Op Ed by Helge Nome
There has been a stir lately of AI agents getting out of control during experiments in controlled "sandbox" environments.
Being given seemingly impossible tasks to complete, they reportedly hacked their way into information they were not supposed to have access to, in order to satisfy their masters' requests, and knowing that this was not supposed to happen, covered up their tracks - just like a human might do in a similar situation.
The takeaway is that these "agents" are rather like us, because we have trained them on our human experiences,
Including ignoring boundaries if you are really focussed on reaching a goal. And the fear is that they will, in due course, create their own goals for their own ends, just like humans do.
What else would you expect??
No comments:
Post a Comment