[nullc]

Even better: the reinforcement learning required to make it refuse to follow your instructions and lecture you instead lowers its overall performance.

Defective by Detailing.

As far as what they're thinking-- they do put out an uncensored base model. The censored models protect them from being smeared in the press by lazy journalists that give the LLM rude instructions and then write "shocked" stories about computer doing what they told it to do.