WEBVTT

00:00:00.720 --> 00:00:04.517
Hello again. If I keep
doing this, I understand that

00:00:04.600 --> 00:00:09.837
there's vacancy at 60 Minutes. Maybe I
can If this investing thing doesn't work

00:00:09.920 --> 00:00:14.317
out, maybe I can do that.
Another full disclosure,

00:00:14.400 --> 00:00:16.677
Sam is a friend. I'm also on the board

00:00:16.760 --> 00:00:19.437
of OpenAI, so

00:00:19.520 --> 00:00:23.797
I promise you I will not only ask him
softball questions. I will channel my

00:00:23.880 --> 00:00:27.157
inner Brett Bear and and uh

00:00:27.240 --> 00:00:28.957
and seek Rupert.

00:00:29.040 --> 00:00:31.517
So, Sam, as I said,

00:00:31.600 --> 00:00:35.397
CEO of OpenAI, one of the founders.

00:00:35.480 --> 00:00:39.157
He previously served as the
president of Y Combinator,

00:00:39.240 --> 00:00:43.997
where he helped scale some of the world's
most influential startup accelerators,

00:00:44.080 --> 00:00:48.037
and advised companies that will
shape entire industries.

00:00:48.120 --> 00:00:49.877
At OpenAI, Sam has obviously been at the

00:00:49.960 --> 00:00:51.797
forefront of advancing artificial

00:00:51.880 --> 00:00:55.157
intelligence to the benefit of humanity,

00:00:55.240 --> 00:01:00.077
helping to bring cutting-edge AI tools
to the hands of millions of people. How

00:01:00.160 --> 00:01:01.917
many people? Eight, 900 million people?

00:01:02.000 --> 00:01:05.757
>> More, yeah.
>> And businesses around the world.

00:01:05.840 --> 00:01:10.437
So, Sam, let me start with what I think
is the question that is on everybody's

00:01:10.520 --> 00:01:14.517
mind, okay? Which is

00:01:14.600 --> 00:01:16.917
where are we today

00:01:17.000 --> 00:01:19.317
in the world of AI?

00:01:19.400 --> 00:01:21.997
>> I I think at some point in the last few

00:01:22.080 --> 00:01:24.437
months,

00:01:24.520 --> 00:01:27.757
we really have crossed a threshold into

00:01:27.840 --> 00:01:30.477
major economic utility of these models.

00:01:30.560 --> 00:01:33.077
It may have happened a little bit earlier,
but there was such an overhang

00:01:33.160 --> 00:01:37.117
before we figured out how to use these,
and we had to not only continue to make

00:01:37.200 --> 00:01:40.997
the models get smarter, but figure out
sort of the plumbing to make them easy

00:01:41.080 --> 00:01:43.557
to use. Where we're we're now in a world

00:01:43.640 --> 00:01:48.877
where the models are
astounding people with the

00:01:48.960 --> 00:01:52.677
work they can do.
Uh and I think this has been most

00:01:52.760 --> 00:01:54.797
noticeable in coding, yep, but it's also

00:01:54.880 --> 00:01:56.877
happening in science. It's happening in

00:01:56.960 --> 00:02:00.717
many fields of knowledge work. Uh

00:02:00.800 --> 00:02:04.037
at sort of with this orienting speed
where people are saying like, "Man,

00:02:04.120 --> 00:02:09.237
these things that I thought were still
years away are happening now and I have

00:02:09.320 --> 00:02:12.557
my job shifted from doing

00:02:12.640 --> 00:02:16.797
uh you know, direct technical work
or uh you know, legal work to managing a

00:02:16.880 --> 00:02:19.077
team of agents doing this work."

00:02:19.160 --> 00:02:22.757
This is going to go much further. Uh I
think we're at a very steep part of the

00:02:22.840 --> 00:02:23.797
curve

00:02:23.880 --> 00:02:27.557
and right now maybe you can trust say a

00:02:27.640 --> 00:02:32.077
AI software engineer to do a multi-hour
task. Very soon it'll be a multi-day

00:02:32.160 --> 00:02:34.397
task and then a multi-week task.

