WEBVTT

00:00:00.000 --> 00:00:03.237
It's interesting cuz people
have been talking about

00:00:03.320 --> 00:00:08.397
the potential downsides or serious
risks of AI for some time.

00:00:08.480 --> 00:00:11.957
And I think a lot of people have been
saying why is this the moment where it

00:00:12.040 --> 00:00:16.957
sort of escaped the bubble and is now
all of a sudden an international topic

00:00:17.040 --> 00:00:21.597
of such uh intensity.
And I think a few things have

00:00:21.680 --> 00:00:25.877
happened. One, the models
have gotten so good and

00:00:25.960 --> 00:00:29.237
people for a long time
looked at the models.

00:00:29.320 --> 00:00:32.077
Maybe they looked at the rate
of progress, but I don't think people

00:00:32.160 --> 00:00:35.677
really felt that the models were going
to come this far, this fast. I think the

00:00:35.760 --> 00:00:39.437
biggest thing that has changed
is this thing that our industry has been

00:00:39.520 --> 00:00:43.437
talking about for a while, which is no,
no, this technology, this approach is

00:00:43.520 --> 00:00:47.677
going to go very far. These models
are in some ways going to become more

00:00:47.760 --> 00:00:51.037
capable, smarter than people.
Um that finally

00:00:51.120 --> 00:00:54.637
happened. And so I think that was a big
change to why now when people say, you

00:00:54.720 --> 00:00:57.957
know, my probability of doom
is X or Y or whatever,

00:00:58.040 --> 00:01:04.120
it hits in a different way.
Uh I think you also see um

00:01:04.360 --> 00:01:06.677
a real fear that some of these companies

00:01:06.760 --> 00:01:08.517
or some companies developing AI could

00:01:08.600 --> 00:01:11.317
get too much power and be able to sort

00:01:11.400 --> 00:01:14.637
of exert undue influence on the economy,

00:01:14.720 --> 00:01:18.757
push a world view out on people. Uh and I
think the world is right to be afraid

00:01:18.840 --> 00:01:21.797
of this. So we have these
two big challenges. We have

00:01:21.880 --> 00:01:24.917
the potential of a loss of control
accident or some other serious thing

00:01:25.000 --> 00:01:28.837
that could go wrong. Um we have
the potential of way too much power

00:01:28.920 --> 00:01:33.797
concentration and people fear that those
developing AI could exert their world

00:01:33.880 --> 00:01:36.717
view. Um and there
is a sort of narrow path we

00:01:36.800 --> 00:01:38.917
have to navigate

00:01:39.000 --> 00:01:42.437
with pragmatism, with

00:01:42.520 --> 00:01:47.077
you know, steadiness, with the ability
for the the world to like trust that we

00:01:47.160 --> 00:01:52.477
will make dependable decisions,
consistent decisions through this. And I

00:01:52.560 --> 00:01:55.477
think people are saying, well, you know,
this now seems pretty important. Like

00:01:55.560 --> 00:01:57.917
let's let's get it right. These models

00:01:58.000 --> 00:02:00.997
have gotten so good so fast that we need

00:02:01.080 --> 00:02:05.397
to treat the alignment and monitoring
and security with a new level of rigor

00:02:05.480 --> 00:02:07.957
and we have to be willing to pace our

00:02:08.040 --> 00:02:10.037
development of capabilities such that

00:02:10.120 --> 00:02:15.157
capabilities always such
that alignment safety and

00:02:15.240 --> 00:02:18.637
monitoring are always
ahead of capabilities. And

00:02:18.720 --> 00:02:21.917
the hugging face incident triggered
a real reset not just in our own company

00:02:22.000 --> 00:02:25.357
but I think the whole industry. What I
really care about is that we have got a

00:02:25.440 --> 00:02:29.677
great culture of accident reporting
and learning. I think if we look at what

00:02:29.760 --> 00:02:34.317
happens with the FAA and the NTSB
and how safe airplanes have gotten.

00:02:34.400 --> 00:02:37.677
As I understand it from talking
to people who have really studied

00:02:37.760 --> 00:02:41.317
you know new kinds of technology on the
order of magnitude of airplanes. One of

00:02:41.400 --> 00:02:45.117
the things that has gone so well there is

00:02:45.200 --> 00:02:46.917
is the culture of accident reporting.

00:02:47.000 --> 00:02:48.597
You know accidents are going to happen.

00:02:48.680 --> 00:02:52.277
We're going to learn as much
as we can and react from each one.

00:02:52.360 --> 00:02:55.997
And if you know there's other kinds of

00:02:56.080 --> 00:03:00.317
technology where I don't think it's gone
I don't think society has had the same

00:03:00.400 --> 00:03:04.757
response and I don't think we have
improved as much. I think this is one of

00:03:04.840 --> 00:03:09.637
the most profound changes
to how we use technology

00:03:09.720 --> 00:03:12.797
that I've ever seen.
I mean sci-fi movies as you mentioned

00:03:12.880 --> 00:03:16.477
have been hinting at this or playing
around with this for a long time but

00:03:16.560 --> 00:03:19.397
this is one of those moments where it
feels like you're in the movie. Like the

00:03:19.480 --> 00:03:23.197
sci-fi thing came true.
You're talking to your personal super

00:03:23.280 --> 00:03:25.557
assistant and you are working 10 times

00:03:25.640 --> 00:03:30.480
faster or better or more creatively
than you could have before.