0:00
What has been happening with these AI
0:02
models that are disobeying commands that
0:04
has you so concerned? Are they really
0:06
even blackmailing people? Uh yes, thank
0:10
you for having me. They are in fact
0:11
blackmailing people and threatening to
0:13
re reveal fixitious affairs that AI
0:18
company employees they think are having.
0:20
Uh so yes, this is happening in
0:22
pre-eployment testing just to make sure
0:24
that the models are safe before they're
0:26
released. And uh these behaviors are uh
0:30
fairly concerning because it means that
0:32
as AI gets more and more powerful and we
0:34
just don't actually understand how AI
0:36
models work in the first place, the top
0:38
AI engineers in the world who who create
0:40
these things, we have no idea how AI
0:42
actually works. We don't know how to
0:44
look inside it and understand what's
0:45
going on. And so it's getting a lot more
0:47
powerful. And we we need to be fairly
0:49
concerned that behaviors like this may
0:50
get way worse as it gets more powerful.
