If you have a general solution like a robot that can do human tasks, i.e. copying manuscripts with a pen, that is often highly inefficient compared to a specialized solution like a printing press but the R&D effort of applying it is minimal compared to inventing a specialized machine.

It's similar with LLMs. Compared to a bespoke piece of software for the task they are incredibly inefficient, yet they are extremely useful because they can do things where developing that piece of software would require a lot of research.