The Delta Learning Hypothesis: Preference Tuning on Weak Data can Yield Strong Gains