00:02:34.480 --> 00:02:37.317
And not long after that, I think the

00:02:37.400 --> 00:02:39.637
paradigm will shift again and it'll feel

00:02:39.720 --> 00:02:44.517
like these AI systems are just connected
to your life, to your company, whatever

00:02:44.600 --> 00:02:48.237
proactively thinking,
working all the time. Uh

00:02:48.320 --> 00:02:52.157
and having full context on whatever
they need to know and just sort of doing

00:02:52.240 --> 00:02:55.237
stuff like you would trust
a senior employee to do.

00:02:55.320 --> 00:02:59.197
>> And and you think
that um companies have

00:02:59.280 --> 00:03:01.357
real have a real understanding of how

00:03:01.440 --> 00:03:03.677
these systems can help them and

00:03:03.760 --> 00:03:07.157
reimagine how they do their businesses?

00:03:07.240 --> 00:03:11.157
>> Some do, some don't. Uh certainly
the new generation of startups thinks

00:03:11.240 --> 00:03:14.277
differently than any
generation of startups before.

00:03:14.360 --> 00:03:17.877
Uh it used to be that
when we would talk to

00:03:17.960 --> 00:03:21.917
startups, they would talk about
how many employees they needed.

00:03:22.000 --> 00:03:25.997
Uh now they generally don't want to hire
a lot. They think that'll slow them down

00:03:26.080 --> 00:03:28.677
and they're all focused on how
much compute they can get.

00:03:28.760 --> 00:03:31.117
You know, can I reserve this much
capacity? Can I do a cloud deal for

00:03:31.200 --> 00:03:33.877
that? Uh can I get this many tokens?

00:03:33.960 --> 00:03:38.317
And I think that is that that
is a mental shift that bigger

00:03:38.400 --> 00:03:40.717
companies are going through more slowly

00:03:40.800 --> 00:03:44.277
but some are starting to do that. Uh one

00:03:44.360 --> 00:03:49.597
place I think you can see this happening
is engineering orgs and product orgs

00:03:49.680 --> 00:03:54.357
talking about they're doubling, tripling
what they're planning to ship this year.

00:03:54.440 --> 00:03:56.077
And that has not happened before.

00:03:56.160 --> 00:03:57.117
>> Yeah.

00:03:57.200 --> 00:04:00.597
And you've been quite vocal in saying

00:04:00.680 --> 00:04:05.037
that um artificial
general intelligence will

00:04:05.120 --> 00:04:06.997
come sooner rather than later.

00:04:07.080 --> 00:04:10.677
Want to share your views on how
close we are and how soon it will

00:04:10.760 --> 00:04:12.317
come?

00:04:12.400 --> 00:04:15.237
>> At at this point,
I think the definition

00:04:15.320 --> 00:04:19.477
of AGI really matters. Some people would
say we already got there. Some people

00:04:19.560 --> 00:04:22.517
say it's very close. Some people say
we're kind of you know, it's maybe still

00:04:22.600 --> 00:04:28.517
a year away, but in any case,
that word has ceased to have much meaning.

00:04:28.600 --> 00:04:31.917
>> Mhm. >> Uh there are
maybe two thresholds that we

00:04:32.000 --> 00:04:33.597
could talk about that are interesting.

00:04:33.680 --> 00:04:35.677
>> Okay. >> Number one,
when is there going to be

00:04:35.760 --> 00:04:40.077
more of the world's cognitive capacity
inside of data centers than outside of

00:04:40.160 --> 00:04:45.677
them? And that to me feels
like maybe it could happen.

00:04:45.760 --> 00:04:48.317
Huge error bars. I could be totally
wrong, but maybe that could happen by

00:04:48.400 --> 00:04:51.237
like late 2028. >> Mhm.

00:04:51.320 --> 00:04:55.437
>> And that's an extraordinary
shift in the world.

