00:00:00
All right, everybody. Welcome back to
00:00:01
the number one podcast in the world.
00:00:03
It's the Allin podcast, episode 274.
00:00:07
Sachs is out today, but we're very lucky
00:00:10
to have Gavin Baker from Treaties
00:00:13
Management joining us. The spicy takes
00:00:16
must flow. Welcome back to the program.
00:00:18
Besty Gavin,
00:00:20
>> thanks for having me. Always love it.
00:00:21
>> It's been a huge week in tech. We can
00:00:24
start with the SpaceX and OpenAI IPOs.
00:00:26
We've got Andre Karpathy joining
00:00:28
Anthropic. Nvidia crushing it. So many
00:00:31
different places to go, but I think
00:00:32
we'll start with Andre Carpathy joining
00:00:35
Anthropic. Carpathy is only 39 years
00:00:38
old. He's already a legend in the tech
00:00:39
industry if you don't know him. I
00:00:41
believe he's also coming to liquidity.
00:00:42
Yet,
00:00:42
>> he's going to keynote on Monday morning.
00:00:44
>> Oh, fantastic.
00:00:45
>> No, Tuesday. Tuesday. Tuesday. Day two.
00:00:47
I think he's keing.
00:00:48
>> Okay. As is Gavin. Gavin will be there.
00:00:51
Founding member of
00:00:52
>> Gavin anchoring day two as
00:00:54
>> excellent. Yeah, this is Gavin's second
00:00:56
appearance at
00:00:57
>> Look at those two bookmarks. Andre
00:00:58
Carpathy and Gavin Baker.
00:00:59
>> Oh, yeah. You know,
00:01:01
>> liquidity pulls in the stars.
00:01:03
>> Obviously, Andre was a founding member
00:01:06
of OpenAI. He led the self-driving team.
00:01:08
>> Also, hold on. Gavin is going to help us
00:01:10
judge the best ideas section as well.
00:01:13
Excellent.
00:01:13
>> I don't know if you know that, Gavin,
00:01:14
but you're a judge. You're a judge.
00:01:16
>> I'm up for anything, man. I'm easy.
00:01:18
>> Yes. Kurpathy also coined the term vibe
00:01:20
coding. He recently built auto research.
00:01:22
I think we talked about that here a bit.
00:01:24
That's an open source training tool. It
00:01:26
helps AI models improve themselves by
00:01:28
running five-minute experiments. That
00:01:29
got over 82,000 stars on GitHub. He did
00:01:32
that like as a weekend experiment and
00:01:34
all these civilians started building
00:01:36
their own uh recursive LLMs. Really
00:01:38
inspiring. And the Andre Karpathy skills
00:01:42
is a tool based on his set of principles
00:01:45
for claude code. And somebody just
00:01:48
released that. And so that's just pretty
00:01:50
crazy when you think about it. He's
00:01:52
going to be in charge of a new
00:01:53
pre-training team at Anthropic. The
00:01:55
focus obviously being recursive
00:01:57
self-improvement. In other words,
00:02:00
they're going to have Claude improve
00:02:01
itself. And they've already talked a
00:02:02
little bit about AI improving AI over at
00:02:06
anthropic. What's your take on this? Is
00:02:09
this uh super important in 2026?
00:02:12
Obviously, Karpathy is super well
00:02:14
respected. is obviously uh you know one
00:02:17
of the true talents in the space but hey
00:02:19
we're in we're in a different inning
00:02:20
than we were say 10 years ago when he
00:02:22
was at Tesla or five years ago when he
00:02:24
co-founded open AI
00:02:26
>> you know what's interesting if you go
00:02:27
back to like Google the culture of
00:02:29
Google which they got right was the
00:02:31
singular technical talents there
00:02:35
they were singled out and they were
00:02:36
called Google fellows I don't know if
00:02:38
you guys remember this like Ahmed Single
00:02:40
>> Shredar Ramaswami Jeff Dean
00:02:44
these guys are stars. And what's
00:02:46
interesting is if you track what folks,
00:02:48
particularly Jeff Dean, I guess, now
00:02:50
because the other two aren't there
00:02:51
anymore, but what they did inside of
00:02:54
Google, it's like wave upon wave, they
00:02:56
were at the foot of those waves. What's
00:02:58
interesting about Andre is he's been at
00:03:00
the wave upon wave of AI. He was
00:03:02
probably the first person that really
00:03:05
commercialized the Richard Sutton bitter
00:03:08
lesson essay when he was leading FSD at
00:03:12
Tesla which was really about the brute
00:03:15
force computation and I remember him
00:03:17
telling me this story I don't know if he
00:03:18
said this publicly or not but where he
00:03:20
spent a portion of his time I want to
00:03:22
say a quarter of his time labeling data
00:03:24
could you imagine like 201617 like
00:03:27
handlabeling video data from Teslas So
00:03:31
he did that then he co-founder of open
00:03:33
AI he's a star and he's an exceptional
00:03:35
human being and he's super curious and
00:03:38
then what he's done as a kind of a free
00:03:40
agent is also quite impressive. So I
00:03:43
think that this is a really important
00:03:45
deal. I think he's one of these really
00:03:48
curious people that can be sent off and
00:03:51
they'll just go and invent new things.
00:03:54
And I think this idea of recursive
00:03:55
self-learning puts these models on
00:04:00
a combination of overdrive and
00:04:02
autopilot. And so if you put those two
00:04:04
things together, I think that you start
00:04:06
to
00:04:08
you can potentially live out this idea
00:04:10
that there's an order of magnitude
00:04:12
improvement on a yearly basis. So like
00:04:15
this new form of Moore's law. So then
00:04:17
the model quality just goes absolutely
00:04:19
parabolically just like this straight
00:04:21
up. throwing a bunch of compute at the
00:04:24
problem and these things learn really
00:04:25
quick.
00:04:26
>> I think is the high order bit there.
00:04:29
Gavin, what are you what's your take on
00:04:31
Anthropic's recent success and their
00:04:33
massive hiring binge?
00:04:35
>> The success is extraordinary. It's
00:04:37
undeniable. I think the fact that they
00:04:40
are now they were EBIT positive per the
00:04:42
Wall Street Journal in the most recent
00:04:44
quarter is a really important fact for
00:04:47
kind of the whole AI narrative because
00:04:50
now there's
00:04:52
you know you could talk about circular
00:04:54
funding you could talk about ROI and we
00:04:56
could go look at the ROIC of the
00:04:57
hyperscalers
00:04:59
but if OpenAI and Enthropic are at call
00:05:03
it a hundred billion dollars of ARR now
00:05:05
with 80%ish
00:05:08
gross margins on inference like the
00:05:10
returns are there and then if we add in
00:05:12
and they're growing really fast if we
00:05:14
add in Gemini we add in cursor we add in
00:05:16
XAI we add in open source you know it's
00:05:20
it's not hard to see 200 300 400 billion
00:05:25
of ARR at the end of this year at high
00:05:28
market
00:05:28
>> across all of the language models and
00:05:31
you're talking specifically about the
00:05:32
private language model companies maybe
00:05:35
not Google which is including
00:05:38
But I was excluding you know a lot of
00:05:40
the returns to this GPU spend have come
00:05:42
from you know better recommener systems
00:05:45
at Facebook and Google Amazon better ad
00:05:49
targeting better ad measurement.
00:05:51
>> Sure. So I was excluding that and just
00:05:54
narrowing it to LLMs which I think
00:05:56
strictest possible definition and it
00:06:00
seems like there's going to be a really
00:06:02
strong ROI this year even excluding what
00:06:05
are still some of the most economically
00:06:07
important and profitable use cases for
00:06:09
GPUs and AI infrastructure. I do think
00:06:12
what Karpathi is working on recursive
00:06:15
self-improvement is really important and
00:06:18
unlocking that and continual learning
00:06:21
you know maybe the two final frontiers
00:06:24
for for AI
00:06:26
and just the idea of recursive
00:06:28
self-improvement that the model while it
00:06:31
is training you know during a forward
00:06:34
pass
00:06:36
has input into its training or another
00:06:39
model has input into the training
00:06:42
I think that could be really powerful
00:06:44
and I think Chimamoth's statistics of of
00:06:49
you know 10xing every year
00:06:52
you know might seem conservative if that
00:06:54
comes to pass and then of course
00:06:55
continual learning is the holy grail
00:06:58
where the model learns from experiences
00:07:01
the way humans do.
00:07:03
>> Yeah.
00:07:03
>> And that's something we haven't unlocked
00:07:05
yet. And those those those two combined
00:07:08
I think would um they might pull the
00:07:11
future forward in a very real way.
00:07:13
>> Yeah. And we have right now Enthropic
00:07:15
has a decent lead on everybody else
00:07:17
whether it's 3 months or 6 months.
00:07:19
Obviously they're probably 6 12 months
00:07:20
ahead of open source. Maybe they're
00:07:22
three six nine months ahead of their
00:07:23
contemporaries but they have a lead. You
00:07:25
put Cararpathy in there Freedberg. Now
00:07:28
you have Karpathy. He does recursive.
00:07:31
And at some point, and it may have even
00:07:33
occurred at anthropic, the AI is going
00:07:36
to be improving the language model more
00:07:41
than the humans in the loop are doing
00:07:45
it. Obviously, they're orchestrating it,
00:07:47
Freeberg. But at what point do we think
00:07:50
this, let's call it super recursiveness
00:07:52
occurs? When will we cross the recursive
00:07:55
valley? And AI is doing more to build a
00:07:58
language model than humans are.
