Intuition and the actual test results tell us that LLMs aren’t the pinnacle when it comes to more generally usable and safer AI. A safe and more generally intelligent AI must be able to understand the consequences of its own actions, and how things behave in the real world. It’s also obvious that our biological brains are much more efficient and flexible than LLMs at learning, and at applying intelligence in a general way, that hasn’t specifically been trained before.
As Yann LeCun also emphasises, this doesn’t mean that LLMs aren’t useful. They’re great at understanding any sort of “language”, the more formally defined, the better; at abstracting large sets of digital knowledge, and at inference to generate new concrete knowledge. That’s also why LLMs are usefully good already at understanding and generating source code based on specifications. One also has to keep in mind the mechanisms behind LLMs though (stochastics; LLMs lack a truly grounded/based, intrinsic understanding of matters; it’s all just stochastically arranged symbols) and thus the limitations of LLMs.