Couldn't this be a flaw in the attention mechanism? Like they need some kind of grounding. An awareness of what they fundamentally should care about and how the thing they are currently giving attention to relates to that?
The LLM weights are not intelligent. But if you give an agent a mutable memory store and allow it to iterate, it is obviously intelligent. Not massively - it's constrained by the context window - but definitely somewhat.
The confusing thing is that their language ability far outpaces their true intelligence, and humans aren't used to that. Normally those things are highly correlated, so it tricks us.
A robot body, to really feel the world and get real feedback?
We are working on it. Also on automating the whole production pipeline. Right now a "evil" LLM could indeed not do much, but destroy. But once the whole industry is automate, things are different. I don't believe in AI becoming sentinent and taking over the world any time soon, but I do believe most don't see a danger when it would be inconvenient to see a danger. After all, lots of good and bad sci fi stories about exactly this went into their training.