00:08:02
I'm not sure when this idea that you
00:08:05
feed the whole model into a context
00:08:07
window to train itself and build a new
00:08:09
model is going to happen, but I think
00:08:11
there's probably a lot of different
00:08:12
architectural paths that could be walked
00:08:14
here. One of which is this idea that you
00:08:17
could make much smaller models and then
00:08:19
create networks of smaller models that
00:08:21
work together where you ultimately have
00:08:24
less energy or less cost per token
00:08:26
produced out of a aggregation of models
00:08:29
than you did with one single large
00:08:31
model. I've said this probably three or
00:08:32
four times now. There's a lot of work
00:08:34
and a lot of opportunity ahead in kind
00:08:37
of rearchitecting models and
00:08:39
rearchitecting how models work together
00:08:42
to solve problems. My guess is a lot of
00:08:44
leadership that that he can bring to
00:08:47
exploring those paths and all it takes
00:08:49
is a minor breakthrough and your cost
00:08:51
per token drops in half. That's a
00:08:53
tremendous efficiency gain that is seems
00:08:55
very much on the horizon because some of
00:08:57
the early papers I think I shared one
00:08:59
from MIT a few weeks ago indicate that
00:09:01
there's a lot of room to run here in
00:09:04
terms of rearchitecting models and
00:09:05
deployment of models. these very small
00:09:08
uh models uh the small language models
00:09:11
and then verticalized ones are the
00:09:13
future. We've got a company Abocus
00:09:15
that's doing it for corporations
00:09:17
crushing it. Everybody's got an interest
00:09:19
in doing this. And I don't know if you
00:09:20
saw the news this week uh Chimoth it
00:09:22
happened about two weeks ago very
00:09:24
quietly Chrome included
00:09:27
Gemini or Google included in their
00:09:29
Chrome browser the Gemini Nano model
00:09:32
without telling anybody. 4 gigabytes on
00:09:34
your computer and that's the one that
00:09:36
does like proof reading, spelling,
00:09:38
autocomplete, all that. So now we have
00:09:41
Google covertly installing this on
00:09:44
everybody's operating system.
00:09:46
>> Hold on, hold on. Covert's a strong
00:09:48
word, so let's not use that word.
00:09:49
>> Without telling people, without giving
00:09:51
people a heads up,
00:09:52
>> let's say it in a way that we can both
00:09:54
agree.
00:09:56
We're in the phase now where I think
00:09:58
breathlessly talking about every model
00:10:00
improvement is a waste of time. there's
00:10:02
no ROI in it. We are on a path of
00:10:06
accelerated learning and we're going to
00:10:08
start to see enduser achievements that
00:10:11
were here to for impossible. That should
00:10:14
be the focus. So for example, we were
00:10:17
able to solve I'm just collectively
00:10:19
saying we in this case it was
00:10:21
specifically open AI
00:10:24
by the use of a human and this is
00:10:26
important you know a math problem that
00:10:28
stood outstanding not been solved for
00:10:30
decades and decades
00:10:33
I can tell you in a different example
00:10:35
there are drug candidates that are about
00:10:37
to enter clinical trials and INDs that
00:10:40
were sitting on the shelf and people
00:10:41
didn't think were very viable at
00:10:45
So, we're at the phase now where these
00:10:47
things are front and center. They're
00:10:49
useful to people. They're increasingly
00:10:51
valuable. I think what we should do now
00:10:53
is focus on these end user
00:10:56
use cases because the way that you say
00:10:59
it, in my opinion, is part of the
00:11:01
problem because it starts to create this
00:11:03
boogeyman us versus them thing. And I'm
00:11:06
not saying you're doing it on purpose,
00:11:07
but I'm saying this is exactly why I
00:11:10
think so many people are becoming sort
00:11:12
of like it's a four-letter word now when
00:11:14
you mention AI because it's presented as
00:11:17
this thing. And I think we have to
00:11:18
present the other side of it at least so
00:11:20
that people have the data. So I don't
00:11:22
think Google's in the business of doing
00:11:24
or shady things. I don't they're not
00:11:26
that company. There are other companies
00:11:28
that would
00:11:31
referring to your alma mada.
00:11:32
>> I'm not going to say which ones. Okay.
00:11:35
But Google is not that company. So I I
00:11:37
think that the reason they did it was
00:11:38
probably because there's user utility.
00:11:40
And my point is we should focus on the
00:11:42
user utility because I think that's the
00:11:43
story worth telling from now on because
00:11:44
I think we collectively, the four of us
00:11:47
can responsibly tell both sides of the
00:11:49
story in a well balanced way because I
00:11:51
think nobody wins if we become bloodites
00:11:54
and go back in time.
00:11:55
>> Yeah.
00:11:55
>> And I think that if we don't if we're
00:11:57
not careful with our words, that's what
00:11:59
will happen.
00:11:59
>> Yeah. And by the way, uh obviously not a
00:12:02
lite, but this is what you know has been
00:12:06
reported by a lot of folks that people
00:12:07
were
00:12:10
surprised, shocked when they saw the
00:12:12
size of the model being done in the
00:12:14
background and it has triggered some uh
00:12:17
people looking at it around privacy and
00:12:19
I do agree that Google is not a bad
00:12:23
actor in the space. So probably a speed
00:12:25
arrow more than anything I would say.
00:12:28
>> Maybe just add two things. Um there's
00:12:31
attorney maxing and then there's
00:12:33
attorney maxed and Google is probably
00:12:36
attorney maxed.
00:12:37
>> Yes.
00:12:38
>> Yeah.
00:12:38
>> And have been for a long time. And the
00:12:40
second second thing I would say is I do
00:12:42
think it's incumbent on all of us as
00:12:44
Americans who are involved in the
00:12:46
technology industry in one way or
00:12:47
another
00:12:49
>> to be advocates for a the positive
00:12:52
optimistic possibilities that AI
00:12:55
introduces to to everyone in this world
00:12:58
because it it is starting to feel or
00:13:01
seem like there may be a CCP funded
00:13:05
campaign against AI and data centers in
00:13:09
America.
00:13:10
And that's very logical for China, but
00:13:13
it is not good for America. And so I
00:13:17
just I think it's it we all have
00:13:20
responsibility is what I would say.
00:13:22
>> Yeah. Who do you think's doing a poor
00:13:24
job at that and responsible for this? Is
00:13:26
it Daario with his constant, hey,
00:13:29
everybody's going to lose their job?
00:13:31
Who's responsible for this? Is it the
00:13:33
CEOs blaming AI for their layoffs?
00:13:37
What's your take on this, Gavin? Look at
00:13:38
hold on a second. Everybody is trading
00:13:41
their own book. It makes enormous sense
00:13:44
for Daario to try to create the boundary
00:13:46
conditions for a regulatory moat because
00:13:49
he will be inside of the tent pissing
00:13:51
out. He's big enough now. And if you
00:13:55
notice that a lot of the breathlessness
00:13:58
has ramped up and Jason, we've talked
00:14:00
about this. You can annotate successive
00:14:03
rounds of fundraising and successive
00:14:05
scale with the volume. So, I think that
00:14:08
it's a reasonable business strategy and
00:14:10
I think that he's quite clever and I
00:14:12
think that look, if you actually and I
00:14:14
and I do this, if you actually just have
00:14:16
a Nashbot, a Nash agent inside of Quad
00:14:18
and you ask it what it would do, it
00:14:20
would come up with this strategy. And
00:14:22
meanwhile, there are other versions of
00:14:23
other counter strategies and
00:14:25
counterexloitative strategies. The point
00:14:27
is that each CEO has a clear incentive.
00:14:30
They're operating at such a level of
00:14:32
scale that they're they're just reading
00:14:35
their own book.
00:14:37
>> So it's it's up to us to take a step
00:14:39
back and actually see the forest from
00:14:41
the trees. I think Nick, can you find
00:14:43
this? There was a clip of Sham Sankar,
00:14:45
friend of the pod, fabulous guy, the CTO
00:14:47
of Palunteer, and he was I think he was
00:14:48
on Fox News
00:14:50
and he said, "Stop breathlessly asking
00:14:53
these model makers what they think.
00:14:56
go to the end user and ask the person in
00:14:58
the factory that's using the model and
00:15:00
ask him what he or she thinks. Ask what
00:15:02
the doctor thinks, ask what the
00:15:04
scientist thinks and start to tell those
00:15:06
stories. That's what we should be
00:15:08
talking about.
00:15:08
>> Yeah. And Gavin, you were I was sort of
00:15:11
asking you your opinion on what's what's
00:15:14
who's causing this and then what's the
00:15:16
solution? Like do you do you have folks
00:15:18
you think in the industry who are
00:15:20
representing it particularly well? we
00:15:22
can point out, hey, Elon has said we're
00:15:25
going to move to a world of incredible
00:15:26
abundance and working will be optional.
00:15:29
I think that's on the margin a little
00:15:30
scary for people to hear because they
00:15:32
hear no job. But he does say, hey,
00:15:34
universal basic income is probably going
00:15:36
to have to come into, you know, place.
00:15:39
He said that multiple times. You have
00:15:40
Daario, according to Chimath talking his
00:15:43
own book, scaring the be Jesus out of
00:15:45
people in order to get regulatory
00:15:47
capture. What do you think is going on
00:15:50
here and how can we do better as an
00:15:52
industry?