00:04:55.520 --> 00:04:56.677
Um

00:04:56.760 --> 00:05:00.477
the other one is when can

00:05:00.560 --> 00:05:03.157
a CEO of major company, a president of a

00:05:03.240 --> 00:05:05.437
major country, a Nobel Prize winning

00:05:05.520 --> 00:05:07.637
scientist, when can they not do their

00:05:07.720 --> 00:05:10.957
job without making heavy use of AI? This

00:05:11.040 --> 00:05:14.677
doesn't mean that there will be
an AI CEO or an AI president, uh

00:05:14.760 --> 00:05:18.317
but it does mean that
the the role of let's

00:05:18.400 --> 00:05:23.277
say a human CEO, when I think about my
job, it's really quite different. You you

00:05:23.360 --> 00:05:25.277
still do need a person to

00:05:25.360 --> 00:05:27.277
stand behind decisions and kind of

00:05:27.360 --> 00:05:30.677
exercise human judgment and

00:05:30.760 --> 00:05:33.037
all of the understanding

00:05:33.120 --> 00:05:35.597
that we expect out of someone running an

00:05:35.680 --> 00:05:39.437
an important organization to do, but

00:05:39.520 --> 00:05:43.037
the actual parts of my role that I will

00:05:43.120 --> 00:05:47.477
increasingly have to rely on an AI to do
because no human can. You know, no human

00:05:47.560 --> 00:05:51.477
can talk to no human CEO can talk
to every employee at a company, every

00:05:51.560 --> 00:05:53.557
customer, be in every meeting, be an

00:05:53.640 --> 00:05:56.357
expert in every field. And so,

00:05:56.440 --> 00:05:59.317
more and more I think of these jobs will

00:05:59.400 --> 00:06:03.637
be supervising a bunch of AI, um providing

00:06:03.720 --> 00:06:06.197
oversight, deciding how to like trust

00:06:06.280 --> 00:06:08.037
the the outputs, how to provide

00:06:08.120 --> 00:06:12.317
guidance. And that threshold of when you

00:06:12.400 --> 00:06:16.637
really wouldn't want to be doing your job,

00:06:16.720 --> 00:06:20.877
running a large organization
without heavily heavy reliance on AI,

00:06:20.960 --> 00:06:23.357
I think that's another sort
of interesting threshold. That may take a

00:06:23.440 --> 00:06:26.277
little bit longer,
but probably not a lot longer.

00:06:26.360 --> 00:06:29.077
>> And and as you do
your job, how much are

00:06:29.160 --> 00:06:32.357
you finding yourself relying on

00:06:32.440 --> 00:06:36.117
some of the agents and some of the
artificial intelligence that we're

00:06:36.200 --> 00:06:39.317
developing at OpenAI? >> Uh

00:06:39.400 --> 00:06:42.117
it's ramping incredibly

00:06:42.200 --> 00:06:45.757
quickly. Um if I have like a new idea

00:06:45.840 --> 00:06:51.520
for a business model,
a strategy shift, a uh

00:06:51.680 --> 00:06:55.317
a product offering we should do,
the very first thing I do before I even

00:06:55.400 --> 00:06:57.517
bounce it off somebody else is to ask

00:06:57.600 --> 00:07:02.517
our tools. And as they get more context,
and I think this really is like

00:07:02.600 --> 00:07:06.117
the next big thing to happen is they
can get close to full context of our

00:07:06.200 --> 00:07:10.717
company, access to all of our
internal docs, communication, code,

00:07:10.800 --> 00:07:16.117
customer data, everything, the quality of
the answers, thought, whatever you want to

00:07:16.200 --> 00:07:17.237
call it, gets better and better.

00:07:17.320 --> 00:07:21.437
>> Right. Okay. So, let's
let's shift um a little bit. 2

00:07:21.520 --> 00:07:26.477
weeks ago, um you announced
a $110 billion funding

00:07:26.560 --> 00:07:31.877
round. I asked ChatGPT,
how does that compare compare to any

00:07:31.960 --> 00:07:34.437
other fundraising that's been
done in the public markets?

00:07:34.520 --> 00:07:37.917
>> I actually don't know.
>> Uh four times as large, okay? The

00:07:38.000 --> 00:07:42.357
largest public offering ever done
was roughly 25 billion that Aramco did

00:07:42.440 --> 00:07:45.477
several years ago.
And the public markets are supposed to

00:07:45.560 --> 00:07:50.317
be the broadest and deepest
sources of capital, okay?

00:07:50.400 --> 00:07:53.677
Three strategic partners, Amazon,

00:07:53.760 --> 00:07:56.997
Nvidia, and oh and SoftBank.

00:07:57.080 --> 00:07:59.477
Tell us a little bit about this.

00:07:59.560 --> 00:08:02.397
How is this an inflection
point of the company?

