They tried to set it up like how Grok checks what Musk’s opinion is on a topic before answering, but it’s impossible with mister-word-salad
In general, one of the secrets about LLMs is that the intelligence they’re displaying is largely in the mind of the person interacting with it and the person watching and interpreting. If you ever watch one of them “on its own” trying to reason its way through a problem without help, it becomes painfully clear.
In this case, there is no intelligence in the source material for them to fall back on, so they don’t have the cheat available. It’s actually a really instructive demo.
That’s why LLMs require operator to already know the answers or can figure out when output is wrong for it to be of any practical use.
