Except LLMs are the worst of both worlds in that respect. In order to work in a robot factory, its output needs to be reliable and repeatable, ideally across as wide a range of inputs as possible. LLMs .... are very much not that. They're also only as 'skilled' as their training data, which thanks to the morally bankrupt scraping of every source the AI companies can get their grubby hands on, is of enormously variable quality - and because of the nature of LLMs, it will never be better than its training data. The average quality of its output will, in fact, be the average of its training data.

It's possible for LLMs to be creative - in the sense that it can output novel sentences - except that as you increase its 'creativity' (temperature) beyond the default that most of the chatbots out there have, the quality plummets. It still can't solve complex problems though, because even if it does have an internal model of how certain things function, it can't come close to the complexity of what humans can hold in their brains - or perhaps cannot abstract portions of their model in the same way - as evidenced by their utter failure to work through any problem that has more than five or so layers. This is a problem that sees diminishing returns with increased parameter count - the primary metric that is driving the enormous data centers being built.

LLMs are a solution looking for a problem, and aside from 'bs for people who don't want to make any decisions in their day-to-day life' and 'scam generator', there doesn't seem to be very many niches that they are actually good at filling.

Everytime someone says to a worker in a factory "figure it out" it involves creativity in some way.