Amazing post, thank you.

I really can't see how security can be solved within a probabilistic model, which is what we'd need to happen here, and that in turn effectively puts a huge limit on the scale at which we can use LLMs.

Lots of food for thought.