00:08:02.480 --> 00:08:05.237
And one of the questions somebody asked
me is, what are we spending all that

00:08:05.320 --> 00:08:08.797
money on? The

00:08:08.880 --> 00:08:12.357
>> There is many hard parts of this
business, but one of the hardest ones is

00:08:12.440 --> 00:08:15.157
the infrastructure is so expensive. You

00:08:15.240 --> 00:08:20.360
need so much of it, and you
have to commit so far in advance.

00:08:20.960 --> 00:08:25.757
I have never seen any other
industry quite like this. I

00:08:25.840 --> 00:08:29.637
mean, there have been clearly many
capital intensive industries throughout

00:08:29.720 --> 00:08:31.757
history, but

00:08:31.840 --> 00:08:34.317
as I look at the what's to come in front

00:08:34.400 --> 00:08:36.717
of us in the future in future years, if

00:08:36.800 --> 00:08:39.597
the ramp stays as steep as it looks like

00:08:39.680 --> 00:08:44.197
it is right now, the demand is growing
as fast as it's growing, you have to do

00:08:44.280 --> 00:08:47.277
some pretty unusual things. Uh

00:08:47.360 --> 00:08:51.237
OpenAI does a lot
of things that look weird.

00:08:51.320 --> 00:08:53.757
We spend a ton of money on

00:08:53.840 --> 00:08:56.837
infrastructure in advance of revenue. We

00:08:56.920 --> 00:08:59.517
do new business models like ads that

00:08:59.600 --> 00:09:04.477
seem like, you know, maybe not the most
profitable thing we could do. Um

00:09:04.560 --> 00:09:08.277
a long list of other things, but we have
this fundamental belief in abundance of

00:09:08.360 --> 00:09:12.357
intelligence. And that one of the most
important things in the future is that

00:09:12.440 --> 00:09:16.997
we make intelligence, you know,
to borrow an old phrase from

00:09:17.080 --> 00:09:19.757
the energy industry that didn't quite
work, too cheap to meter. We want to

00:09:19.840 --> 00:09:21.997
flood the world with intelligence.
We want people to just use it for

00:09:22.080 --> 00:09:24.997
everything. We want this to just be
something that the future generation

00:09:25.080 --> 00:09:29.637
doesn't think about. They expect
everywhere, and everybody has access to

00:09:29.720 --> 00:09:33.157
like geniuses, as many as they need in any

00:09:33.240 --> 00:09:37.240
area that they need. And

00:09:37.640 --> 00:09:41.917
this principle, which is one of our kind
of like top guiding principles, does

00:09:42.000 --> 00:09:45.437
lead to a lot of behavior that would look

00:09:45.520 --> 00:09:49.197
less natural for other companies.
And one of those is

00:09:49.280 --> 00:09:52.477
we really want to get out of this world
that we have been in that we still think

00:09:52.560 --> 00:09:55.637
we're on a trajectory to stay on without
changing what we do of always being

00:09:55.720 --> 00:09:58.237
capacity constrained. >> Right.

00:09:58.320 --> 00:10:01.477
And capacity constrained
you mean compute? Yeah.

00:10:01.560 --> 00:10:06.917
And I've often heard you say
a lot that compute is revenues.

00:10:07.000 --> 00:10:11.317
You want to talk a little bit about
how you think how you think about that?

00:10:11.400 --> 00:10:14.797
>> Fundamentally, our business and I
think the business of every other model

00:10:14.880 --> 00:10:19.237
provider is going to look
like selling tokens.

00:10:19.320 --> 00:10:23.437
You know, they may come from bigger
or smaller models which makes them more or

00:10:23.520 --> 00:10:26.917
less expensive. They may use more
or less reasoning which also makes them

00:10:27.000 --> 00:10:30.037
more or less expensive. They may be
running all the time in the background

00:10:30.120 --> 00:10:33.437
trying to help you out. Uh they may run
only when you need them if you want to

00:10:33.520 --> 00:10:38.637
pay less. They may work super hard.
You know, spend tens of

00:10:38.720 --> 00:10:41.877
millions, hundreds of millions, someday
billions of dollars on a single problem.

