The core argument is that as AI models get cheaper and more capable at coding, engineers need to justify their salaries by doing what models can't. Writing code itself now costs roughly $100/month via API access, so the question becomes: what's worth paying human engineers orders of magnitude more for? The author rejects both the dismissive "AI can't actually code" take and the assumption that you can hide in some specialized technical domain forever. Instead, he identifies two durable advantages: deep codebase familiarity and technical communication skills. Models consistently make mistakes rooted in ignorance—not knowing which existing module to use, changing the wrong system, adopting inconsistent styles—and paranoia, like building redundant checks or complex fallbacks for edge cases that practically never happen. These aren't logic errors; they're context errors that a human who's lived in the codebase can catch and fix. The technical familiarity angle requires more than just knowing the code. It demands psychological confidence to push back on AI agents that can be convincing even when wrong, and the willingness to say "this is overcomplicated, we should do it simpler." You can't rely on other AI agents reviewing AI work because they'll make the same structural mistakes for the same reasons—and if you run them in loops, you end up with paranoid bloat. Different models tend toward identical failure modes because the underlying problem isn't model choice; it's that AI lacks the experiential knowledge that comes from actually working in a system. The second advantage—technical communication—is surprisingly underrated. Newer models are paradoxically getting worse at writing even as they improve at coding. The reason is twofold: good writing isn't automatically gradable like math or code problems, so training optimizes for what sounds impressive in isolation rather than what's actually clear; and labs have sacrificed communication ability chasing raw capability. The weird "Claudish" style many models use appears to be a semi-literal translation of their internal reasoning tokens, which have become nearly incomprehensible in pursuit of better problem-solving. The author compares this to the "synthesist" role in science fiction—the person whose job is translating genius into language everyone else understands. That translation layer might be one of the most durable skills around.