Latest / Elon Musk Podcast / Elon Musk Reveals xAI & SpaceX Masterplan - Full Musk Speech

Transcript

Elon Musk Reveals xAI & SpaceX Masterplan - Full Musk Speech · Elon Musk Podcast

0:00
Welcome. To the XAI all hands, very
0:04
exciting presentation for you. We're we're going to start off
0:06
by recapping the incredible progress that the XAI team has
0:10
made in just 2 1/2 years. It's really remarkable in
0:14
pursuit of our goal of understanding the universe.
0:16
So just going over our accomplishments since inception,
0:20
it's important for bear in mind that XAI is only 2 1/2 years
0:24
old, basically a toddler, and we've nonetheless achieved an
0:27
incredible amount in a very short period of time.
0:30
So our competitors are 510, some cases 20 years old.
0:36
They have much larger teams. They started off with formal
0:39
resources. And yet nonetheless we have
0:41
achieved #1 in many arenas in, in just a few years.
0:45
So we've achieved #1 in, in voice, in image and video
0:50
generation. I think we're now at this point
0:52
are actually generating more images and video based on the
0:56
last numbers I saw then all of our competitors combined, we are
1:01
winning in terms of forecasting, which is one of the key metrics
1:04
of intelligence of so up the Grok 420 forecasting model beat
1:09
all the other AIS in forecasting and we've talked many
1:13
leaderboards. We've got now a great app with
1:17
the with the imagine with the core Grok.
1:20
We've made radical improvements to the X app and we've launched
1:24
A Grokopedia which is on its way to far exceeding Wikipedia.
1:28
And ultimately the magnitude are more comprehensive and more
1:31
accurate and have more information as well as video and
1:35
and and image data that simply isn't there on Wikipedia.
1:39
So it's, it's intended ultimately to be Encyclopedia
1:42
Galactica, a distillation of all knowledge of yeah, all
1:46
knowledge. And we're, we're the first to
1:50
achieve 100,000 H 100 GPU training cluster and we're now
1:56
about to achieve the first hundred, I should say 1,000,000
2:00
H 100 GPU equivalents in training.
2:02
So a really an incredible amount of work in a very short period
2:06
of time. And it's important to consider
2:08
for competitiveness of any technology company what matters.
2:12
It's not the position at any point in time, but what is your
2:15
velocity and acceleration? And if you're moving faster than
2:19
anyone else in any given technology arena, you will be
2:22
the leader. And XEI is moving faster than
2:24
any other company. No one's even close.
2:26
So let's go to our team. As we grow as a company, a
2:31
natural thing that happens is you reorganize the company as it
2:34
scales up. So when you first have a
2:36
startup, you might have just a few dozen people and they all
2:40
just chat amongst themselves. As you grow to several 100
2:42
people, you have to then add more structure, just like an
2:46
Organism that grows from a single, like we all just grew
2:49
from a single cell and then to a BLOB of cells.
2:51
Then you get organ differentiation limbs, you grow
2:54
a tail, hopefully a tail disappears and then you become a
2:58
baby. You go through these stages.
3:00
And so we're organizing because we've reached a certain scale.
3:05
We're organizing the company to be more effective at this scale.
3:10
And naturally, when this happens, there's some people who
3:12
are better suited for the early stages of a company and unless
3:16
suited for the later stages. And so and for the people that
3:21
have departed, I'd just like to say thank you for your kind of
3:23
contribution. Thank you for getting us this
3:25
far and we wish you very well in your future endeavors.
3:28
So now going on to the new structure of the company, the
3:33
companies organized in four main application areas.
3:36
There's there's Grok main and voice, which is really the main
3:39
Grok model. It's most called Grok Main.
3:42
Then there's a coding specific model, There's an image and
3:45
video model, which is Imagine and then Macro hard, which is
3:48
intended to do full digital emulation of entire companies.
3:52
And then we've got the infrastructure layers.
3:55
So I'd like to invite members of the teams come up and talk about
3:57
each of their areas. Hey thanks, you know so grok
4:05
main and voice. Are going to be merged.
4:07
Into one team and you know on voice 1 anecdote is September
4:11
2024 opening. I had this product you could
4:13
talk to advanced voice mode, and we had nothing, no model.
4:17
Of course, no product. We started much after that and
4:20
in a span of few months, six months, we developed the model
4:23
in house from scratch without a bunch of people who knew audio
4:26
and had a product that was surpassing opening in six
4:28
months. Fast forward six more months and
4:30
now we have Grok in more than two million Teslas.
4:34
We have a Grog voice agent API. You can do all kinds of amazing
4:37
things. In a span of one year, we went
4:38
from nothing to being leaders. That kind of stuff is only
4:41
possible in a place like XCI. We have small teams, committed
4:44
mission, focus, lots of compute, and we really, really want to
4:48
keep pushing same story on the chat models.
