> the day we manage to run a full LLM on biological neurons, even if using conventional code under the hood

Doesn't make sense to me to use conventional code, shouldn't it be a matter of connecting the biological neurons in the same way as the simulated neurons of the NN implementing the LLM?