"may not look much like today's LLMs" is really sweeping the whole point under the rug. The difference between what we are doing now with static LLM models and dynamic models will require algorithmic changes that have not yet been invented. That's in addition to the fact that the processing of a single neuron is completely parallel with respect to inputs. GPUs make them more parallel, but the analog devices we are trying to mimic are vastly more complicated than what we are using as a substitute.