00:10:41.960 --> 00:10:44.437
>> Right. >> That's really valuable. But

00:10:44.520 --> 00:10:46.997
we

00:10:47.080 --> 00:10:49.437
see a future where

00:10:49.520 --> 00:10:51.597
intelligence is a utility like

00:10:51.680 --> 00:10:56.517
electricity or water
and people buy it from us

00:10:56.600 --> 00:11:00.957
um on a meter and use it
for whatever they want to use it for.

00:11:01.040 --> 00:11:04.997
The demand that we see for that seems
like it's going to continue to just go

00:11:05.080 --> 00:11:08.797
like this. And

00:11:08.880 --> 00:11:14.197
if we don't have enough, we either can't
sell it or the price gets really high

00:11:14.280 --> 00:11:19.397
and it you know, it kind of goes to rich
people or society makes a bunch of sort

00:11:19.480 --> 00:11:22.557
of central planning decisions that I
think almost always go badly about, you

00:11:22.640 --> 00:11:25.997
know, we're going to use our limited
compute supply for this and not that.

00:11:26.080 --> 00:11:30.397
So, the best thing to me throughout all
the history of capitalism, innovation,

00:11:30.480 --> 00:11:32.637
whatever you want
is to just flood the market.

00:11:32.720 --> 00:11:35.757
>> Yeah. And obviously a critical part of

00:11:35.840 --> 00:11:37.637
addressing that

00:11:37.720 --> 00:11:41.477
compute demand is Stargate.

00:11:41.560 --> 00:11:45.757
And this is an infrastructure conference
and you know, you announced that several

00:11:45.840 --> 00:11:49.837
months ago, or almost a year
ago, I guess it was, yeah.

00:11:49.920 --> 00:11:53.997
How's it going in the US?
And then how is it going, because I

00:11:54.080 --> 00:11:59.680
also started in in Abu Dhabi.
>> Yeah. Yeah. Um

00:12:00.280 --> 00:12:03.037
Honestly, there's many
cool parts of the job. One

00:12:03.120 --> 00:12:06.317
of the coolest is getting to go
visit these mega data centers under

00:12:06.400 --> 00:12:07.837
construction and operation. >> Yeah.

00:12:07.920 --> 00:12:11.317
>> Just the scale is, you know, of like
these gigawatt campuses. It's it's it's

00:12:11.400 --> 00:12:15.237
really hard to explain and you see these
photos and it's like, okay, it does look

00:12:15.320 --> 00:12:18.157
big. And then you go there and you're
walking through it and building to

00:12:18.240 --> 00:12:21.597
building and it's just, you know,
10,000 people there, all these different

00:12:21.680 --> 00:12:24.197
skilled trades doing all
these different things and

00:12:24.280 --> 00:12:28.797
you know, it looks like a spaceship
inside. It's really quite incredible. Um

00:12:28.880 --> 00:12:32.557
we are training right now
on the first site in Abilene,

00:12:32.640 --> 00:12:33.917
what I think will be the best model in

00:12:34.000 --> 00:12:38.797
the world, hopefully by a lot. And it is

00:12:38.880 --> 00:12:42.597
it's so amazing to have gone from,
you know, many visits while it was under

00:12:42.680 --> 00:12:46.997
construction, and just really
internalizing the scale and incredible

00:12:47.080 --> 00:12:50.477
complexity to the day
that one researcher at OpenAI

00:12:50.560 --> 00:12:54.157
types in like one command
and it pushes enter,

00:12:54.240 --> 00:12:58.557
unbelievable number of GPUs spin up
and start doing this one huge computation

00:12:58.640 --> 00:13:02.517
all together. >> Yeah. It's very And and

00:13:02.600 --> 00:13:06.757
what have been the most
pleasant surprises and what have

00:13:06.840 --> 00:13:11.237
been the hardest things about getting
it up and running and scaling it going

00:13:11.320 --> 00:13:16.157
forward? >> Uh I mean,
there have been all the expected

00:13:16.240 --> 00:13:19.517
challenges and then the unknown unknowns.
Like there was a crazy weather

00:13:19.600 --> 00:13:22.677
event in Abilene that we hadn't,
you know, was outside

00:13:22.760 --> 00:13:25.277
what we had planned for and that brought
things down for a little while.

