I would start worrying if AI models can understand, reason, learn and incorporate new information on-the-fly without retraining or just stuffing information in

I would start worrying if AI models can understand, reason, learn and incorporate new information on-the-fly without retraining or just stuffing information in context windows, RAG, etc. The worry would also depend on the economics of the entire model lifecycle, as well as the current state of mechanical automation. We aren't getting that with next-token generators. I don't think we'll get there by throwing shit at the wall and seeing what sticks, either, I think we'll need a deeper understanding of the mind and what intelligence actually is before we can implement it on our own, virtually or otherwise. Similarly, we're pretty good at creating purpose-built machines, but when it comes to general/universal purpose, it's still in its infancy. The hand is still the most useful universal tool we have. It's hard to compete with the human mind + body when it comes to adapting to and manipulating the environment with purpose. There's quite literally billions of them, they create themselves and their labor is really cheap, too. There's my serious answer.