What are the primary bottlenecks in modern transformer architectures, and how would you propose addressing them for long-context tasks at Zyphra?
