That's arguably the most interesting about the discussion on LLMs. Can't they reason? If they do, reason is an emerging fonction. 2 years ago, before inference was truly added to LLMs,a Honk Kong university paper wrote that ChatGPT 3.5 reasoning was correct 64% of the time on a specific task (you asked it to make the same reasonning 100 time, 36% of the time it would be wrong for no reason). I'd like to see how modern LLMs with added inference matrices and a lot of helpers before the actual transformers perform on this test.

If a consensus arrives in 5 years and we decided that yes, LLMs can in fact reason, reasonning would be an emerging capacity, and that would be incredibly interesting.