4:50
You know, we've always been at the forefront of reasoning,
4:53
starting from Grok 1.5, Grok 2, Grok 3, and we want to really
4:57
move to a world where it's no longer about this question
4:59
answering, we want to build and everything up.
5:01
So you should be able to come to it and really get done whatever
5:04
you want, you know, ask a legal question, make a slide deck or
5:07
or you know, solve a puzzle, stuff like that.
5:10
Yeah. So I really think on the
5:11
product. Side we're.
5:12
Really going to see a huge. Transformation happening in a
5:15
very short period of time. We're going to see the magnitude
5:19
of amount of work that all knowledge workers are going to
5:22
be. Able to produce.
5:23
Increase tenfold in the next. Short period of a few months,
5:27
the models that we are. Building out are incredibly
5:29
amazing. And we have a lot on the way and
5:32
we're. Really excited to share.
5:33
That with with you all. And on a product side?
5:35
The goal is to just build that portal that allows you to
5:38
accomplish all of your work. And how do?
5:40
We amplify everyone to achieve. Much, much more than what they
5:43
can accomplish alone. And we're building that out and
5:46
it's it's going to be an incredibly easy to use
5:48
experience that just works seamlessly too.
5:51
Being said, we are. Hiring we're looking for.
5:53
Intelligent and smart people. This is not an easy place to
5:56
work, guys like this is it's a grind.
5:58
But we have I guess like interstellar ambitions.
6:01
So it's it's not going to be easy, right?
6:03
So I will say haven't come to X AI, it has been an.
6:08
Opportunity for lifetimes. Work among really smart and
6:10
really passionate people. The vibes here are amazing and
6:14
it's truly an environment where if you're a smart person and you
6:17
want to get shit done, you can get shit done.
6:19
There isn't like organizational overhead getting your way or
6:23
kind of, I don't know. Like having to write docs.
6:25
And all this kind of stuff, you just do stuff.
6:27
At least for me, I just, you know, just you can do things
6:30
here and that's amazing. And I invite more people to come
6:33
here and just do an awesome thing.
6:35
Yeah. So with the Grok main, the sort
6:38
of main foundation model, the intent is that it's genuinely
6:42
useful in a wide range of areas. So if you're doing engineering
6:46
or law or medicine, any, anything it is useful to you in
6:52
in your job that's essential to understanding the universe and
6:57
and making things as useful as possible.
6:59
Like when Grok gives you an answer that you can count on it,
7:02
right. All right.
7:03
Thank you. Thanks.
7:08
Hey everybody. I'm macro.
7:10
So the world changed a lot recently in terms of coding, the
7:14
coding models. I was always complaining, people
7:16
were trying to convince me to use a coding model and I was
7:18
like trusting it and I wasn't really convinced.
7:21
But as of recently, the models, they, they actually produce
7:25
good, decent quality code. I mean, you still need to review
7:28
and give feedback, but you can it's, it's easy to see how they
7:32
can accelerate you quite a lot. It's not only about coding.
7:35
It's like they understand your intuition like much better than
7:39
before. Like now when you are when I
7:41
describe a problem, I only have to phrase it like I would to
7:45
another colleague engineer who has already seen the code base.
7:47
That's a huge change before you kind of need to hand hold a
7:50
toddler to make a change. And they don't only write your
7:53
code, but they also can debug your code.
7:55
So now we have, I do like what we do, like hours of grog code
7:59
running continuously to make sure that a more complex changed
8:05
the training system actually works in production.
8:07
So it's easy to see for us that this is not only about
8:10
accelerating as ourselves writing code and making US10X
8:13
more productive, but we are really on this path for
8:16
recursive self improvement where the current generation of grow
8:20
code is training the next generation of croc code.
8:23
And we see that this path, yeah, an exponential take off here,
8:26
this path will continue. So we are doubling down on
8:29
coding and making coding one of the highest priority efforts in
8:31
the company. So if you're out there and
8:33
you're excited about coding and you're either very good at
8:37
training modeling or you're a really good low level software
8:40
engineer interesting in systems design, this is the place to
8:43
work like we have a million H 100 equivalents to train the
8:49
best coding model in the world right now.
8:50
So please join us. Yeah, I'm Gordon.
8:54
I work here with macro on coding.
8:56
So it'll become more and more obvious to us like in all the
8:58
time, like we are on a path to singularity, at least on coding.
9:02
So we decided like, you know, have our best engineer in the
9:05
company macro to lead the coding and we'll build the best coding
9:08
model for everyone to empower everyone to build.
9:11
And for me, like the man like limiting factor is probably
9:15
computer energy, whether it can run the best model to support
9:18
everyone, to empower everyone. And with specs.
