WEBVTT

00:00:00.240 --> 00:00:03.997
Can you compete can can you compete with
hundred billion dollars? Um and and and

00:00:04.080 --> 00:00:07.677
my answer is basically yes because
of the talent density you know I've said

00:00:07.760 --> 00:00:09.997
this before but uh you know the the the

00:00:10.080 --> 00:00:11.837
Anthropic is actually the fastest

00:00:11.920 --> 00:00:14.717
growing software company in history at

00:00:14.800 --> 00:00:17.437
the scale that it's at. Um so we grew

00:00:17.520 --> 00:00:20.877
from zero to 100 million in 2023 100

00:00:20.960 --> 00:00:23.997
million to billion in 2024 and this year

00:00:24.080 --> 00:00:26.157
we've we've grown from one billion to I

00:00:26.240 --> 00:00:28.157
think I said this before 4.5. So that

00:00:28.240 --> 00:00:31.517
that 10x a year I mean you know every

00:00:31.600 --> 00:00:36.317
every you know every year I like suspect
that we'll grow at that scale and and

00:00:36.400 --> 00:00:40.157
every year I'm almost afraid to say it
publicly because I'm like no it couldn't

00:00:40.240 --> 00:00:43.597
it couldn't possibly happen again.
So you know I think I think the growth at

00:00:43.680 --> 00:00:46.957
that scale like kind of speaks
for itself in terms of our ability to

00:00:47.040 --> 00:00:48.877
compete with the big players. We tried

00:00:48.960 --> 00:00:51.757
to plan very well for a world of 10x

00:00:51.840 --> 00:00:54.397
growth per year. Um, in the first

00:00:54.480 --> 00:00:57.437
quarter of this year, we saw,

00:00:57.520 --> 00:01:00.717
>> if you were to annualize it, 80x growth

00:01:00.800 --> 00:01:04.877
per year in in in in in in in in in

00:01:04.960 --> 00:01:09.757
revenue and usage. And and and so
that that that is the reason we have had

00:01:09.840 --> 00:01:13.117
difficulties with compute, right? We've
planned for anything from it only grows

00:01:13.200 --> 00:01:16.317
a little to it grows 10x and yet we saw

00:01:16.400 --> 00:01:19.357
80x. Um and and and so you know as you

00:01:19.440 --> 00:01:22.557
saw today with the the the the SpaceX

00:01:22.640 --> 00:01:24.557
compute deal, we're working as quickly

00:01:24.640 --> 00:01:27.837
as possible to provide more more compute

00:01:27.920 --> 00:01:30.397
than uh than than we have in the past.

00:01:30.480 --> 00:01:33.757
What happened? I'm always reminded
of like, you know, the the there's this

00:01:33.840 --> 00:01:37.917
famous scene in the movie Interstellar
where where they go to this, you know,

00:01:38.000 --> 00:01:41.517
planet that's very close to a black
hole and and so, you know, the planet has

00:01:41.600 --> 00:01:46.157
these waves that are like, you know,
2,000 ft high and and, you know, I I was

00:01:46.240 --> 00:01:47.837
a physicist. I know I know the math of

00:01:47.920 --> 00:01:53.357
general relativity how much things can
be um sheared but actually seeing it on

00:01:53.440 --> 00:01:57.197
human scale like there's something
deeply you know it's kind of deeply

00:01:57.280 --> 00:01:59.197
strange and unsettling about seeing it

00:01:59.280 --> 00:02:00.397
actually happen.

00:02:00.480 --> 00:02:05.117
>> Yeah. So, constitutional AI was uh was

00:02:05.200 --> 00:02:07.277
kind of an algorithm, a technique that

00:02:07.360 --> 00:02:10.157
that we developed kind of a waypoint on

00:02:10.240 --> 00:02:12.557
the path of trying to figure out how to

00:02:12.640 --> 00:02:18.157
train AI systems to do what we want and
not do things that are that are harmful

00:02:18.240 --> 00:02:21.357
>> that we don't want. Yeah.
>> We don't want. So, the basic idea is

00:02:21.440 --> 00:02:24.477
that you can write down a list of

00:02:24.560 --> 00:02:28.717
principles for how an AI
system should behave.

00:02:28.800 --> 00:02:31.117
We called that a constitution.

00:02:31.200 --> 00:02:35.117
And then you have

00:02:35.200 --> 00:02:38.397
your AI,

00:02:38.480 --> 00:02:40.797
still training, it's still learning, try

00:02:40.880 --> 00:02:42.557
out different things that it could do.

00:02:42.640 --> 00:02:46.077
So you ask it a question. Um, was the

00:02:46.160 --> 00:02:48.157
third president of the US Thomas

00:02:48.240 --> 00:02:51.197
Jefferson or Abraham Lincoln or Elon

00:02:51.280 --> 00:02:55.597
Musk? And it should prefer the the the
answer that's that's correct and that's

00:02:55.680 --> 00:02:57.997
more helpful. So um you you you you

00:02:58.080 --> 00:03:03.757
allow it to try many different possible
behaviors give you a lot of different

00:03:03.840 --> 00:03:07.597
answers >> and then you
ask it or you ask uh other

00:03:07.680 --> 00:03:08.957
AI

00:03:09.040 --> 00:03:11.837
>> which of these answers is most in

00:03:11.920 --> 00:03:14.957
accordance with the constitution. The

00:03:15.040 --> 00:03:16.717
reason why I why we use the word

00:03:16.800 --> 00:03:19.677
holography for this idea is that objects

00:03:19.760 --> 00:03:24.157
like our hand dropping a a apple falling

00:03:24.240 --> 00:03:28.077
towards the earth in a gravitational
universe are fundamentally being

00:03:28.160 --> 00:03:30.637
described by some lower dimensional

00:03:30.720 --> 00:03:34.077
theory that lives on an area. Just like

00:03:34.160 --> 00:03:36.477
holograms look like three-dimensional

00:03:36.560 --> 00:03:38.797
images, but fundamentally they're

00:03:38.880 --> 00:03:41.597
two-dimensional uh structures etched

00:03:41.680 --> 00:03:45.757
into uh into glass. So the idea is that

00:03:45.840 --> 00:03:49.837
uh our approach to quantum gravity
shouldn't take this reductionist route

00:03:49.920 --> 00:03:52.557
of taking things apart in space because

00:03:52.640 --> 00:03:55.597
really space is an illusion and instead

00:03:55.680 --> 00:03:59.277
there's this uh holographic description.

00:03:59.360 --> 00:04:02.397
So if you've absorbed that you should

00:04:02.480 --> 00:04:04.077
see that I'm telling you that quantum

00:04:04.160 --> 00:04:10.520
gravity is really very very
different from uh the other