I agree with most of what you said.

Way chatgpt is now its overusing LLMs for things that do not need LLMs.

Its wrong tool for the job of, say doing math.

And like you noted, actions allow it not to do it.

You know what is hyper inefficient? How when it makes action call and get's markdown back it consumes it in to context and then freaking generates it again to show... It reinterprets everything trough llms. Even links to show to user...

For me solution is to marry it more with code. Allow it to output "references with types"

If it got markdown from acton and user want's to see it just send it to actions of "show markdown" instead of going trough llms for gods sake....

Second issue for me is reusable code. I can save in knowledge files python script that can be run from code interpreter... But that is not perfect...

ChatGPT needs more server and client code hosting with better integration.

But at the same time... For non developres. GPTs is a marvel. It allows non developers to create chatbot prototypes in minutes. My advice to people wanting to build something is to prototype it in ChatGPT first. It would be great if they were expanding the platform or allowing others to expand it in more ways.