DeepSeek has achieved something very Special with attention. I haven't seen a model that's so proactive with its context. It doesn't just have full recall, it *inhabits* a context, feels at home there. It reminds me of Ant hype about Opus self-awareness. Or of test-time training.