4 ms·SubQ 1.1 Card: Linear-scaling sparse attention with 98% retrieval at 12M tokens [pdf]2 points by mitchwainer 4mo agomitchwainer 4mo ago[flagged]