9:21
Now we are one king and we will win on the compute and we are
9:25
win with space compute and also like for every engineer, right?
9:29
So if you are like writing kernel, if you're writing
9:32
compiler, just think about like whether it's still worth it.
9:35
Maybe you should join us, you know, for coding effort to
9:37
automate yourself a letter like to speed yourself up.
9:39
Umm, yeah. I think it's like really amazing
9:42
year. Umm, basically what a year to be
9:44
alive. And I can't really feel the AGI.
9:47
Feel the AGI, at least for coding, yeah.
9:49
Yeah, I think actually things will move maybe even by the end
9:53
of this year to where you don't even bother do it during coding.
9:57
The AI just creates the binary directly and the AI can create a
10:01
much more efficient binary than can be done by any compiler.
10:05
So just say create optimized binary for this particular
10:08
outcome and, and you actually bypass even traditional coding
10:12
that there's, there's no that that's an intermediate step that
10:15
actually will not be needed probably by I'd say the end of
10:19
this year. And we do expect graph code to
10:24
be state-of-the-art in two to three months.
10:26
So it's happening very quickly. Also do imagine so you know, I
10:35
mean what will you do right after post AGI, right?
10:37
You probably do like digital life.
10:39
So that's what we are doing here as well.
10:41
And we have the Imagine team like started pretty much from
10:43
scratch like 6 months ago. We have a few people we decided
10:47
we have to do the imaging, we'll do the video Gen. like yeah,
10:50
look at what we achieved today. Like, you know, like 2 weeks ago
10:53
we released like Imagine, we won.
10:54
We actually topped the leaderboard across like many of
10:57
them. And people really loves our
10:59
product, love our model. And we have many more releases
11:02
actually this month and next month.
11:03
So yeah, to me, there's like really high chance, like we
11:08
actually made view the metaverse before Meta, Yeah.
11:12
Also has to try to to talk about like, you know, the metrics we
11:15
have the product. Yeah, yeah.
11:16
Like like Gorong said, it's only been six months since we started
11:19
working on Imagine. We had, we had no code
11:21
internally for diffusion at all six months ago.
11:24
And basically now we've launched Imagine on every product surface
11:28
that we have, including seamlessly integrating into X.
11:31
So you can open the X app right now, you can long press on any
11:33
image, you can edit the image, you can make a video out of the
11:35
image. We also ran a contest recently
11:38
where we had some really funny submissions that I'm sure many
11:40
of you have seen. So Imagine it's growing
11:42
extremely, extremely fast. And it's because of the speed at
11:45
which we iterate. Basically, we do multiple
11:47
product updates every day. We do model updates every other
11:49
week. And effectively what this has
11:51
led to is now users are generating close to 50 million
11:54
videos every day using Imagine. And just to reiterate what Elon
11:58
said earlier, that to the best of our knowledge, that is more
12:01
than every other provider combined, which again, is an
12:03
astonishing place to be compared to where we were six months ago.
12:06
We are also generating 6 billion images in the last 30 days.
12:10
Nano Banana, you know, Google recently posted that, you know,
12:13
1 billion images were generated using nano banana in 30 days.
12:16
So you know, we're six times that, right?
12:19
And really the goal is it's not like we don't just want to win.
12:23
We want to win like, like over a long period of time and have
12:26
sustained greatness. And so the goal with imagine is
12:28
to take anything that you can, you know, imagine and turn it
12:31
into reality. And so that's that's what we're
12:33
going to, you know, that we're going to speed run.
12:35
That basically is the goal, Yeah.
12:37
Hey, I'm hunting. We, as we keep scaling our model
12:41
capabilities, building visual worlds that's indistinguishable
12:44
from reality, we're also building systems that unlocks
12:48
much more possibility than what we have right now.
12:52
They will be able to generate the videos that's much longer
12:54
than what we have right now with stories or with souls of your
12:58
imagine. And by the end of the year, we
13:00
likely will be having models that allow you to generate
13:04
videos of 10 minutes or 20 minutes in one shot without any
13:09
intervention. You just need to give your
13:11
imagination and our model, our agents will do it for you.
13:14
And and moreover, those are the videos we generate and we're
13:19
also going to allow rendering those.
13:22
We're already the fastest in generating the videos and we're
13:25
going to keep pushing the extreme where we're going to
13:27
render those videos in real time.
13:30
And you will be able to imagine, build and interact with your own
13:34
world and the world will respond to you in real time.
13:38
And it is exciting future that we are going to build with
13:41
ourselves. Absolutely.
