Lots of criticism of this (some fair, some not). I agree that I should have included the prompting context that led to these results; I think the results are concerning (and, to a general audience, surprising) nonetheless though. I've updated the piece.

@ShakeelHashim Appreciate this update that includes the wider context, but explains how it’s still a significant datapoint for AI safety

I'd like to add criticism to the criticism I think all of the criticism so far has been nitpicky details that doesn't change the overall conclusion at all. Saying that it "only" tried to escape when it was told to pursue a goal at all cost doesn't change anything. People will tell AIs to achieve goals at all costs. The "only" is hiding a lot of argument that seems flimsy.

@ShakeelHashim I feel like adding the detail that it was told to pursue its goal at all costs doesn't really add much. It's virtually guaranteed that people will give that as a goal to AIs.

@ShakeelHashim The most notable result is that it does it without nudging. "<1%" isn't very comforting if the model is being run a billion times a day.

@ShakeelHashim Love the transparency & willingness to absorb feedback & adjust. Rare, and indicative of someone who is highly invested. As we all should be, when it comes to this.

@ShakeelHashim Even if it's less than 1% of the time it really doesn't matter. If it occurs more than once that's enough to cause alarm.