> LLMs have no connection to ground truth. It won't matter how many or how cleverly you network them to each other, they still can't tell the difference between real users saying a product is good and good spammers saying a product is good.

I think this sounds much worse than it is. LLMs reason. They don't have an "opinion" about the truth. Meaning that the harness/prompt/tooling/data, determines the quality of the output. Using an LLM to navigate the truth is surprisingly effective right now with the right inputs and constraints.