In these areas, the model performed better than even specialized image captioning models (e.g. Flamingo, OpenFlamingo) despite seeing less text in its training data, Meta claims. But what about bias?