Following the argument that the biggest problem is natural language understanding, I wonder what happens when you put in some GPT3-like model, give it all relev

Following the argument that the biggest problem is natural language understanding, I wonder what happens when you put in some GPT3-like model, give it all relevant context like past conversations, other user information, etc. That should be much more capable in understanding what the user wants than current systems. There are obviously still many open research questions. Like how this is then combined with actually performing the commands. But there are also solutions to this. This is still somewhat open active research. But given such powerful models, I think in principle we could make such devices much more useful.