switching databases (e.g., Postgres vs. MongoDB) does not address the core issues. When claims suggest otherwise, it’s often more of a sales pitch than a realistic assessment of current capabilities.