The language model's body of "knowledge" tends to fall off outside of functionality commonly covered in tutorials. Writing a "hello world" program is no problem; proposing a design for (or, worse, an addition to) a large application is hopeless.
Hard disagree. I've used GPT-4 to write full optimizers from papers that were published long after the cutoff date that use concepts that simply didn't exist in the training corpus. Trivial modifications were done after to help with memory usage and whatnot, but more often than not if I provide it the appropriate text from a paper it'll spit something out that more or less works. I have enough knowledge in the field to verify the corectness.
Most recently I used GPT-4 to implement the paper Bayesian Flow Networks, a completely new concept that I recall from the comment section on HN people said "this is way too complicated for people who don't intimately know the field" to make any use of.
I don't mind it when people don't find use with LLMs for their particular problems, but I simply don't run into the vast majority of uselessness that people find, and it really makes me wonder how people are prompting to manage to find such difficulty with them.