Really enjoyed @SemiAnalysis_ deep dive on @TileRT_AI & ultra-high interactivity inference. It's a great explainer of why memory is the bottleneck for low latency decode, and how much engineering it takes to get a GPU to behave like a dataflow machine.