13:43
My prediction is that most of AI compute is going to be real time
13:48
video understanding real time video generation and and we
13:51
expect to be the least in that. It's worth emphasizing these
13:54
points that, you know, six months ago we didn't even have
13:57
we had basically nothing in very weak in video and image
14:02
generation and editing. And we're in six months to
14:05
number one spot and in fact, generally more videos and images
14:08
and everyone else combined. We're going to do the same thing
14:10
with coding and we're going to do the same thing with macro
14:13
hard. And I think people will be
14:16
pretty impressed with the Grok full Green 2 models coming now.
14:19
That's it's a it's a significant improvement.
14:22
And that's really just that that's, that's the small version
14:25
of our new model. So we will have a medium and a
14:29
large version that are even more intelligent.
14:36
All right. Hi everyone.
14:37
I'm Toby and I work on Macro Heart, the most serious of all
14:42
product names. So arguably giving computers to
14:46
humans was a good idea. So we're doing the same thing
14:48
thing for AI. It's kind of like Inception,
14:50
we're giving computers to computers.
14:52
So Macro Heart is building a fully capable, digital, real
14:56
time, very important human emulator.
14:59
So it's able to do anything on a computer that a human is able to
15:02
do, including using advanced tools and engineering and
15:06
medicine. So there should be rocket
15:08
engines fully designed by AI. And in a sense, it's one of the
15:12
last few remaining areas where AI is significantly worse than
15:16
humans, which is why I think it's one of the most exciting
15:19
areas to actually innovate in and actually change the change
15:23
the field. Hi, everyone.
15:24
So, yeah, my name's John. And yeah, so we're building
15:27
these strong reasoning models, which are now going to control
15:30
our CLI. Like we're actively using these
15:32
every day. They are like tremendous, like
15:35
productivity boost to the whole team.
15:36
I know the voice team is like killing it on that.
15:39
And you know, this is the reason why we need the compute.
15:41
You know we need the large scale computer on these models to
15:44
boost our own productivity. But you know, 80 to 9095% of the
15:49
world world software has a GUI. So that's like, you know.
15:53
Great representation and you know, to truly make people's
15:56
lives easier, we need to develop models that are capable of
15:59
solving day-to-day tasks on GUI. So macro hearts, you know, we
16:04
will emulate a company where the output is digital.
16:07
And so this is the obvious next step for agents.
16:10
Macro hard will enable true end to end orchestration across the
16:13
desktop and it will lead to immense economic prosperity.
16:16
Umm, so yeah, we're entering an era where we need to tackle the
16:20
hardest of tech problems. But in order to solve this, we
16:22
need to hire the best people. So, you know, think of the
16:25
smartest people that you've worked with and I'm put them
16:28
forward for, for a position here.
16:29
And if you can't think of anybody, like, go through your
16:32
phone book, go for your LinkedIn, you'll be surprised
16:35
like how big your actual network is.
16:37
And they just need 3 properties, obviously, that we want to
16:40
optimize for. Are they clever?
16:42
Can they solve hard problems? And the second property is, are
16:45
they driven? Do they have the ambition?
16:47
Do they want to win? And the third is, are they a
16:49
nice person? Like, do you want to actually
16:51
work with them? Yeah.
16:53
So thank you. Yeah, the macro hard project is
16:56
overtime like she will probably be our most important project
17:00
because what we're talking about is emulation of entire human
17:04
companies. So when you look at the most
17:07
valuable companies in the world, they are their output is
17:11
digital, so they don't actually make hardware.
17:15
So it should be possible to completely emulate any company
17:18
that where the output is digital.
17:20
And this will usher in an age of prosperity likes which we could
17:24
barely imagine at this point. You need to imagine to imagine
17:27
it. So this is a big, this is a big
17:29
deal. And This is why the words macro
17:31
hard are painted on the roof of the training cluster, because
17:34
that's what it's going to bolt. It's also pretty funny, yeah.
17:37
Glad to be a joke. It's me again.
17:45
You might remember. Me from Micro Heart and computer
17:47
use from a long time ago, but I also actually work on core
17:51
product infrastructure and API. In fact, this is what I've done
17:54
for most time at XAI. So anytime you use any of our
17:58
products like group.com API authentication, you go to status
18:02
dot X dot AI. This is done by the core product
18:05
infra team and a large portion of them actually sit in London
18:08
and we work with Jaime over there.
18:10
So we keep the lights on at peak hour, 4:00 PM.
18:13
Maybe we get paged at night when stuff goes down.
18:17
Also, thank you to anyone in Palo Alto getting paged.
18:20
There's really important work, reliability, security, core
18:23
product infrastructure. So if you actually, if you're
18:26
really interested in solving difficult distributed problems
18:29
with like, messy data, this is the team to join.
18:34
Hey, everyone. My name is Diego.
18:36
Yeah. So I think one of the main
18:38
bottlenecks in this next year for these models is going to be
18:40
very high quality evals and training data.
