I understand some of the hype. LLMs are pretty amazing nowadays (though closedai is unethical af so don't use them).

I need to program complex cryptography code for university. Claude sonnet 3.5 solves some of the challenges instantly.

And it's not trivial stuff, but things like "how do I divide polynomials, where each coefficient of that polynomial is an element of GF(2^128)." Given the context (my source code), it adds it seamlessly, writes unit tests, and it just works. (That is important for AES-GCM, the thing TLS relies on most of the time .)

Besides that, LLMs are good at what I call moving words around. Writing cute little short stories in fictional worlds given some info material, or checking for spelling, or re-formulating a message into a very diplomatic nice message, so on.

On the other side, it's often complete BS shoehorning LLMs into things, because "AI cool word line go up".