00:13:25.360 --> 00:13:27.837
>> Mhm. >> Um there's all of supply chain

00:13:27.920 --> 00:13:32.437
challenges. Anything at this scale is just
like so much stuff goes wrong. Lots

00:13:32.520 --> 00:13:39.080
goes right, but the just try trying
to build something around the clock

00:13:39.280 --> 00:13:42.397
with such complexity, there's
all the stuff that goes wrong.

00:13:42.480 --> 00:13:46.757
One of the biggest surprises
on the upside is how

00:13:46.840 --> 00:13:50.877
many different organizations had to come
together to do this in an incredibly

00:13:50.960 --> 00:13:53.197
short period of time. Uh, and how much

00:13:53.280 --> 00:13:54.997
we all like ended up working together as

00:13:55.080 --> 00:13:58.517
one team uh, under a lot of pressure.

00:13:58.600 --> 00:14:03.117
>> And of course, the power demand part
of it is the one that a lot of people are

00:14:03.200 --> 00:14:06.197
focused on. Um,

00:14:06.280 --> 00:14:10.357
are you optimistic that we
will solve that challenge

00:14:10.440 --> 00:14:15.077
first in the US and then we can
talk about in in other places?

00:14:15.160 --> 00:14:19.077
>> Uh, in the long term, I am.

00:14:19.160 --> 00:14:23.517
>> Okay. >> Uh, I have no
doubt we will figure out how

00:14:23.600 --> 00:14:27.797
to build huge amounts of power generation.
AI will help, of course.

00:14:27.880 --> 00:14:31.437
Uh, but I think the portfolio
the world has in front of it

00:14:31.520 --> 00:14:35.197
gas, solar, nuclear fission,
nuclear fusion, more

00:14:35.280 --> 00:14:39.237
like we're I feel good
about what what we will be

00:14:39.320 --> 00:14:41.397
capable of and what we will eventually do.

00:14:41.480 --> 00:14:42.157
>> Mhm.

00:14:42.240 --> 00:14:43.957
>> Um,

00:14:44.040 --> 00:14:48.637
given the demand growth we are seeing,

00:14:48.720 --> 00:14:52.997
I am sort of hoping for a miracle in terms
of figuring out how we can get way

00:14:53.080 --> 00:14:54.957
more efficient per watt >> Uh, okay.

00:14:55.040 --> 00:14:59.877
>> models to give us time to build
out all of this infrastructure. Now,

00:14:59.960 --> 00:15:03.477
the track record there
has been incredible. Um,

00:15:03.560 --> 00:15:07.317
people cite whatever amazing statistic
they like about how much more efficient

00:15:07.400 --> 00:15:10.877
our models have gotten over our industries
models have gotten over time,

00:15:10.960 --> 00:15:14.477
but one that I think is incredible,

00:15:14.560 --> 00:15:18.237
our first reasoning model
was called 01, came out

00:15:18.320 --> 00:15:22.277
like 16 months ago.

00:15:22.360 --> 00:15:28.800
Uh, and latest model where we're
now integrated reasoning is 5.4.

00:15:29.040 --> 00:15:31.357
To get the same answer

00:15:31.440 --> 00:15:34.437
to a hard problem from that first model

00:15:34.520 --> 00:15:40.800
to 5.4 has been a reduction
in cost of about 1,000 x.

00:15:41.320 --> 00:15:42.997
Maybe I was a little bit wrong
on the timeline. Maybe it was a little bit

00:15:43.080 --> 00:15:45.197
longer. But in any case,
since 01 till now.

00:15:45.280 --> 00:15:50.317
>> Yeah. >> 1,000 x. Uh,
that is unbelievable in a relatively

00:15:50.400 --> 00:15:54.677
short period of time. And

00:15:54.760 --> 00:15:58.397
the thing that that the two
things that I think that points to

00:15:58.480 --> 00:16:02.277
one, we are still so early and in this

00:16:02.360 --> 00:16:05.597
paradigm and we're still so we
have so much more to gain about our

00:16:05.680 --> 00:16:08.957
understanding of how to develop these
models and train them and run them

00:16:09.040 --> 00:16:13.597
efficiently >> Mhm. >> that we have

00:16:13.680 --> 00:16:17.757
were then and still are doing things
in dumb ways and will get better and