18:42
And one of the ways we solve that is by taking the world's
18:44
foremost experts in these rich domains, bringing them here and
18:48
having them evaluate them up. We do this for domains like
18:50
medicine, finance, law. We have voice actors.
18:53
We have video editors who contribute daily to making rock
18:57
better. And yeah, we're going to be
18:59
continuing to work on very high quality evals over the next few
19:01
months. We have some exciting stuff in,
19:04
you know, the frontier of useful tasks in finance and law.
19:08
You know, we're trying to build evals that are are useful in
19:11
training data that represents useful work and not necessarily
19:15
proxies of intelligence without a lot of the open source evals
19:18
do today, yeah. Yeah, I'd like to like we're,
19:24
we're shifting from using these sort of common Internet evals,
19:28
which I think are actually not a real indicator of usefulness, to
19:32
having expert tutors in each domain.
19:36
So every domain of engineering, medicine, law, whatever the case
19:38
may be. And the, the actual eval is,
19:41
does the expert in that arena or does our group of experts in
19:45
that arena, human experts agree that Grok is extremely useful
19:50
and that the results are correct.
19:51
That's the. That's actually the only eval
19:53
that really matters. Exactly in you'll see this in
19:57
GRAC 420. But we've made some improvements
19:59
because of that type of data in truth seeking and kind of
20:02
minimizing political bias. Other responses are much more
20:04
cogent. Yeah, that's exciting.
20:07
And we are also working on Gracopedia.
20:09
So the the goal of Wikipedia is to create a distillation of all
20:12
human knowledge. I kind of like to think of this
20:14
as like a modern day version of the Library of Alexandria.
20:17
And in the quest to build Encyclopedia Galactica, will one
20:21
day be cult. We've gone from essentially
20:23
having nothing to around 6,000,000 articles.
20:26
For context, Wikipedia is around 7 million English articles.
20:29
And yeah, we're we're improving on hallucination and our goal is
20:35
essentially for Rock 5 to not have to search out of the data
20:37
center. So yeah.
20:46
So in the ML infra team, we are building the training inference
20:50
and tooling team. It is tooling software for the
20:52
company. So it's giving you an example.
20:54
When we were training Grok 3, we built the pre training framework
20:57
for this and it these are some some of the coolest system in my
21:01
opinion that you can build as a software engineer.
21:03
So it's like we have 100 KH, one hundreds at the time and they
21:07
were just delivered and we didn't quite have the software.
21:10
We thought we'd have the software, but then AT30K scale
21:13
we realized actually the software is not quite working.
21:16
And it took a major almost, I was like halfway rewrite of the
21:21
software because there's so much going on in a data center that
21:24
you can't actually account for switches are switches are
21:28
flapping, links are flapping, switches are going down, GPU's
21:32
are just burning through. You have numerics issues and
21:35
it's a system where you want really 100 KH one hundreds to
21:38
behave in lockstep. So a training step is like 5
21:41
seconds and you're going 5 seconds in lockstep, but during
21:44
that 5 seconds everything can happen.
21:46
So you need to write a system that makes progress despite all
21:49
these things that can happen in the environment.
21:51
And we did this successfully and was one of the coolest times in
21:53
my life where the system was actually running and it was
21:56
running at the same time my son was born.
21:58
So there was extra excitement. Umm, but these problems like you
22:02
don't find anywhere else, like nobody has this kind of compute
22:06
and also nobody has this kind of talent density.
22:09
So at the time, to give you a perspective, we were like an
22:11
overall team in pre training. We were probably like 15 people
22:14
out of that, maybe like seven people were working on the
22:17
actual training system. And we still maintain that
22:20
talent, talent density in the team.
22:22
So if you're interested in working on these problems and
22:25
you don't want to be just like part of a bigger organization
22:27
where you're 1 of like 1000 people working on this, then
22:30
this is the place. Like we are still a very small
22:33
team with me as Leon Min from the RL and inference team.
22:36
Hi, I'm Lemmy. So at our team, we run our
22:38
reinforcement learning training job and the production inference
22:41
has a large scale on the Earth and probably soon in space.
22:44
Uh, and we are kind of already designed a lot of thing to, to
22:50
make it more resilient and scalable.
22:52
So we're building a system to scale from 100K chips to
22:55
millions of chips. And we organize every aspect of
22:59
the stack like parallelism, pre fuel decode and make resilient
23:04
to every numb and unknown hardware failure.
23:07
So if you are system hackers obsessed with extreme
23:10
performance and reliability. So here is you'll find the most
23:14
interesting problems to work with.
23:16
And I think actually like very similar to all kind of things
23:20
like you, it's it's very important for you to 1st see the
23:23
problem and then you will develop the solutions that no
23:27
one else can develop before. I'll handle to the tooling team.
