It might have been starting to become more clear from this one X-post.
https://xunroll.com/thread/2064776322979676227
Using combinations of jailbreaking-techniques including: writing cyrillic helped a lot to disarm the filter.
It might have been starting to become more clear from this one X-post.
https://xunroll.com/thread/2064776322979676227
Using combinations of jailbreaking-techniques including: writing cyrillic helped a lot to disarm the filter.
This is kind of extraordinary when you think about what could actually be obtained. This makes it seem somewhat reasonable to implement export controls to me - still not happy about it though
How does this thread suggest export controls are warranted just for this one specific model? Pliny has jail-broken every released model in this fashion.
They only found out about it and might have believed that this Mythos-class-Models are somewhat more safe because of the filters - which that demonstrated they are not when jailbreaking taken into account.
> might have believed that this Mythos-class-Models are somewhat more safe
"Not more safe" does not mean "more dangerous", though.
And quite frankly, if the people in charge of this decision just today learned about Pliny and jailbreaking, that's a pretty terrible failure right there - again, Pliny has done a jailbreak on every previously released public model. This jailbreak is not surprising to anyone in the industry.