lol he believed some bullshit hype story from Anthropic about their new model so smart it will blackmail you, which would certainly be a big step up from generating text with subtly wrong and fake details.
Can’t both aspects happen? That it can be both ‘dumb’ and cunning if not enough ethics has been programmed into the model (if that’s possible?) I’m not a techie person.
Yes, although dumb and cunning are not words I would use. These models are highly capable of psychologically manipulating humans in all kinds of ways, but also make very basic mistakes from time to time. Obviously a really bad combination for society.
The creators never intended it to do anything other than make money for them. It can’t intend, it’s just good at guessing which token comes next. If you prompt it so that what follows is likely to be blackmail, that’s what you get. It’s a bit of a miracle ‘guardrails’ work as well as they do, but it’s a statistical beast, you’re never going to get all the outliers.
