At the end of the day, it seems like LLMs create a tool that you can write software yourself to get the tool to do things for you. It seems like it will not get rid of any jobs, but create new ones for software developers to automate prompt engineering.
I agree with you on the current generations of models. With more feedback loops like executing code and adapting based on the results the scope of capabilities will get a lot wider.
Maybe. Certainly this technology is profound. But I was thinking about this in the context of the library in developing currently to do spatiotemporal modeling. What would an LLM do to use this model? At what point would it be like star trek and you would say "computer, can you identify a safe landing site for our shuttlecraft where it won't be hit by lightning on the lightning planet we are orbiting." And then somehow the LLM knows all the data available, the correct physical model to use, the correct statical model to use, the characteristics of the shuttlecraft that might make it dangerous in some areas even though the physical model says there's no lightning there, etc. Like at what point does it have the context to do all of that, not hallucinate anything, and give you a confident result that you can then trust. I feel like that is a long way off and the current hype is just hype.