Current methods for training vision-language models have shortcomings, and the discussed task provides a clear incentive to improve them.

“And I think that just points to a lot of shortcomings in the ways that we're training vision language models and provides like a clear incentive of, uh, how we should be making these models better.”