That'd be too restrictive, but to get improvement in a particular domain you definitely need to train specifically for it. Just cramming more random internet text in a bigger model has stopped being a effective way of scaling since at least mid 2023.