It has been shown that in the infinite limit of width of a neural network such as in current LLMs, any function can be approximated with arbitrarily high precision. It has also been shown that current LLMs can act as Turing machines capable of a full set of actions necessary to do complete computing behaviors. As such, it is logically impossible for such a computing system to be literally incapable of the functions necessary for AGI. Is your position then that the cost of sufficient scaling would be infeasible?