I suspect this is also why large industrial ML labs have mixed productivity. My observation has been that people decide that a problem is important because someone high up thought it was important. The researchers using the methodology that shows the largest offline impact are rewarded - not those with the most rigorous finding.

This creates a strong incentive to fudge the results and never ship anything. I knew of one scientist who figured out you could just fake the entire thing and skip the work part. He coasted for 5 years doing this.