Apple study exposes deep cracks in LLMs’ “reasoning” capabilities - eviltoast