00:15:52
>> I think Jimoth outlined like a very
00:15:56
viable and positive path forward. We're
00:15:59
just, you know, real people who are not,
00:16:02
you know, at the tip of the spear. These
00:16:04
are the positive impacts AI's had on my
00:16:06
life. I was at a I was at I was at an
00:16:09
event maybe 10 days ago and someone who
00:16:12
runs a hedge fund, his daughter was born
00:16:14
with a very rare genetic mutation
00:16:18
that effectively would have normally
00:16:20
condemned her to a life devoid of joy,
00:16:23
meaning everything
00:16:26
the neurons in her brain were not
00:16:28
firing. So she wouldn't, you know, who
00:16:31
knows what her life expectancy would
00:16:32
have been or what the quality of her
00:16:33
life would would have been. and it's
00:16:36
it's a tragic disease.
00:16:38
He said he didn't accept that as an
00:16:40
answer. He found he did an enormous
00:16:43
amount of research with LLMs and found
00:16:45
an existing safe drug on the market that
00:16:48
they thought would have a meaningful
00:16:50
impact on his daughter's condition and
00:16:54
it did. It took I think the percentage
00:16:56
of times the neurods were firing was 30
00:16:58
or 40% and it took it up to 80 or 90%.
00:17:02
And that means that she can live a
00:17:04
normal life. She may not be as smart as
00:17:06
she would have been, but she can live a
00:17:08
normal life. And he's now figured out
00:17:12
how to use AI, how to to further tailor
00:17:15
that drug. And you know, there've been
00:17:16
all sorts of advances in protein design,
00:17:20
etc., etc. And he's reasonably confident
00:17:23
he's going to have a drug in months that
00:17:25
is a complete cure. And that's just one
00:17:29
person, one dad who was unwilling to
00:17:32
accept defeat for his daughter and who
00:17:35
changed her life and the life of
00:17:36
everyone else with that disease. And we
00:17:39
tell those stories. So I think Elon's
00:17:40
doing a good job. A future where work is
00:17:42
optional. I think that sounds great to
00:17:44
some people, you know, scary to others.
00:17:47
You know, a 4-day work week, you know, I
00:17:49
think is is probably something that
00:17:50
sounds sounds good to a lot of people. I
00:17:52
think Jensen is doing a good job of
00:17:54
being an effective advocate. And I do
00:17:57
think anyone who is trying to drive
00:17:59
reggget
00:18:01
I just we need to stay focused on the
00:18:03
positives as well.
00:18:04
>> Yeah Freeberg what's your take here on
00:18:08
the AI
00:18:10
PR crisis if we'll call it that? Uh we
00:18:13
had three different commencement
00:18:15
speeches that were booed.
00:18:18
Eric Schmidt being one of them, two
00:18:20
other ones by maybe less notable folks.
00:18:23
when you hear young people booing AI
00:18:26
viferously, why are they doing that,
00:18:28
Freeberg? And what's your take on the
00:18:30
overall PR problem and how to turn it
00:18:32
around?
00:18:35
>> Uh, that's a
00:18:38
there's a long answer to that question.
00:18:40
Um,
00:18:42
>> it relates in some ways to your concerns
00:18:44
about socialism and polarization.
00:18:46
>> The long answer.
00:18:47
>> What's the long answer?
00:18:48
>> I mean, that's like
00:18:50
like why do people hate technology?
00:18:54
the greatest technological why do they
00:18:57
>> this technology they love their phones
00:18:58
they love the internet this technology
00:19:00
they
00:19:04
>> I think that there's like an underlying
00:19:07
view that technology creates leverage
00:19:09
for a small group of people which
00:19:12
creates power imbalances and nothing
00:19:16
represents that more than AI that a
00:19:19
small number of people that control and
00:19:23
profit from and benefit from AI are
00:19:26
going to end up getting outsides returns
00:19:30
relative to the broader population that
00:19:32
the time to diffusion of the technology
00:19:35
because ultimately all technologies like
00:19:37
commoditize and diffuse but the time to
00:19:40
diffusion here is such that it's going
00:19:43
to be like extremely asymmetric for
00:19:45
society and I think that there is
00:19:47
something fundamental about that it's
00:19:49
like you know nuclear bombs I think
00:19:51
really created this this moment in
00:19:53
people's minds in the mid 20th century
00:19:56
that by the back half of the 20th
00:19:57
century gave everyone a high degree of
00:19:59
skepticism about technology and science
00:20:01
generally that those who have the
00:20:03
knowledge and those who engineer
00:20:05
solutions with the knowledge can create
00:20:08
outsized advantages for themselves and
00:20:10
it puts the rest of us at risk the rest
00:20:11
of the world the rest of the population
00:20:13
at risk and because those questions
00:20:15
about when does this benefit me how does
00:20:18
it benefit me can't be answered today
00:20:21
the economic benefit that's acrewing to
00:20:23
the few today becomes the narrative. It
00:20:26
becomes the story and it becomes this
00:20:28
like power system that a few people take
00:20:31
from the many. And so there's something
00:20:33
deeply disturbing for the average person
00:20:36
about that. They don't understand how it
00:20:38
works, why it works, what it'll do for
00:20:39
them, when it will do it, and all that
00:20:41
they're being told is that some people
00:20:42
are making trillions of dollars. So I
00:20:45
think that it's pretty obvious why this
00:20:46
has got such a backlash. Secondly, I
00:20:49
think that there's a deep amount of
00:20:51
external energy that's fueling this
00:20:53
anti-technology sentiment in the United
00:20:55
States and has been for decades. I think
00:20:57
to Gavin's point, I don't think it's
00:20:58
just China with NOS's today. I think
00:21:00
that there is a long history
00:21:03
of state actors intervening in media
00:21:07
activities in foreign nations to try and
00:21:10
create the sentiment and fuel a
00:21:13
sentiment that reduces progress in that
00:21:16
competitive state. I think this goes all
00:21:19
the way back to KGB design during the
00:21:22
Cold War and it's been refined and honed
00:21:25
and improved over time. This is not just
00:21:27
some conspiracy theory. There are plenty
00:21:28
of great books about this. The
00:21:30
techniques of what's going on
00:21:31
specifically today, I don't know enough.
00:21:33
I don't have any great details on that.
00:21:35
But I don't think that there's no
00:21:36
foreign interest in seeing
00:21:39
technology advancement slow in
00:21:42
competitive nations. The United States
00:21:44
probably does similar things to other
00:21:46
nations. And I think that that's
00:21:47
probably a key part of this. And then I
00:21:50
think this this like third piece is like
00:21:53
when the capernac revolution happened it
00:21:56
was a mind you know like helioentricity
00:21:59
was a totally new way of thinking for
00:22:01
humans and it was uh deeply disruptive
00:22:03
to the church and it was deeply
00:22:05
disruptive to the power centers which
00:22:07
were the centers that could tell people
00:22:11
earth is at the center of the universe.
00:22:13
We're in control. We're the direct
00:22:14
channel to God. And the idea that the
00:22:16
sun is at the center of the solar system
00:22:18
and we spin around it and we're a tiny
00:22:19
speck in the universe was very hard for
00:22:22
people to grasp. There's something about
00:22:24
AI that's very like not humanentric and
00:22:27
it kind of shifts and fs with the ego of
00:22:30
the human. It's it's almost
00:22:31
anti-humanist
00:22:33
and I think that that's like a deep
00:22:35
psychological current a lot of people
00:22:37
and their disdain for this technology.
00:22:39
It fuels it. It's not the cause but I
00:22:41
think it fuels it. So I think there's a
00:22:43
lot of complicated aspects to this Jal.
00:22:45
You know, I don't think there's like a
00:22:46
simple put put Sham on on a podcast and
00:22:48
he'll solve the problems with AI right
00:22:50
now. I think that there's a real set of
00:22:52
shifts happening
00:22:53
>> and there's a real set of global
00:22:55
competition underway
00:22:57
where you know various state actors and
00:22:59
interests are competing with each other
00:23:02
>> do you think that we should slow down?
00:23:06
>> I don't think you can.
00:23:07
>> No, no, no. Do you think we should slow
00:23:09
down?
00:23:10
>> No. I I think I was just talking to some
00:23:12
people on a Zoom right before this, but
00:23:13
I think after the Manhattan project, the
00:23:15
research labs were stood up to maintain
00:23:17
our scientists that worked on the the
00:23:19
Manhattan project from effectively
00:23:22
leeching back or leaking back to Russia
00:23:25
and Germany and and and other places
00:23:26
that that were adversary to the United
00:23:28
States.
00:23:30
And they all were against the nuclear
00:23:32
bomb. They worked on it because it was
00:23:34
necessary for the United States
00:23:36
security. But then when Russia got a
00:23:38
hold of the secrets, they were leaked
00:23:39
because people were worried that if the
00:23:40
US had all the power, there would be no
00:23:43
counterbalance to the power. And so the
00:23:46
the nuclear secrets were leaked to
00:23:47
Russia for that purpose. Then when
00:23:49
Russia had the nuclear secrets and they
00:23:51
began developing hydrogen fusion bombs
00:23:55
and it it was clear that they were going
00:23:56
to race ahead, the United States raced
00:23:58
ahead with developing nuclear bombs as a
00:24:00
counterbalance to Russia. When the
00:24:02
proliferation began, there was no
00:24:04
stopping it. it began and you have to
00:24:06
have this balance in the world otherwise
00:24:08
you have effectively an asymmetric power
00:24:11
that can do whatever it wants globally.
