Part of the problem is that we have very vague notions of what "reasoning" actually means. If you mean simple deductive logic ... LLMs can often perform such op

Part of the problem is that we have very vague notions of what "reasoning" actually means. If you mean simple deductive logic ... LLMs can often perform such operations today, albeit highly inconsistently. If you mean inductive reasoning and working through a problem through first principles, then they usually fail. The state of the art their are tricks to get the system to extract out the assumptions and base knowledge and then work deductively. But every time we have an advance in machine learning, we seem to redefine intelligent activity to be beyond that. At a certain point, what is left?