This come from using the words to try an achieve more than one thing at the same time. Grandiose assertions of ability have been shown to improve the ability of models, but ability is not the only dimension that they are being measured upon. Prioritising everything is the same thing as prioritising nothing.

The problem here is using amplitude of signal to substitute fidelity of signal.

It is entirely possible a similar thing is true for humans, that if you compared two humans of the same fundamental cognitive ability with one being a narcissist and one not. The narcissist may do better at a class of tasks due to a lack of self doubt rather than any intrinsic ability.

Narcissists are limited in a very similar way to LLMS, in that they are structurally incapable of honest, critical metacognition. Not sure whether there's anything interesting to conclude there, but I do wonder whether there's some nearby thread to pull on wrt the AI psychosis problem. That's a problem for a psychologist, which I am not.