Can I talk to gzip in natural language and have it produce novel output not contained within any of its source files? If not, I think your comparison is deeply flawed. LLM's are not simply compression algorithms.
Maybe if someone were to build it. You can't talk to LLMs in natural language either. They have a very precise query language. The natural language component is an additional feature bolted onto the front.
Also the output of LLMs is not novel, it's a derivative work of the training dataset. LLMs can't produce anything not present in the training set. They are operating on parameterized classifications of artwork instead of the artwork itself, which is the new technology; but they can't do anything except combine those existing Legos in new ways.
We also talk to databases in what was considered 'natural language' for that era - SQL. The LLMs are no different, which is why we have prompt engineering same as we have database engineering. Just because the database is processing the dataset and storing it in a proprietary way, approachable via query language, does not change the fact that the original data is still in there, and everything that comes out is derivative.
