LLMs have a long way to go. There are less neurones than LLMs parameters, so a neurone is more efficient than one parameter at the moment.
It also means LLMs can maybe have more space for optimisation. (A neuron is different from a parameter though, so the comparison could not stand)