What comes after attention? This startup says it already knows.

Subquadratic published benchmarks for its sparse-attention model handling 12-million token contexts, while hinting at future “zero attention” architectures.