Tab 1
Scout, Shortlist, Attend
MiniMax treats the KV cache like an internal retrieval corpus: block it, score it cheaply, keep the nearest block, then spend exact attention only on the shortlist.
Tab 2
Pattern Lab
Switch the matrix. Dense attention paints the whole causal triangle; MiniMax tries to preserve the useful parts while keeping the kernel shape regular enough to matter on real GPUs.
Tab 3
What the Speedup Means
The hard numbers in the episode sit on one side of the story. The other side is where the uncertainty still lives: cross-GPU portability and rare distant evidence recovery.
Tab 4
Where MiniMax Sits
Not the first sparse-attention idea, but one of the clearest attempts to strip routing down until deployment stops fighting back.