One of the big limitations of the transformer architecture used by all LLMs is that it does not allow learning from interactions. The learning/training phase is entirely separate from inference the actual use of the trained model. Animal (including human) intelligence learns from interactions. This provides the sanding. Current LLMs cannot do that at all.
One of the big limitations of the transformer architecture used by all LLMs is that it does not allow learning from interactions. The learning/training phase is entirely separate from inference the actual use of the trained model. Animal (including human) intelligence learns from interactions. This provides the sanding. Current LLMs cannot do that at all.
Precisely, without consequences or embodied experience, there is no way that I can see for them to overcome that. At least for the time being.