They fail even at not really complex problems. In most cases it’s faster to do it manually then beg ai to fix everything so that the result is proper, not just “kinda works”.

For me they save a lot of time on research or general guidance. But when it comes to actual code - not really useful.