some recent findings:

1/ LLMs trained on "A is B" are not able to deduce that means "B is A"

2/ LLMs are not good at self reflection

3/ Have you extensively used agents ? Because you would have quickly realized how inefficient they are and how their failure modes are impredictables.

4/ Something better is coming in the future, I have no doubt about that. But not with current autoregressive models.