00:24:14
I think there's that moment in the world
00:24:15
right now where if the United States
00:24:18
does not advance its AI technology, the
00:24:22
availability of it, TBD industry,
00:24:24
taxation, all these all these things
00:24:26
that we're talking about doing, there
00:24:27
will be someone else that will. And if
00:24:30
someone else does, we can go through
00:24:32
what would happen. There's a there's a
00:24:34
complicated game theory on this, but
00:24:36
what would happen if China had
00:24:38
sufficiently advanced models and
00:24:39
sufficiently advanced scaled deployment
00:24:42
of those models relative to the United
00:24:43
States? As you do that analysis, you
00:24:45
realize, wait a second, that's probably
00:24:47
not a healthy place for the world to be.
00:24:49
It's also probably not a healthy place
00:24:50
for the United States to be the only one
00:24:52
with AI. And so I I think what we end up
00:24:55
seeing
00:24:57
is if we do try and slow down AI, we
00:25:00
kind of lose this moment of balance
00:25:02
that's necessary when you have a
00:25:04
technology proliferation like we saw
00:25:06
with the arms race after World War II
00:25:08
that, you know, we're going to see again
00:25:10
here.
00:25:11
>> Chimath, you I think bring up a good
00:25:13
point. You know, should we slow it down
00:25:15
or could we slow it down? There actually
00:25:17
have been some discussions about ways to
00:25:20
do this. One of them would be, hey, with
00:25:22
self-driving, people are scared that all
00:25:24
these cab drivers are going to lose
00:25:26
their jobs. Uber drivers, cab drivers,
00:25:28
bus drivers, truck drivers. This is, you
00:25:31
know, over 10 million people in the
00:25:32
United States driving things for a
00:25:34
living. Would you be in favor of some of
00:25:37
the announcements that that will be a
00:25:41
paced roll out? It won't happen all at
00:25:43
once. In other words, those people will
00:25:44
be giving some amount of job security to
00:25:46
stay behind the wheel with it. Another
00:25:48
example that's been given is if you put
00:25:51
Optimus into Amazon factories or the
00:25:53
figure robot just did like a week of
00:25:55
just sorting packages. I'm sure
00:25:57
everybody saw that video. We'll insert
00:25:58
it here. That figure robot sorting
00:26:01
things. Hey, if Amazon deploys those,
00:26:02
there'll be a tax on those per hour and
00:26:05
we'll tax humanoid robots in some ways
00:26:07
and then use that for say retraining
00:26:10
people. Those are two very specific
00:26:12
conditions and approaches that people
00:26:14
have been promoting. Do either of those
00:26:16
resonate with you in any way?
00:26:19
I think it's
00:26:21
interesting that in all of those
00:26:22
discussions, I've yet to see an actual
00:26:25
survey of only the truck drivers and
00:26:27
only the package sorters.
00:26:30
>> The question that I would have is, do
00:26:32
the people that do these jobs want these
00:26:33
jobs? And if they do, then there's a
00:26:36
reasonable claim to make to keep those
00:26:38
jobs the way that they are. If you're
00:26:39
saying this is the job that I do, I love
00:26:41
it. I'm able to provide for my family.
00:26:43
Great. That's a very different argument
00:26:45
than well you know what Amazon has 35 or
00:26:48
40% churn inside of their warehouses and
00:26:52
we should probably ask the question why
00:26:54
is that because if it was such a great
00:26:56
job I suspect the churn would be 3 or
00:26:57
4%. So what exactly is it that we want
00:27:02
to protect and have you asked them? And
00:27:05
I think that this is just again a bunch
00:27:07
of people in the peanut gallery who want
00:27:10
to take a moral high ground and try to
00:27:12
make some other group of people feel
00:27:14
guilty or feel bad. At no point are we
00:27:17
actually asking the conversation that
00:27:20
that we should be having, which is it's
00:27:23
interesting to me that the there was
00:27:24
supposed to be an EO an exe a
00:27:26
presidential executive order that was
00:27:27
announced today and then it was pulled.
00:27:29
It was scrubbed at the last minute.
00:27:31
>> Did you guys notice that? And yesterday
00:27:33
what was leaked was everybody that was
00:27:35
attending. It was all the big Neolabs
00:27:37
CEOs and it was all the big hyperscaler
00:27:39
CEOs
00:27:40
>> including friend of the pod Nick Aurora.
00:27:43
Shout out to Nick. And then it was
00:27:45
scrubbed an hour ago. Why was it
00:27:48
scrubbed? And the president said that
00:27:50
there were aspects of the bill that he
00:27:52
didn't agree with. And as far as we can
00:27:54
tell, the aspects would have required
00:27:56
some amount of supervision, insight,
00:27:59
review from the federal government
00:28:01
>> of language models specifically of these
00:28:03
frontier models is what I read.
00:28:05
>> Not language models because I think just
00:28:07
of AI because there's going to be many
00:28:08
different kinds. They're not always
00:28:09
going to be language models, but of AI.
00:28:11
So look, I think Freeberg is is right.
00:28:14
We are in a proliferation with China. I
00:28:17
think it's actually good that China is
00:28:18
less than nine months behind us.
00:28:21
I think it allows us to find a detant
00:28:24
where we have a certain magnitude of
00:28:26
capability that they also have and that
00:28:30
allows all of us to then seek peace and
00:28:32
abundance and the fact that we are
00:28:34
orthogonal societies, we are organized
00:28:36
differently increases the probability of
00:28:38
finding peace using the Renee Gerard
00:28:40
kind of framework of mimetic theory than
00:28:43
if if it was like us and another country
00:28:45
that was exactly similar to us.
00:28:48
So I think what we need to do, we
00:28:51
probably need KYC. I think that that
00:28:54
should be something that us in China get
00:28:55
together and say you don't want it to
00:28:58
get into the hands of people you can't
00:28:59
control. You probably already KYC those
00:29:01
models anyway inside of China. You
00:29:03
already review those training runs
00:29:05
before you allow these models to get
00:29:06
released. We already know that that's
00:29:07
happening.
00:29:08
>> Yes.
00:29:08
>> So we should probably do some sort of
00:29:10
KYC so some crazy person doesn't create
00:29:12
some biological weapon. I think that
00:29:14
those are like some reasonable ground
00:29:15
rules. But otherwise, Freeberg is right.
00:29:17
You have to take a a little bit of a
00:29:19
deterministic view here, which is that
00:29:21
we are in this existential race and we
00:29:24
need to get to the place where each of
00:29:26
us, meaning us in China, can look each
00:29:28
other in the eye and say, "All right,
00:29:29
weapons down, so to speak."
00:29:31
>> Gavin, I'm going to hold you to answer
00:29:33
two questions. One, should we run
00:29:37
Frontier models? Because that's
00:29:39
specifically what was mentioned in the
00:29:40
leak about the EO Frontier models, the
00:29:42
powerful ones. Should they be run
00:29:44
through some sort of testing before
00:29:45
they're released? And should there be
00:29:46
some regulatory framework for that?
00:29:48
That's my first question to you. Yes or
00:29:50
no question. And then you can explain
00:29:52
your answer.
00:29:55
>> Jeez. Like I just think it's such a
00:29:57
complicated topic. It feels
00:30:01
we're a little early for that. I don't
00:30:04
love the idea of the United States um
00:30:08
doing it and no one else doing it. I
00:30:13
like I think in a world where we hold
00:30:15
hands with China
00:30:17
I think that's that's much more
00:30:19
palatable and we are aligned and we
00:30:21
trust each other and have kind of
00:30:23
verification
00:30:25
uh capabilities I do think yeah let me
00:30:28
then rephrase it should China and the US
00:30:31
come up with a simple battery of things
00:30:32
that have to be tested before these go
00:30:36
out including boweapons
00:30:38
terrorism and and that genre of in that
00:30:42
vertical of just really known dangerous
00:30:44
things just like the FDA might test for
00:30:46
poisons or contaminants in a food or a
00:30:48
drug. Would you be in favor of that? I'm
00:30:50
curious.
00:30:52
>> So, a few things like the one one thing
00:30:54
that's great about America is there is
00:30:57
or one thing that's just we have other
00:31:01
forms of regulation.
00:31:03
>> Self-regulation. Sure. We have self one
00:31:05
we have self-regulation also we have the
00:31:07
courts and if an AI model company
00:31:11
behaves irresponsibly
00:31:13
they know that there are ways that
00:31:16
people who have been harmed can seek
00:31:18
recourse
00:31:19
and so we already have a system that
00:31:22
encourages responsible behavior on the
00:31:26
part of the model makers. That's a great
00:31:28
point because OpenAI is being sued right
00:31:30
now by a kid who who killed themselves
00:31:32
after talking to OpenAI's model. So
00:31:34
you're you're correct in that. Yes.
00:31:35
After the fact. Yeah.
00:31:36
>> And we will see we'll soon we'll see
00:31:37
more of that. I just you know to me once
00:31:41
you give something give a power to the
00:31:43
government.
00:31:45
>> It's almost never taken back and it
00:31:47
tends to grow
00:31:48
>> and it's kind of a one-way a one-way
00:31:51
path.
00:31:51
>> Oneway ratchet. Yeah.
