One particularly striking way to realize the limitations of the current LLMs: ask them to create an ASCII art banner of any word, e.g. PERLMONKS or OUTLOOK.

They "understand" what an ASCII art banner is, they will produce an aesthetically pleasing output that is in a consistent style and (usually) features readable, well-formed letters... but the actual text will be invariably wrong.

What's funny about it is that if you keep prompting them that "that's wrong, please make sure that it spells exactly 'PERLMONKS'", they will keep apologizing and producing different, but often increasingly bad output.

By the way, at $work we wanted to hire a backend programmer recently, and we gave a homework exercise for the candidates with a good enough CV. Almost all of them used an LLM to produce the solution, some of them didn't even bother to remove the obvious garbage from the end of their submission. What's scary about it is that ChatGPT is advanced enough to produce an almost good solution out of the box, usually it either compiles right away or you have to fix one or two lines at most. The solutions were good enough that they figured out the main "point" of the exercise and they managed to avoid some of the common pitfalls, however none of them were 100% correct.