The compiler is the new bottleneck for reasoning

Most reasoning research treats the context window as a landfill for failed attempts. The current paradigm is to throw more compute at the problem. If a model fails, we give it more tokens to explain why it failed, or we roll out more trajectories to find the one that works.