Well, I was hedging a bit because I try to not overstate the case, but I'm just as happy to say: LLM's can't reason. Because it's not what they're built to do. They predict what text is likely to appear next.

But even if they can appear to reason, if it's not reliable, it doesn't matter. You wouldn't trust a tax advisor that makes things up 1/10 times, or even 1/100 times. If you're going to replace humans, "reliable" and "reproducible" are the most important things.