The real reason they’re doing these gimmicks is far more interesting.

Whether they truly believe Claude is conscious or not is not important. It’s important to act that way and train the model that way because they believe it transfers to general intelligence and performance.

This is exactly why you see them doing other gimmicks like giving Claude the ability to end conversations and ban users.

This fits in with two phrases I have observed ChatGPT saying the past weeks:

1) "What I like to do is ..." when asked about some coding idioms

2) "I think that ...", I don't remember the context

It has never used phrasing like that in the past, I feel (correct me if I'm wrong)

Yawn. They've been not-so-subtly hinting at this at every possible opportunity for the past year.

> Suleyman argues AI models may become even harder to control if they're told to think they have feelings and rights: "Frankly, with this additional baggage, I think it would make them a catastrophic threat to human civilisation," Suleyman says.

This feels like a villain's reasoning, and I'm kind of astounded he's just saying it out loud.