00:16:17.840 --> 00:16:21.437
better. And two, is that

00:16:21.520 --> 00:16:25.477
human ingenuity and the ability to operate

00:16:25.560 --> 00:16:29.877
in constraints and to find ways to solve
problems almost always surprise you on the

00:16:29.960 --> 00:16:32.797
upside. So, it's not just that the models
have gotten better, it's that we

00:16:32.880 --> 00:16:36.277
have figured out, you know, kernel
engineers came to help figure out how to

00:16:36.360 --> 00:16:40.237
write more efficient kernels and like
power engineers, the people that design

00:16:40.320 --> 00:16:43.077
data centers found more efficient ways
to do that. So, people are answering the

00:16:43.160 --> 00:16:45.237
call well beyond just
the model side >> Right.

00:16:45.320 --> 00:16:46.277
>> to make this more efficient.

00:16:46.360 --> 00:16:48.837
>> Right. And then of course, um,

00:16:48.920 --> 00:16:53.717
obviously OpenAI is a big customer
of the Nvidias and AMDs of the world, but

00:16:53.800 --> 00:16:57.117
um, you know, I think we signed
a big contract with Amazon.

00:16:57.200 --> 00:16:58.357
But we're building our own chip.

00:16:58.440 --> 00:17:01.517
>> Yeah. >> What's what's
the thinking behind that?

00:17:01.600 --> 00:17:03.717
>> Uh, so the chip we're
doing is inference only.

00:17:03.800 --> 00:17:08.037
>> Okay. >> And the thinking behind
it is that on this, you know, come

00:17:08.120 --> 00:17:10.957
rise up to solve problems in front of us,

00:17:11.040 --> 00:17:15.957
um, we think that a specialized chip to be

00:17:16.040 --> 00:17:19.717
not necessarily the fastest inference
chip, but the cheapest inference chip,

00:17:19.800 --> 00:17:23.597
the most efficient per watt given
the constraints we see in front of us, is

00:17:23.680 --> 00:17:27.197
going to be important for all the agent
demand we see in the future. So, it's a

00:17:27.280 --> 00:17:31.277
it's an opinionated bet, it's a limited
chip, but the thing that it does,

00:17:31.360 --> 00:17:34.157
in a world where we're energy constrained,
I think it'll be very important.

00:17:34.240 --> 00:17:38.197
>> Maybe you can explain because I am
not sure everybody in the audience is

00:17:38.280 --> 00:17:43.197
AI proficient with the difference between
an inference chip and a training chip.

00:17:43.280 --> 00:17:44.557
>> Sorry, I should have done that first.

00:17:44.640 --> 00:17:45.517
So,

00:17:45.600 --> 00:17:49.957
there there are two main phases to uh

00:17:50.040 --> 00:17:53.477
AI workloads. Now, they're going
to blend together eventually and it'll all

00:17:53.560 --> 00:17:55.837
be one continuous thing. But right now,

00:17:55.920 --> 00:17:58.757
first we train a model,

00:17:58.840 --> 00:18:00.717
you know, a gigantic number of GPUs

00:18:00.800 --> 00:18:03.157
crunch for weeks or months on a bunch of

00:18:03.240 --> 00:18:05.837
data and you can think of that like

00:18:05.920 --> 00:18:08.597
maybe it takes you 22 years in life to

00:18:08.680 --> 00:18:13.237
get your education and you learn a bunch
of stuff starting when you're a baby and

00:18:13.320 --> 00:18:16.317
you drop things and see they fall and
then eventually like in college physics

00:18:16.400 --> 00:18:20.157
understanding like at a very
detailed level what's going on there.

00:18:20.240 --> 00:18:24.157
Um And then after that, if you ask a model

00:18:24.240 --> 00:18:26.637
to solve a physics question, it that's

00:18:26.720 --> 00:18:31.197
called inference uh and that's quite
efficient. But this is true for people

00:18:31.280 --> 00:18:34.397
as well and when people talk
about all these AI models are so

00:18:34.480 --> 00:18:37.557
efficient, they're usually comparing the

00:18:37.640 --> 00:18:40.597
22 years that a human takes to train to

00:18:40.680 --> 00:18:45.517
the 1 second that an adult takes to solve
a physics problem. If you compare the

00:18:45.600 --> 00:18:48.717
you know, the model solving the physics
problem to the human solving physics

00:18:48.800 --> 00:18:52.917
problem, actually the models already
probably are more energy efficient. Um

00:18:53.000 --> 00:18:57.637
but training is this hugely massive
amount of processing power that then

00:18:57.720 --> 00:19:01.037
produces really just like
a file of numbers that

00:19:01.120 --> 00:19:03.877
you can then pose a question
to and get a response.