23:30
Hello, I'm Ashley from the tooling team.
23:33
Every software needs to have a great interface to be able to
23:37
make it useful. So as the tooling team, we are
23:39
responsible for building the platforms, frameworks, and
23:42
infrastructure which is required for humans as well as agents to
23:46
be able to use our products. We started by building out the
23:49
human data platform. This is the place where we
23:51
collect all of our human data and eventually expanded on to
23:55
build our internal engineering platform through which we
23:58
basically run deployments, run evaluations, or like look at
24:02
like what training results exist.
24:03
So if you really care about building a good interface or
24:07
providing a really useful framework for researchers, for
24:10
agents as well as our tutors, then we should definitely join
24:13
our team. Hi, everyone, I'm you know from
24:15
the JAX team. So now JAX at X AI, it's a
24:18
really small team with a couple of engineers that working on Jax
24:22
GPU to optimize our ultra large scale GPU training.
24:27
So you can imagine that training at scale can be very
24:29
complicated. Even you run Hello world at
24:31
scale, it can be complicated, right?
24:33
So then we're actually responsible for supporting the
24:38
entire companies from pre training foundation models, RLS
24:41
and also multimodal to scale things to from from first from
24:48
10K Andre Canton, probably 1,000,000 H 100 equivalent GPU
24:53
scale. And we to implement a lot of,
24:56
you know, practical optimizations, we have to
25:00
customize the entire JAX stack from compiler and runtimes and
25:04
there will be a lot of interesting problems.
25:07
And also if you really want to, you know, obsessed on optimizing
25:13
the entire at scale, we are probably the best place to go
25:17
because, you know, we really have a very large scale GPU
25:20
clusters and we have a lot of interesting problems to work
25:22
with. Hey, I'm Kanjal from the kernels
25:24
team. Basically the kernel team sits
25:26
at the very bottom of our training and serving stack.
25:28
Our code runs inside the million equivalent GPU's that we have.
25:32
And if you look inside the GPU, there's hundreds of thousands of
25:34
threads. And these threads are trying to
25:36
talk to each other to multiply matrices, compute attention
25:39
scores, and some of them even talk to the million other GPU's
25:42
that we have. And this is the low level system
25:44
that we have and we like optimizing every single
25:46
microsecond in this and we care deeply about squeezing every
25:49
last couple of performance from this Gpos.
25:51
So if you like this low level systems problems algorithms,
25:54
please join us. As you know, try to bring in
26:02
Heiner and Spencer who are actually at our Tuber compute
26:07
cluster in Memphis. Hey, Heiner.
26:13
I'm from the computer network infrastructure team.
26:17
We are mainly based in the supercomputer.
26:21
So the data center here Memphis from on the planet and it is
26:26
still growing our keep all this computer and the used to work
26:35
well left ingredients. Actually just put the mic really
26:39
close to your mouth because you're the ambient noise is
26:41
high. So keep.
26:46
The computer up and running. Trying the next model of.
26:48
Truck and serve TI so. Users so which it used to work
26:51
well a lot of them greens have to come together mainly software
26:54
and hardware. So there's all these to be CPUs,
26:58
NICs, switches of hundreds of thousands of operating systems
27:01
running as one big supercomputer.
27:02
And what we need is sports to really understand the notes,
27:05
really understand that we may, and really understand how.
27:07
Computers work on deep level. That is you.
27:09
Reach out the next. And I'm ending up with them all.
27:12
Right, so we have. 300,000. TV 300 platform excuse here
27:18
today still growing, still building 847 miles of fiber per
27:23
data hall 12 data halls you want to be.
27:25
Part of the world's largest. Supercomputer come join us all
27:29
right so it's quite marvelous what we've been able to do in
27:32
less than. One years time here.
27:34
We have once we're. Completely finished.
27:37
We'll have north of a GW of power Online is running.
27:40
We'll have the largest Tesla Megapack system in the world.
27:43
Large spin in Hawaii. Or SA and Zach is really quickly
27:47
going to talk a little bit about actually constructing the data
27:50
center. So behind me you can see Data
27:52
Hall 11. So one of the most incredible
27:54
things about what we're doing here at Macro Hearts, how fast
27:57
we do it, right? So like they were saying before,
27:59
over 850 miles of fiber and every single data hall over
28:02
27,000 and over a 200,000 connections.
28:07
So all of this. That you can see behind me was
28:10
put up in less than six weeks. We do that.
28:12
Over and over over again. We massively parallelize them.
28:15
It's pretty much the most complex and consistent type of
28:19
engineering, design and construction project.
28:22
You possibly. Imagine.
28:23
So come join. Us, yes.
