The guy in the video didn't give a reference to the "new Harvard study", but I found a study from 1978 that seems remarkably similar:
Kerr, R., & Booth, B. (1978). Specific and varied practice of motor skill. Perceptual and Motor Skills, 46, 395-401.
In this study, children between 8 and 12 practiced throwing a beanbag at a target. The "specific" group only practiced with the target at one set distance, (for the 12 year olds it was 4 ft), while the "varied" group practiced throwing either at a nearer (3 ft) or a more distant (5 ft) taget, but never at 4 ft. The test at the end was to throw at a target 4 ft away, and the surprising result was that the "varied" group, which had never practiced at 4 ft, performed better than the "specific" one.
A later study, with adults, was performed in 2018 by ChƩla R. Willey and Zili Liu. You can access it here. They came to a more nuanced conclusion:
Our results suggested that after 5ā7 weeks of training, the specific groups tended to undershoot at longer distances and overshoot at shorter distances while the varied group tended to center their throws around the target at all distances. However, the overall magnitude of error (regardless of over- or undershooting) was similar across groups. We found some support for the hypothesis that the varied group could better generalize to untrained distances, but this advantage was found mainly for the longest distance and disappeared by a posttest held two weeks after practice.
For me, applying these findings to working on leaps in piano playing, the bottom line is this:
Variation is certainly good, but I see no need to purposefully play wrong notes in order to achieve this variation. It's rare to find a piano piece containing leaps that are all between the same two notes. La Campanella is a good example: the leaps are continuously changing in size, so if you practice it as written you will be experiencing enough variation without having to play a load of wrong notes, that sound ugly anyway!