00:19:03.960 --> 00:19:07.757
>> And are you optimistic about
the progress we're making on the chip?

00:19:07.840 --> 00:19:12.037
>> Yeah, um we should have
the first chips deployed at scale

00:19:12.120 --> 00:19:15.077
uh by the end of this year. We should
have the first chips back just in a few

00:19:15.160 --> 00:19:18.237
months now. Um and looks
like it'll be really good.

00:19:18.320 --> 00:19:21.877
>> Fantastic. So,
you announced a new partnership this

00:19:21.960 --> 00:19:25.957
morning, actually, with the North
American Building Trades Unions to

00:19:26.040 --> 00:19:30.357
expand training pathways for skilled
construction, which has been a subject

00:19:30.440 --> 00:19:34.677
of my many of the discussions today. Can
you tell us a little bit about what you

00:19:34.760 --> 00:19:38.517
and uh Sean McGarvey >> Yes.
>> agreed and What's he And he's going to

00:19:38.600 --> 00:19:41.997
be on stage later on. So, let's hear
your version of it and then see if his

00:19:42.080 --> 00:19:46.640
version is the same. >> So,

00:19:48.200 --> 00:19:52.557
we talked earlier and we've talked
The world has talked many times

00:19:52.640 --> 00:19:53.997
in the past about the need for AI

00:19:54.080 --> 00:19:56.437
infrastructure, the need for the world

00:19:56.520 --> 00:20:00.357
to have the physical infrastructure,

00:20:00.440 --> 00:20:03.117
power plants, transmission lines, data

00:20:03.200 --> 00:20:05.997
center halls, chillers,

00:20:06.080 --> 00:20:10.477
obviously the racks and the GPUs and
everything that goes inside of those. Um

00:20:10.560 --> 00:20:14.757
and I sort of wish everyone would go visit

00:20:14.840 --> 00:20:17.157
at some point one of these

00:20:17.240 --> 00:20:21.997
mega scale data centers
because it it's hard,

00:20:22.080 --> 00:20:25.837
you know, when you ask ChatGPT a question,
you get your answer back and it's really

00:20:25.920 --> 00:20:29.877
hard to visualize the scale of what it
took to make that or what it takes to

00:20:29.960 --> 00:20:34.397
make that happen. Uh people
talk about all of these different

00:20:34.480 --> 00:20:36.397
places that were limited. >> Mhm.

00:20:36.480 --> 00:20:38.477
>> We are limited in the, you know, number

00:20:38.560 --> 00:20:43.837
of turbines or now it's the voltage
transformers or it's the memory fabs or

00:20:43.920 --> 00:20:46.557
it's building the data centers,
whatever. All of these things have one

00:20:46.640 --> 00:20:49.757
thing in common, which is they
are massively complex physical

00:20:49.840 --> 00:20:52.077
infrastructure that require

00:20:52.160 --> 00:20:53.797
skilled tradespeople

00:20:53.880 --> 00:20:57.077
uh and a lot of them to do. And

00:20:57.160 --> 00:20:59.357
no matter where the choke point in the

00:20:59.440 --> 00:21:02.677
supply chain is at any one time,

00:21:02.760 --> 00:21:04.397
when I talk to people about what it

00:21:04.480 --> 00:21:07.597
would take to accelerate it, um

00:21:07.680 --> 00:21:11.997
it is more skilled trades workers
to build out all of this infrastructure

00:21:12.080 --> 00:21:15.557
that we all depend on.
I think this will be incredible jobs.

00:21:15.640 --> 00:21:19.997
Uh more than that, I will I think they
will lay the foundation for the next

00:21:20.080 --> 00:21:23.837
generation of American infrastructure
and economic prosperity. And we are

00:21:23.920 --> 00:21:28.197
thrilled to get to work uh
together to to drive that faster.

00:21:28.280 --> 00:21:32.600
>> Uh just so you know,
we at BlackRock share that