Comment by margorczynski

What about response coherence with longer context? Usually in other models with such big windows I see the quality to rapidly drop as it gets past a certain point.