For me, self-hosted agentic LLMs will have "made it" if I can run useful/intelligent agentic workloads on 8GB VRAM (or less)