MoE feels a lot more like engineering to me. You're routing around the problem rather than actually solving it. The real math gains are things like quantization schemes that change how information is actually represented. Whether that distinction matters long term probably will depend on whether we hit a capability wall first or an efficiency ceiling first.