LLMs definitely present dangers. At the same time, I think there might be more present, tangible risks such as indirect model biases in decision making and enfeeblement when models are used as substitutes to everyday life functions such as AI emotional companions and over reliance to copilots.

Lol agree mostly, some applications are hard to run with all dependencies and configurations needed, but overall yeah

Yeah, it's like what @raphaelmilliere.com said, I don't think it creates an actual, robust physics engine that runs everything by the laws of Newton, Maxwell, etc. It's more like that the model might learn a "physical intuition" of how different objects and phenomena play out during its training