28:25
You know the other really awesome thing about this is that
28:27
everything is completely vertically integrated within
28:30
this team. From architecture, mechanical,
28:32
electrical starts all the disciplines and we also care a
28:35
lot about efficiency while we're designing all of this too, so
28:38
it's not just about getting the most compute online the.
28:41
Fastest, but also achieving the highest.
28:43
PUE in the industry of using as much power smoothing technology
28:48
as we can and being really good partners in the community here
28:51
in Memphis. With the Tesla Mega packs and
28:55
things that we have. Going you can check them out.
28:57
XAI Memphis back to you. All right, thank you.
29:04
All right, So that was live, live from the front lines in
29:06
Memphis. So fundamental to any AI
29:11
company's success is the computer advantage.
29:13
And what we've demonstrated over and over again is that XAI can
29:16
actually deploy more AI compute faster than anyone else.
29:20
And actually, as Justin Wong of CEO of Video has said many times
29:25
in interviews, there is no one faster at getting AI compute
29:28
online than XAI. So congratulations, guys.
29:34
Yeah, this is what it looks like.
29:36
So that's a really phase one, which is 330,000 Grace
29:41
Blackwells with macro hard written on the building.
29:44
That's an image edit. It actually is on the roof of
29:46
the building. And then macro harder will be
29:49
the building that you can see which has got the macro harder
29:52
with rockets on it. And that will be another 220,000
29:56
GB, three hundreds. So all of this will be training
29:59
or the models that you that you experience.
30:02
So the it's a absolutely fundamental obviously to have
30:06
large scale training compute in order to get the best models.
30:09
Yeah, I'm sort of reminded of the Jose mean where you see one
30:13
guy digging and there's like 7 people watching and one of the
30:16
big differences between X AI and other companies is we are
30:20
actually Jose. Hello.
30:21
All right, I'm Nikita. You might know me as a part time
30:24
ship poster, full time customer support for X.
30:28
So we're now reaching over a billion people across our family
30:32
of apps. Every time news breaks, it just
30:35
becomes evident that this is the most important communication
30:38
tool of our time. It's where the the the most
30:41
influential people come convene. It's where truth is
30:45
crystallized. Everything is downstream of X.
30:48
The reason they say this is going to hit Facebook in a week
30:51
because it happens here. And I think we're only beginning
30:56
to realize its full potential. We had a remarkable year for the
30:59
app. We rolled up our sleeves and got
31:02
a ton done. January was our biggest month
31:06
ever for the app in terms of engagement and then February is
31:11
on track to beat that. Much of the credit lies with the
31:14
algorithm team. They've been putting in crazy
31:17
hours and it's clearly paying off, but there's still a huge
31:20
amount of work to be done on the top of funnel side.
31:24
First time downloads are up over 50% every month and we're
31:28
exhibiting right now like basically the growth rates of an
31:31
early stage consumer product. We also made a ton of headway
31:35
and solving one of the like 20 year old problems of the app,
31:38
which was ramping up new users. New users are now spending 55%
31:43
more time per day in the app than they were six months ago.
31:48
And on the core product side, we're hitting our stride to not
31:53
only did we rebuild the algorithm, we rebuilt our
31:55
onboarding flows and we're seeing double digit increases on
31:58
all our key metrics. We rebuilt notifications, our
32:01
web browser, X chat, basically every surface of the app has
32:06
been rebuilt to be better than ever.
32:08
And it's clear that if we're focused, we can move mountains
32:12
and evolve this platform. Just last month, we did a little
32:15
push on articles and articles published are up 10X articles
32:23
read or UP17X. And on all other fronts like
32:28
under over the holidays, we did a big push on subscriptions.
32:31
We just crossed a billion dollars in Arkansas.
32:33
Are there I I think with the X app, you know, the there's very
32:37
few unknowns like the path for us to win and become, you know,
32:42
the number one app in the world. We're it's it's we we know what
32:46
to do. The ball's in our court.
32:49
It's it's for us to win and it's just a matter of us to
32:52
executing. Yep.
32:53
And yeah, so we've evolved the what used to be the old Twitter
33:02
DM stack, which was unencrypted, basically just text to a fully
33:06
encrypted messaging system that is allows you to do audio and
33:10
video calls, has, you know, all the things you'd want from any
33:14
messaging app, the distributing messages, screen screenshot
33:17
blocks, like there's a whole all the features that you'd want
33:19
want in an app. We and we will be open sourcing
33:23
the code for this in the next few months as we are open
33:26
sourcing the recommendation algorithm code so people can
33:28
actually see what we're doing. Nothing beats nothing beats
33:33
transparency for believing in in a company.