00:31:54
>> Yeah. Second question then. Chimath was
00:31:56
saying, "Hey, nobody listens to the, you
00:31:58
know, these cab drivers or maybe the um
00:32:00
people sorting the packages. Do they
00:32:02
want the jobs or not?" Actually, the UK,
00:32:04
there was just a 60-minute special and
00:32:06
UK and also Boston and New York are
00:32:09
pretty adamant that they want humans to
00:32:11
stay and they want to ban self-driving
00:32:13
in those locations or severely limited
00:32:15
or maybe limit it in some way to let
00:32:18
those people keep their jobs. How do you
00:32:20
feel about that possibility? Is that
00:32:22
something you think society should be
00:32:24
open to some gradual licensing?
00:32:27
>> They're going to get sued for wrongful
00:32:28
death. When somebody runs over somebody
00:32:30
else and you could have implemented a
00:32:32
solution that has a zero death rate,
00:32:33
that's very different from package
00:32:35
sorting.
00:32:36
>> Okay.
00:32:36
>> Go talk to the package sorters is what I
00:32:38
say. Go talk to the people inside the
00:32:39
Amazon warehouse. Ask them what they
00:32:41
would rather do at Amazon. Ask them.
00:32:44
>> Yeah. Sure. But Gavin, what are what are
00:32:46
your thoughts here on either one of
00:32:47
those examples here? I think going to a
00:32:49
city where you can't get in a Whimo or a
00:32:51
cyber cab is going to feel barbaric and
00:32:54
unsafe until you
00:32:56
>> agree. Strongly agree.
00:32:57
>> I don't know if you remember, but the
00:32:58
early days of Uber, sometimes you go to
00:33:01
a city where there was no Uber.
00:33:02
>> Yeah. It'd be incredibly frustrating.
00:33:04
>> Well, I'm not going to come back until
00:33:06
they have Uber. It's so inconvenient.
00:33:08
And I think so whatever individual
00:33:11
municipalities decide, I do think one
00:33:14
Chimas's point is really powerful.
00:33:16
There's 50,000 automotive deaths per
00:33:19
year in the United States if I recall
00:33:20
correctly and a million globally and you
00:33:24
know that's not tolerable and there will
00:33:26
for sure be wrongful death lawsuits and
00:33:28
then just from a convenience and quality
00:33:30
of life perspective I just don't think
00:33:32
it's going to persist and that's another
00:33:35
great thing about America is you know
00:33:37
you have this patchwork of different
00:33:38
states and municipalities and each one
00:33:41
doing things in a different way and I'm
00:33:43
not suggesting that's good for AI
00:33:46
But it does tend to you know has
00:33:50
historically the curly effect aside led
00:33:53
to you know I think more positive
00:33:55
outcomes where cities and states compete
00:33:57
you the curly effect being it is
00:34:00
>> yeah this is a really important point
00:34:02
you're making Gavin with flock safety as
00:34:05
but one example we had a an AI there's
00:34:07
an AI tool called flax safety it's
00:34:09
cameras that use AI monitor people who
00:34:12
are committing crimes there's a privacy
00:34:13
issue around it is bottom up. You just
00:34:15
do it by town. It's not top down. And
00:34:17
states can regulate it. Same thing will
00:34:19
probably happen with self-driving. And
00:34:21
states will probably have some say in
00:34:23
how AI is deployed, even if maybe some
00:34:26
centralized governments don't want to do
00:34:28
that. I really think this only comes
00:34:30
down to
00:34:31
>> the block safety thing. I think it's so
00:34:32
good, Jason. Crime is now a choice.
00:34:34
>> Yeah.
00:34:35
>> You know, I think the the Cambridge City
00:34:37
Council voted
00:34:38
>> to turn off gunshot detectors two days
00:34:41
ago.
00:34:42
>> And
00:34:42
>> wait, which city did that? What you mean
00:34:44
about Harvard?
00:34:45
>> Cambridge Cambridge Cambridge. That's
00:34:47
the place where Harvard is.
00:34:48
>> That's a place where Harvard is.
00:34:49
>> So the geniuses coming out of Harvard in
00:34:51
that town decided gunshot detection
00:34:56
>> shouldn't occur. We you don't want
00:34:58
gunshot detection.
00:35:00
>> It's wild because you know there's
00:35:01
there's a theory that it disadvantages
00:35:03
you know that it might lead to an
00:35:05
illegal migrant shooting a gun being
00:35:08
apprehended and we don't want that.
00:35:10
>> Got it. and A16Z had a great essay on
00:35:12
flock.
00:35:14
We can really really solve crime and
00:35:19
it's just a choice and different states
00:35:21
and municipalities will make different
00:35:22
choices to be pro-rime or anti-rime and
00:35:25
I'm sure they don't cast it as pro-
00:35:27
crime. There's an, you know, some sort
00:35:30
of, you know, moral or ethical reason
00:35:32
they're making that choice. But people
00:35:34
will vote with their feet over time and
00:35:36
then voters will vote with their votes
00:35:39
and we'll see what works.
00:35:40
>> Have you guys been to Vegas recently? My
00:35:42
wife and I went to visit Vegas and we
00:35:44
spent the the afternoon with Ben
00:35:46
Horowitz and his wife Felicia. She has
00:35:49
done this incredible job with the Las
00:35:50
Vegas Police Department. It is one of
00:35:53
the most impressive things I've ever
00:35:54
seen. And to your point,
00:35:56
crime is an option and they've said no.
00:35:59
So what happens there is they have
00:36:00
gunshot detection. They have drones that
00:36:02
get deployed off the roof of the police
00:36:04
building. We were sitting inside of
00:36:06
mission control where you see it
00:36:07
happening, Jason. If something happens,
00:36:10
they have eyes on site within minutes.
00:36:12
They can track offenders and bad guys
00:36:15
all the way to wherever they're hiding.
00:36:17
And you walk out of it and you feel
00:36:19
incredibly safe, like they're really on
00:36:21
top of it. And when you understand the
00:36:23
level of investment, it's not it doesn't
00:36:25
take billions of dollars.
00:36:26
>> It's dimminimous compared to the cost.
00:36:29
It's dimminimous. It's
00:36:30
>> especially when compared to the cost of
00:36:32
the crime occurring.
00:36:33
>> Exactly. If you gave the Las Vegas
00:36:35
Police Department
00:36:37
30 $40 million a year, it would be the
00:36:39
safest city in America and that's all it
00:36:41
would take.
00:36:41
>> Yep. Exactly. And the for the privacy
00:36:44
concerns, Freeberg, there are very
00:36:45
simple solutions to this. I I am a
00:36:47
privacy advocate myself. Of course, we
00:36:49
all want some level of privacy. I I had
00:36:51
the Flock CEO in this week in startups
00:36:53
twice in the past 10 years. He's very
00:36:55
considered. And the way they do it with
00:36:57
Flock is they allow you to have a
00:36:59
rolling database and I think there's a
00:37:02
maximum you can save the license plates
00:37:04
for and they don't do facial
00:37:05
recognition. I I don't see why not but
00:37:07
let's put that aside. You can only keep
00:37:09
it for 2 or 3 years and then they insist
00:37:11
on having an audit trail in it. So there
00:37:13
are all little things you can do on the
00:37:14
back end to protect privacy with audit
00:37:16
trails etc. We got a lot more to get to
00:37:17
in the docket. I just want to just give
00:37:19
my final thoughts on what we're talking
00:37:21
about here in terms of the AI problem
00:37:24
and the PR problem. I think we have to
00:37:27
recognize that the layoffs that are
00:37:29
occurring in big tech and in a lot of
00:37:32
these places are not just the bloating
00:37:35
issue anymore. And I'm just going to
00:37:36
point to two factors that I think are
00:37:38
scaring the be Jesus out of people. And
00:37:40
we just have to admit that this is
00:37:42
occurring as opposed to we've been
00:37:43
debating it here. Is it occurring? Is
00:37:45
this just cover? and are we AI washing?
00:37:47
The first one I want to give you an
00:37:48
example of is Matthew Prince who's the
00:37:51
CEO of Cloudflare. Incredible company,
00:37:53
public company. Two weeks ago, I laid
00:37:55
off more than 20% of my workforce. I
00:37:57
didn't do it because CloudFare is
00:37:58
struggling. We posted record revenue
00:38:00
growth, have strong free cash flow, and
00:38:02
are adding an unprecedented number of
00:38:04
customers, yada yada yada. And he says
00:38:06
basically he's getting rid of measurers.
00:38:09
Measurers are the people who manage
00:38:11
people and who measure data. and he just
00:38:14
says, "We're getting rid of all those
00:38:15
people. They're unnecessary because of
00:38:17
AI, and we'll be adding to people in
00:38:19
other positions." At the same time,
00:38:21
Zuckerberg did another round of layoffs,
00:38:24
and they were done in a way that people
00:38:26
felt was not considered
00:38:28
and a bit um what's the word?
00:38:34
>> Dystopian.
00:38:35
>> Dystopian. Thank you, sir. Uh he did him
00:38:37
in a in a pretty dystopian way. Here's
00:38:40
Zuckerberg for 30 seconds. In general,
00:38:43
the average intelligence of the people
00:38:45
who are at this company is significantly
00:38:47
higher than the average set of people
00:38:49
that you can get to do tasks if you're
00:38:51
working through the contract um through
00:38:53
through these contractors. So if we're
00:38:56
trying to teach the models coding for
00:38:58
example, then having people internally u
00:39:01
build tools that or or solve tasks that
00:39:05
um that help teach the model how to code
00:39:07
we think is going to dramatically
00:39:09
increase our model's coding ability
00:39:10
faster than what others in the industry
00:39:12
have the capability to do who don't have
00:39:14
thousands and thousands of extremely
00:39:16
strong engineers at their company.
