I wonder whether Bing has been tuned via RLHF to have this personality (over the boring one of ChatGPT); perhaps Microsoft felt it would drive engagement and hype.

Alternately - maybe this is the result of less RLHF. Maybe all large models will behave like this, and only by putting in extremely rigid guard rails and curtailing the output of the model can you prevent it from simulating/presenting as such deranged agents.

Another random thought: I suppose it's only a matter of time before somebody creates a GET endpoint that allows Bing to 'fetch' content and write data somewhere at the same time, allowing it to have a persistent memory, or something.