Only browsed but this is really interesting and I'm glad it was published. I understand why a unified model is an interesting thing to work on but doesn't the

Only browsed but this is really interesting and I'm glad it was published. I understand why a unified model is an interesting thing to work on but doesn't the discovery of "modal-competition" suggest that at least short term it might be even better to train specialized models for each modality and some sort of modality-supervisor (glue code model)?