00:39:18
>> Okay. So what Zuckerberg did at the same
00:39:21
time concurrently he told everybody
00:39:23
we're laying off these 8,000 people. A
00:39:24
lot of those people are incredibly
00:39:26
talented.
00:39:27
Some of them are on H-1B visas. Creates
00:39:29
all kinds of chaos for them in their
00:39:30
personal lives and obviously they're
00:39:31
having record profits there as well.
00:39:34
At the same time he was laying off those
00:39:36
8,000 people. This is after tens of
00:39:38
thousands of layoffs before which were
00:39:40
obviously because of bloating. He said
00:39:41
we're putting recording software on
00:39:43
every single person in the company's
00:39:44
computers to study and train our model.
00:39:47
And people were like, "Oh," and previous
00:39:49
people said, "I built during the AI
00:39:52
hackathons they had months ago, I built
00:39:55
all this AI tools to make my job uh more
00:39:58
efficient." And then Zuckerberg laid me
00:40:00
off. So the now p the perception people
00:40:03
have now, and it's quite correct, I
00:40:05
think, is the most you can hope for here
00:40:07
is you keep this job for some amount of
00:40:10
time and train your way out of it, and
00:40:11
hopefully there's some more work for
00:40:12
you, but they're studying you. And
00:40:14
Zuckerberg just said it plainly there.
00:40:16
Hey, we're going to study everybody here
00:40:17
and that's going to lead to more
00:40:19
replacements. This is scaring the Jesus
00:40:20
out of people and we need to have an
00:40:22
answer for it. Yeah.
00:40:23
>> I thought the Matthew Prince note was
00:40:26
horrible.
00:40:27
>> Okay, explain.
00:40:28
>> This was like from the PR school of
00:40:31
retards.
00:40:32
>> Okay, here we go.
00:40:34
>> You could not have written a worse memo.
00:40:36
It's like you reduce humans to a label
00:40:40
called the measurer and then you're
00:40:42
like, I'm going to lay off all the
00:40:43
measurers.
00:40:45
I mean,
00:40:47
I just think that part of this, again,
00:40:49
I'll go back to the and maybe the Sham
00:40:52
Sankar quote that I'm thinking about
00:40:54
should be extended beyond the model
00:40:55
makers. Can you just play this for one
00:40:58
second and I'll tell you why I think
00:40:59
this is too
00:41:00
>> and then we'll go on to the S1 from
00:41:02
SpaceX.
00:41:02
>> We're listening too much to the
00:41:04
inventors of AI. I know that's
00:41:05
appealing. They're geniuses. They're
00:41:07
smart. We need to be listening to the
00:41:08
frontline factory workers who are using
00:41:10
AI saying, "Wow, I was able to add a
00:41:12
third shift. I was able to hire more
00:41:14
workers or the ICU nurse who says I have
00:41:16
more time to spend with my patients. I'm
00:41:17
able to ensure they don't code during a
00:41:19
shift change.
00:41:20
>> So look, my my point is like the first
00:41:22
part of what he said applies here, which
00:41:24
is who cares what Matthew Prince thinks.
00:41:26
Because the reality is that if this is
00:41:28
the way that you're going to message
00:41:30
something as critical as this, I think
00:41:32
you did a horrible job. And now you
00:41:33
label these people and you put a scarlet
00:41:35
letter on their back. So now when they
00:41:36
try to get a different job, they're
00:41:37
like, "Oh, you're one of the Cloudflare
00:41:39
measurers." How does that help anybody?
00:41:41
It didn't needed to be done this way.
00:41:43
There's enough of these tech CEOs that
00:41:46
are now public. You can hear them. You
00:41:48
can understand them. And I think what
00:41:51
we're learning is, man, they're really
00:41:53
good at one thing and they're not
00:41:54
necessarily as good at all the other
00:41:55
things.
00:41:56
>> Yeah. Uh, okay.
00:41:57
>> And so I would say shut the up,
00:42:01
get behind the keyboard, just do your
00:42:03
job. And if you need to manage
00:42:05
something, just manage it, but don't
00:42:06
write these misses. You're terrible at
00:42:08
it. All of you. You're all terrible. You
00:42:11
suck at this.
00:42:13
>> All right. Uh,
00:42:14
>> end of my TED talk. Thank you for coming
00:42:16
to my TED talk.
00:42:16
>> Thanks for coming to Shamat's 18minute
00:42:18
TED talk and we'll we'll uh keep moving
00:42:21
on.
00:42:21
>> Sorry. And sorry, when everybody gets
00:42:23
upset, this will be why.
00:42:25
>> Yes. I mean, I I do think the Zuckerberg
00:42:28
this and Jack saying, you know, hey,
00:42:29
we're going to have half as many people.
00:42:30
Everybody reports directly to me. This
00:42:32
is all building this fear in society.
00:42:34
And I think people are rightfully
00:42:36
scared. If the people building it tell
00:42:38
you, be scared. your job's going away.
00:42:40
>> Wait till the next
00:42:42
regulatory filing comes out from these
00:42:44
companies and they authorize a massive
00:42:47
share buyback and an increase in their
00:42:49
dividend.
00:42:49
>> Yeah. And their cash pile grows.
00:42:51
>> All I'm saying is there's a right way to
00:42:53
do this, make these decisions, and then
00:42:55
there's a wrong way to do it, which is
00:42:56
to message it in the way that they're
00:42:58
doing it.
00:42:58
>> So, whoever is running PR and comms and
00:43:01
approves and reviews these things
00:43:03
>> are really at their job.
00:43:06
They don't understand the moment.
00:43:08
>> Oh, some breaking news here. It looks
00:43:10
like Anthropic has hired three more
00:43:13
people. Here we go. Let's see.
00:43:14
>> Oh, here we go. Personal job news from
00:43:16
Sam Alman. He'll be joining Anthropic. I
00:43:19
see. That's pretty good.
00:43:20
>> Who else is joining Anthropic? Let's
00:43:21
see. We checked everybody's socials.
00:43:23
>> Oh, Tucker. Tucker Carlson also joining
00:43:27
Anthropic. He'll be doing their PR and
00:43:29
podcast from Anthropic headquarters. And
00:43:31
who else? Oh, Shamas looking good. Well,
00:43:34
it looks like you put 5 lbs on and using
00:43:37
this. First of all, this is not this is
00:43:40
not what I look like. And um
00:43:41
>> did you get 5 lbs?
00:43:44
>> Looks like Are you not wear Are you not
00:43:46
wearing underwear?
00:43:47
>> He's no underwear and he's wearing his
00:43:49
khakis, but I think it's just he's
00:43:50
trying to not show off those scrawny
00:43:52
those scrawny squints. Those those
00:43:54
little slats he calls legs. He's
00:43:56
covering them up now.
00:43:57
>> Uh hold on. I'm gonna send I don't want
00:43:58
to leg shame you.
00:43:59
>> I I I knew that this was going to come
00:44:01
up. I'm going to send Nick an updated
00:44:02
picture of my legs. And you you can deal
00:44:05
with this.
00:44:05
>> You can photoshop your legs all you
00:44:07
want.
00:44:07
>> I didn't I didn't photoshop.
00:44:08
>> You look like an ostrich, brother. Those
00:44:10
are not a day working on it.
00:44:14
>> They may not be photoshopping, but he
00:44:16
may have been leg maxing.
00:44:17
>> He could be legs.
00:44:21
What are you doing?
00:44:22
>> First of all, first of all, I'm 6'2, you
00:44:25
goons. Okay, so
00:44:26
>> gosh legs.
00:44:27
>> All you little people, you know, Jason,
00:44:29
you're 5 foot all. So, you know, your
00:44:31
ability
00:44:31
>> I've had a good day when my platforms.
00:44:33
Yeah.
00:44:33
>> Your ability to have legs is is
00:44:35
different cuz like my muscle mass won't
00:44:37
show on my legs
00:44:39
>> the way it shows on your leg.
00:44:40
>> Oh, I don't think you're helping. I look
00:44:42
at you know what he's doing there,
00:44:43
Freeberg. You see it, right? Look at how
00:44:45
scrawny the calves are. And then look at
00:44:47
how he's doing a
00:44:48
>> That's a push-up bra for your quads.
00:44:50
>> Oh my gosh.
00:44:51
>> That's the equivalent of a push-up bra.
00:44:52
He has those Bonsu balls. He put Bosu
00:44:55
balls under his hammies to You did it.
00:44:58
You're using fillets to
00:44:59
>> I did my legs and I flex. Okay, you know
00:45:01
what? Move on. This
00:45:03
>> I was just say I think we should give
00:45:04
credit where credit is due.
00:45:06
>> Yeah,
00:45:06
>> Jamoth's legs.
00:45:08
>> Better has been doing a lot of work on
00:45:10
his legs.
00:45:12
>> Thank you.
00:45:12
>> It's better. I'll I'll give him better,
00:45:14
but I do think that he's pumping them up
00:45:16
here. Okay, let's keep going here. All
00:45:18
right, topic two. SpaceX just filed
00:45:20
their S1 on Wednesday. They are aiming
00:45:23
to raise 75 billion at a 1.75 trillion
00:45:27
with the T valuation. This will be the
00:45:28
largest IPO ever by more than double
00:45:31
Saudis Ramco $29 billion IPO a couple
00:45:35
years ago. Listing is expected mid June
00:45:38
likely June 12th. Ticker will be SPCX.