33:37
So we're going to be the only recommendation algorithm that
33:42
actually open sources so you can see what it what it does and how
33:44
it's evolving with, with Grok Chat, it will also be open
33:48
source. So you can actually see if there
33:50
are any vulnerabilities. There will be no hooks for
33:52
advertising or anything else like that in in Grok Chat, which
33:55
is really intended to be a generalized communication
33:58
system. And in the next few months we'll
34:00
be releasing A standalone X chat app.
34:04
So if you just want to do messaging, you can just you can
34:06
do that. You don't you don't have to go
34:07
to the the core product and we'll have desktop sharing and
34:13
multi user so you can do video calls with lots of people.
34:16
It's really intended to be a a fully functional communication
34:21
system with Xchat for X money. We're we've actually had X money
34:27
live in closed beta within the company and we expect in the
34:30
next month or two to go to a limited external beta and then
34:36
to go worldwide to all X users. And this is really intended to
34:40
be the place where all the money is the the central source of of
34:45
all monetary transactions. So it's, so it's really going to
34:48
be a game changer. And the reason we say 1 billion
34:52
users is actually over a billion users is that while our monthly
34:55
users are on average around 600 million, the number of people
35:00
who have the X app installed is well over a billion.
35:03
It's just that most people only occasionally come to the X app
35:06
when there's some major world event.
35:08
But as we give people more reasons to use the the app,
35:12
whether it's for communications, for Rock or for X money,
35:18
whatever the case may be, we want it to be such that if you
35:21
want to, you could live your life on the X app.
35:24
And as you make it more and more useful, we'll obviously give
35:26
people reasons, compelling reasons to use the app every day
35:30
and have my expectation is well over a billion daily active
35:34
users. Now, in order to understand the
35:37
universe, you must explore the universe.
35:39
There's only so much you can learn from from just being on
35:43
Earth with telescopes and colliders on Earth.
35:46
Ultimately, you have to go out there and you have to explore
35:48
the universe to understand it. And that's the motivation behind
35:52
the combination of SpaceX and XAI, is to accelerate humanity's
35:57
future in understanding the universe and extending the light
36:00
of consciousness to the stars. So in the grand scheme of
36:03
things, when you look at how much energy Earth is actually
36:06
using for civilization, we're only right now using, call it
36:10
roughly 1% of the potential energy of Earth.
36:13
And if we wanted to use even a millionth of the sun's energy,
36:17
that would be roughly a million times more energy than
36:20
civilization currently uses. The only way to access that that
36:24
energy, the energy of the sun is to extend beyond Earth.
36:27
Earth is really a tiny, tiny dust mote in, in a vast
36:30
darkness. You know, the sun is 99.8% of
36:34
all mass in the solar system. So you, you have to expand
36:37
beyond the tiny dust mote that is Earth to, to make any
36:42
significant dent in using the sun's energy like says you'd
36:45
have to expand roughly a million times just to get to 1 millionth
36:48
of our of our sun's energy. And then going beyond that,
36:51
exploring, extending to the Galaxy and maybe someday even to
36:55
other galaxies. So the the next step beyond
37:00
Earth data centers is our Earth orbital data centers.
37:04
And we'll be launching with SpaceX orbital data centers at
37:08
the 100 to 200 GW per your level, not cumulative, I mean,
37:12
per year. And ultimately we see a path to
37:15
maybe launching as much as a terawatt per year of compute
37:19
from Earth. But what if you want to go
37:21
beyond a mere terawatt per year? In order to do that, you have to
37:24
go to the moon South by having factories on the moon, building
37:28
AI satellites and having a mass driver, which is the kind of
37:31
thing you really learn about in read about in science fiction.
37:34
But we're going to make it real. We're actually going to have a
37:37
mass driver on the moon. And if you do that, you can go
37:41
several orders of magnitude greater.
37:42
You can go to 1000 gigawatts or more per year and ultimately get
37:48
to maybe a millionth and then a thousandth and maybe even a few
37:52
percent of the sun's energy. It's difficult to imagine what
37:55
an intelligence of that scale would think about, but it's
37:58
going to be incredibly exciting to see it happen.
38:00
I really want to see the mass driver on the moon that is
38:04
shooting AI satellites into deep space.
38:06
It's go like shroom shroom, just the one after the other.
38:09
I can't imagine anything more epic than a mass driver on the
38:12
moon and a self-sustaining city on the moon.
38:14
And then going beyond the moon, Mars I going throughout our
38:18
solar system and ultimately going being out there among the
38:21
stars and visiting all these star systems.
38:24
Maybe we'll meet aliens. Maybe we'll meet see some
38:27
civilizations that lasted for millions of years and we'll find
38:30
the remnants of ancient alien civilizations.
38:33
But the only way we're going to do that do that is if we go out
38:35
there and we explore. And this is the path to making
38:37
it happen. Thank you.