If found that stack languages are a massive weak spot for LLMs. It requires accurate tracking of what is on the stack and for some reason that doesn't really work. So even the best models end up making the same basic mistakes over and over again.