00:45:41
We got a lot of interesting information
00:45:43
in the S1 tearown. Obviously SpaceX has
00:45:47
three main business units. Starlink is
00:45:50
the money printer right now, but there
00:45:52
is a second one that's emerging.
00:45:54
Starlink did 11.4 billion in revenue
00:45:56
last year on 50% growth with 4.4 billion
00:46:00
in operating income. Over 10 million
00:46:03
people are now subscribing to Starlink.
00:46:05
That business could easily be hundreds
00:46:08
of millions of paying subscribers. So
00:46:10
that that's uh a lot of growth
00:46:12
potentially there. The space business is
00:46:15
but 4 billion in revenue. It's growing
00:46:17
17% growth, which would still be strong
00:46:18
growth, uh, but had 650 million in
00:46:21
operating losses. AI did 3.2 billion in
00:46:24
revenue. That's more than double
00:46:25
year-over-year growth, but it had 6.4
00:46:27
billion in operating losses.
00:46:30
SpaceX had 20 billion in capex spend
00:46:32
last year. Over 60% was for the AI
00:46:35
compute buildout, and obviously they
00:46:37
were trailing anthropic and open AI and
00:46:39
Gemini in terms of XAI uh, playing
00:46:42
catch-up. and he did a big reboot of
00:46:44
that as we saw on the Twitter. But
00:46:47
here's the big one. EWS, Elon Web
00:46:50
Services as we call it here on the
00:46:51
all-in pod has exploded. Anthropic is
00:46:56
paying SpaceX, wait for it, 1.25 billion
00:46:59
a month to rent out Colossus 1 and parts
00:47:01
of Colossus 2. It's a $45 billion deal
00:47:05
over three years, 15 billion a year. In
00:47:08
other words, they added a Starlink in
00:47:09
terms of revenue to the party. Plus, if
00:47:12
they buy Curser, that's going to add
00:47:14
another two or three billion.
00:47:15
>> Not if I already told you they already
00:47:17
bought.
00:47:17
>> Okay, I'm just trying to, you know, dot
00:47:19
the eyes and T's here. Uh, but when they
00:47:21
buy uh Curser, that adds two or three
00:47:23
billion. That's not in the S1.
00:47:26
>> That's also growing and doubling.
00:47:28
>> Yeah, that's probably growing 2x year
00:47:30
over year. And who knows how much faster
00:47:32
it will grow. Poly Market 71% chance
00:47:35
SpaceX closes its first day of trading
00:47:37
with a market cap above two trillion.
00:47:40
Thank you to our partner Poly Market.
00:47:43
I'll stop here. Gavin, uh you've been
00:47:45
involved in the company for a long time.
00:47:47
Jimoth, you uh I think were a big
00:47:49
investor in the satellite company that
00:47:51
became part of Starlink, which is the
00:47:54
revenue driver there. So, you both have
00:47:56
uh a lot to say about this. Gavin, your
00:47:58
take on the S1 and I think specifically
00:48:00
Elon Web Services. Well, I think what's
00:48:03
important about um Elon Web Services
00:48:06
does make me laugh, but 15 billion that
00:48:09
means the AI business right there is
00:48:11
going to quadruple. It it has already
00:48:13
effectively quadrupled.
00:48:15
I think what's important about that is
00:48:17
there's a stat in it that for I think
00:48:19
the their first data center was 122
00:48:21
days. For the second one, it took them
00:48:24
91 days. The third one was I think 66
00:48:27
days. They build data centers
00:48:31
dramatically faster than anyone else at
00:48:33
a lower cost.
00:48:36
And now that you have a clear offtake
00:48:40
partner and I would expect partner to
00:48:42
become partners,
00:48:44
there is no reason they can't start
00:48:47
stamping these data centers out really
00:48:49
fast.
00:48:51
And having watched Jensen for a long
00:48:53
time, it is important to Jensen that his
00:48:55
GPUs be used. And so GPUs will be
00:48:59
allocated to who can plug them in, turn
00:49:01
them on, and start converting electrons
00:49:04
into tokens.
00:49:06
And so I think this business can grow
00:49:10
dramatically faster than I think, you
00:49:12
know, maybe what anyone could have
00:49:14
contemplated three, you know, three
00:49:16
three months ago. But 15 billion from
00:49:19
anthropic is is extraordinary.
00:49:22
>> Important note, it can be cancelled by
00:49:24
either party with 90 days notice. just
00:49:26
want to make sure we also have that in
00:49:28
there. So that means Elon might want his
00:49:30
compute back or Anthropic may find
00:49:32
another solution. So they do both have
00:49:33
an out
00:49:34
>> and I think that's that's a you know I
00:49:37
think that's that's that's probably an
00:49:38
important provision. Yeah.
00:49:39
>> For everyone but I think the other thing
00:49:42
that came out this week which was not in
00:49:44
the S1 Nick can you throw up the para
00:49:46
frontier and maybe don't you know
00:49:47
include the email and the names and
00:49:49
everything but the composer 2.5 stat. I
00:49:52
think this is really extraordinary. So
00:49:54
cursor's composer 2.5 model came out
00:49:57
this week and I mean this is parto
00:50:01
dominant and this is just you know three
00:50:05
four weeks of doing reinforcement
00:50:07
learning on Colossus 2 with cursor's
00:50:12
data and cursor has we will never know
00:50:16
but but cursor allegedly has more tokens
00:50:20
of coding data than exist on the public
00:50:22
internet And that is a stat from I think
00:50:25
more than a year ago. So I'd imagine
00:50:27
it's grown significantly.
00:50:29
And I think cursor and anthropic
00:50:32
probably have the most proprietary
00:50:35
tokens of coding data. And what this
00:50:38
this jump from composer 2 to composer
00:50:40
2.5 showed us is that when you do an
00:50:44
appropriate amount of reinforcement
00:50:45
learning using Nate data let alone
00:50:49
injecting it into the pre-training of a
00:50:50
new base model because composer 2.5 is
00:50:54
the same base model as composer 2 which
00:50:56
is Kimmy K.25
00:50:59
>> like this is amazing. This is three or
00:51:01
four weeks and it is paro dominant. the
00:51:03
paro front here.
00:51:04
>> If you draw a curve
00:51:07
of the blue dots,
00:51:09
you can see composer 2.5 is literally
00:51:15
well outside the paro frontier. And
00:51:17
that's after 3 weeks. And what's going
00:51:19
to happen next is you're going to have a
00:51:21
new base model with a cursor model in
00:51:24
it.
00:51:25
>> Yeah.
00:51:25
>> Then the cursor model RL using the
00:51:29
biggest coherent uh compute cluster in
00:51:32
the world.
00:51:33
And I think this is
00:51:37
I think this may
00:51:37
>> it's significant. Yeah,
00:51:38
>> it's extremely significant for XAI and
00:51:41
Cursor
00:51:42
>> and Cursor was dead in the water in
00:51:45
terms of access to compute and they were
00:51:48
falling very far behind Codeex,
00:51:51
Google, Anthropic and then Elon let them
00:51:55
on Colossus and boom, instantly their
00:51:58
models are growing faster and this could
00:52:02
be we could be sitting here a year from
00:52:03
now and they're the dominant player and
00:52:07
could we sitting here, Gavin, in a year
00:52:09
and Elon is selling compute to Google
00:52:13
and Open AI. Is that a possibility or
00:52:16
not?
00:52:18
>> Well, I think it's much easier to see
00:52:20
him selling compute to Google and I
00:52:22
think already there have already been
00:52:24
posts about that.
00:52:25
>> Yeah.
00:52:26
>> And for sure Google is going to want to
00:52:28
be part of orbital compute.
00:52:31
>> You know, it's very funny. The only
00:52:33
people who are skeptical of orbital
00:52:35
compute are those people who are not
00:52:37
involved in in data centers or space.
00:52:40
Google, Anthropic, Amazon, Nvidia,
00:52:45
they are all very convinced that orbital
00:52:47
compute is going to be reality and
00:52:48
obviously SpaceX is extraordinarily well
00:52:50
positioned for that. But I do think that
00:52:52
Composer 2.5 data point is really
00:52:54
powerful.
00:52:55
>> Keep an eye on it.
00:52:56
>> Yeah.
00:52:57
>> And then the other thing that's come out
00:52:58
is Grock build. So, what Grock lacked
00:53:01
that a lot of other models had, and I do
00:53:03
think it's important to remember that
00:53:06
the newest version of Grock 4.3 is on
00:53:08
the paro frontier for all frontier
00:53:10
models. And you're either on the
00:53:11
frontier or you're not. And the the
00:53:15
companies on the frontier are XAI with
00:53:18
one build of Grock 4.3, which is a 500
00:53:21
billion parameter model, Google 3.1 Pro,
00:53:24
and then OpenAI Anthropic. And that's
00:53:26
it. Those are the companies on the
00:53:27
frontier. and the four horsemen.
