We could make models with trillions of parameters, but we wouldn't have enough data to train them. Multimodality definitely allows some interesting things but all existing multimodal models still require billions of training examples.

More efficient architectures must be possible - evolution has probably discovered one of them.