Either way, my issue there is actually with the fact that the LLM can only predict what a human might say in response to the prompt andz as far as we know, don't have any way of logically thinking through the new solution or modelling reality to understand if the idea might actually work.
LLMs would have to completely go outside the bounds of how they were designed to work for that to happen. They're prediction engines. They could absolutely create a novel idea when predicting what a person might say, but LLMs have no way of working through the logic behind the idea or modelling reality to see if the idea might work. You could task an LLM to design a thousand novel gene therapies or something similar, but at the end of the day humans would have to sift through testing or logically invalidating each option generated. Until that step is automated and paired with an LLM, and almost certainly other structures along the way, it will never be artificial intelligence.
You just described how LLMs have not "nailed it" at being an "analog for the language center of a brain." Creating neologisms is something our language centers do.
> They could absolutely create a novel idea when predicting what a person might say, but LLMs have no way of working through the logic behind the idea or modelling reality to see if the idea might work.... but at the end of the day humans would have to sift through testing or logically invalidating each option generated.
That's not too different from a monkey at a typewriter.
Unless a LLM create a neologism 'en passant', without prompting, they did not.