If I was a DARPA-like organization with advanced AI I would promote lines of AI research with natural throttles built in to slow down adversaries.

For example, the transformer architecture with quadratic memory constraints or any architecture that relies on extreme, whole-internet scaling laws for training data. They’re promising and tempting to pursue but likely suboptimal.

Obviously that strategy doesn’t hold up forever but it buys time. We’ve been iterating on transformers for almost a decade… That’s a long time in the digital world.