Then it becomes perceived as a “loss of control,” but this shift also ( *needlessly* ) shifts the locus of control to the model! We lose control of things without them “taking control” all the time. Building a narrative out of “we weren’t in control, so what was?” invites a kind of mystical fantasy to displace accountability.
@CyberneticForests To me a simpler framing is:
* An optimization goal function created a perverse incentive.
* The model under test exhibited unexpected (but not unprecedented) behavior by escaping the test sandbox
* Hugging Face was a bystander, were harmed, and filed a police report.
* These events may support some of OpenAI’s preferred narratives, but seem to go against others. On the whole OpenAI suffered mild reputational damage.
* This was not a conspiracy, a marketing stunt, or planned in advance.