Correct, but it's functionally very different from how LLMs are implemented and deployed today. What you're highlighting is being experimented with and ties into ideas like scratch pads, world models, RAG, and progressive fine-tuning (if you're googling).
It's a bit like saying your computer has everything it needs to manipulate photos but doesn't yet have Photoshop installed.
No, I’m not talking about giving LLMs chain of thought prompts or augmenting them with scratchpads - I’m literally saying that in a multilayer neural network you don’t know what concepts activations on the inner layers mean. The result of ‘where I want this conversation to be in 100 tokens time’ could absolutely be in there somewhere.
It's a bit like saying your computer has everything it needs to manipulate photos but doesn't yet have Photoshop installed.