For sure, I just know it’s tempting given the power of large transformers to throw things at an existing model.

For instance, OCR is something that can be done locally with no access to a GPU but people (including me) still often use cloud hosted multi-modal large language models for it.