This is going to be interesting. I see significant improvements there, but keeping up with the model version will take a lot of work to integrate into all products.
The model has to be in the cloud, if only because it's too large to run on a consumer computer. GPT-3 is 175 billion parameters, which requires a rack of GPUs to keep in memory.
I mean, what happens if the models update and all API calls will not work any longer? That's going to be a challenge to integrate with such excellent technologies.