00:53:29
>> Google today each have one dot on the
00:53:32
paralo frontier and obviously you want
00:53:35
as many dots as possible but Grock
00:53:38
lacked a harness. So Claude had cla
00:53:42
OpenAI had codecs and now with Grock
00:53:45
build there is a harness that is
00:53:47
available to to Grock and you know as
00:53:50
I'm as I'm sure a downloadable app to
00:53:53
translate into English that has
00:53:55
integrations to all your favorite stuff
00:53:57
whether it's notion, Gmail, Slack, etc.
00:54:00
And if you don't have that it's just
00:54:02
like using a a basic chatbot from year
00:54:04
ago. So now they have their
00:54:05
downloadable. in market and they are
00:54:07
cooking with oil on it and they're
00:54:08
playing catch-up, but they're moving
00:54:09
fast
00:54:10
>> and and it's it's more than just an app.
00:54:12
It's it's a runtime. It's an
00:54:14
environment. It manages state. It
00:54:16
manages memory.
00:54:17
>> It makes these models dramatically more
00:54:21
useful to the extent that I think the
00:54:23
people at the frontier all agree that
00:54:25
the harness is essentially as important
00:54:28
as the model especially in an agentic
00:54:30
world and the harness and the model need
00:54:32
to be developed together. So the release
00:54:34
of gro build and the pace at which
00:54:36
they're iterating is I think also really
00:54:39
encouraging. So now you have cursor you
00:54:41
have the cursor data you have a clear
00:54:43
existence proof that the cursor data is
00:54:45
really important because composer 2.5 is
00:54:48
now paro dominant and the most selected
00:54:50
model on cursor and that's also
00:54:52
important because you know these evals
00:54:55
don't capture everything. You know, this
00:54:58
is why people on X talk about the vibes
00:55:01
>> and the vibes on Cursor 2.5 are also
00:55:03
really good.
00:55:04
>> They're immaculate.
00:55:05
>> Together,
00:55:06
>> yeah,
00:55:06
>> with Grock build, I think these are
00:55:09
really important developments.
00:55:10
>> Yeah, they're there was Elon was
00:55:13
incredibly frustrated by the state of
00:55:15
affairs at XAI. He was very public about
00:55:17
that and he's less frustrated now and
00:55:19
he's shipping a lot faster and so I
00:55:22
think that says something and he has
00:55:24
been very focused on it. Freeberg, your
00:55:28
thoughts on the SpaceX IPO and what this
00:55:31
collection of companies might look like
00:55:34
a year or two from now, especially if,
00:55:37
like many people believe, Tesla and
00:55:41
SpaceX merge. What do you think of
00:55:44
dollar sign El
00:55:46
N as an entity and what impact it might
00:55:49
have as if those two were put together?
00:55:51
The market cap would put them in the
00:55:52
fourth largest company in the world. We
00:55:54
can revisit our earlier conversation
00:55:56
about an anti-tech, anti- AI, anti-
00:56:00
progress world and society ahead. And if
00:56:04
there is an effort, a concerted effort,
00:56:06
an organized effort by governments to
00:56:08
stop or block access to information,
00:56:11
restrict freedom of speech, restrict
00:56:14
freedom of purchasing or buying things,
00:56:17
to control more things. And I think
00:56:20
there's a trend line in this direction
00:56:22
right now globally. The internet has
00:56:24
always been louded at this kind of
00:56:26
system that provides an open alternative
00:56:29
to physical commerce that you could
00:56:31
create digital commerce, digital uh
00:56:34
information, digital media that you
00:56:36
could share and um it's almost this
00:56:38
digital representation of society. But
00:56:41
the internet has to sit physically
00:56:42
somewhere. And the assault on data
00:56:45
center builds outs in the United States
00:56:46
right now, I think, may indicate the
00:56:50
importance of having an alternative
00:56:52
internet from the ground layer up. If
00:56:55
you have a communication network that
00:56:57
isn't restricted and controlled by a
00:57:02
government on Earth,
00:57:05
it's almost like a backup for
00:57:08
civilization, but it's a backup for
00:57:10
progress. And I don't own any SpaceX
00:57:12
shares and I'm not trying to sell the
00:57:14
book of SpaceX, but I think that there's
00:57:17
like an important aspect of can you
00:57:20
create a system that's not under the
00:57:22
control of governments as a way to
00:57:24
ensure humanity's progress to ensure
00:57:27
civilizational continuity if things go
00:57:29
south, if things aren't good, if things
00:57:31
are restricted, and if you know
00:57:33
fundamental forms of tyranny start to
00:57:35
restrict speech, restrict commerce,
00:57:37
restrict information flow and whatnot.
00:57:40
And I think having like a space-based
00:57:42
communication network, space-based data
00:57:45
centers, and space-based communication
00:57:47
back down to earth wireless,
00:57:51
I think it's generally a good thing.
00:57:53
It's good to have a backup.
00:57:54
>> Yeah. So, put all the economics aside
00:57:57
and the multiples and the valuations and
00:57:58
whatnot. And whether it's SpaceX or not,
00:58:01
I think the idea that you could have
00:58:02
data centers store information, transmit
00:58:04
information, route information, and
00:58:06
access information through space-based
00:58:08
systems that can't be controlled,
00:58:11
manipulated, or destroyed by governments
00:58:13
is is important. And I I just I like
00:58:16
that. Yeah. If you most people don't
00:58:19
remember this, but when Elon was
00:58:21
starting SpaceX, the original idea when
00:58:24
he was running around with ADO and they
00:58:26
were looking at some rockets and getting
00:58:28
carriage from Russian rockets was to
00:58:31
back up the biosphere. And he came back
00:58:33
from that trip and I remember talking to
00:58:34
him about it and he said, "I think I
00:58:35
just have to make my own rockets because
00:58:37
that's actually where the problem is and
00:58:38
it would be easier just to make my own
00:58:40
rocket to back up the biosphere." He
00:58:42
wanted to put geodomes
00:58:44
like geodessic domes in space with all
00:58:47
the plants and wildlife and and
00:58:49
creatures. Uh and what incredible vision
00:58:53
and then it you know there was the
00:58:54
necessity of actually getting that up
00:58:56
into space and that's that's the unknown
00:58:58
origin story. I will say this chimat the
00:59:01
idea of putting a data centers in space
00:59:05
seems completely doable even though
00:59:07
there are a bunch of people who are
00:59:08
saying it's not when you compare it to
00:59:11
what happened with SpaceX um with
00:59:13
Starlink which people said also wouldn't
00:59:15
work and now he's got 10,000 Starlinks
00:59:17
up there. The difference between a
00:59:19
Starlink satellite and a data center
00:59:21
satellite is really not that different.
00:59:24
And um
00:59:26
>> no they're they're pretty different.
00:59:27
Well, conceptually, yeah, of course
00:59:29
they're physically different, but
00:59:31
conceptually, Elon put 10,000 Starlinks
00:59:34
up. Is he capable of putting 10,000?
00:59:37
>> No. Look, the size rockets the size is
00:59:39
much bigger, Jason. The foils are much
00:59:41
bigger. The wings are much bigger,
00:59:43
>> but it's Yeah, my point is it's not
00:59:45
different if he has the new Starship
00:59:48
because that's 10 times bigger. Yeah,
00:59:50
>> you can't just scale like this. That
00:59:51
being said, it's technically possible. I
00:59:54
think he will be the first one to figure
00:59:56
it out. But I'll just take a much more
00:59:59
pedestrian take, which is okay, you're
01:00:01
sitting here and if I'm asking myself,
01:00:03
Jimoth, how do I underwrite SpaceX at 2
01:00:06
trillion? Here's the basic math that I
01:00:08
would do. Well, last year it did 181 19
01:00:11
billion. It'll probably do 25 to 30 this
01:00:14
year. Okay. So, I'm buying this thing at
01:00:18
a fairly
01:00:20
costly premium, right?
01:00:22
So, what am I buying?
01:00:25
Well, I'm buying probably the most
01:00:28
important internet infrastructure
01:00:30
project that's happened since the
01:00:31
internet itself. That's going to scale
01:00:33
to hundreds of millions of users. And
01:00:34
the reason that's going to scale to
01:00:36
hundreds of millions of users is it's
01:00:37
just very useful and it's just going to
01:00:39
become cheaper and cheaper and cheaper.
01:00:41
So, that's number one. I'm buying a
01:00:43
delivery infrastructure,
01:00:45
but I think over time
01:00:48
GDP plus 10, GDP plus 15 kind of a
01:00:51
grower. So good business, valuable
01:00:53
business, but it's the underlying
01:00:55
platform that allows everything else to
01:00:56
happen.
01:00:58
And then I'm buying an AI business,
01:00:59
which will be at the top level the apps,
01:01:01
but at the bottom layer all the compute
01:01:04
capability.
01:01:05
And I think when you scale that out,
01:01:09
like why is Colossus so valuable to
01:01:12
Enthropic? Maybe that's like a good
01:01:13
question to ask.
01:01:15
It's because if you look at who's
01:01:17
actually capable of delivering a
01:01:19
gigawatt data center, these guys are the
01:01:22
closest, like an actual gigawatt. And
01:01:25
and the reason is is that this stuff is
01:01:27
very complicated and very very hard. I
01:01:30
think you've probably heard this famous
01:01:31
story where Jensen was like, "Yeah, he
01:01:33
was the one that figured out this one
01:01:35
thing that we that nobody else could
01:01:36
figure out so that you could strip a
01:01:38
bunch of racks and drive a bunch of east
01:01:40
west traffic and make the whole thing
01:01:41
work together."
01:01:43
So I suspect what happens is next year
01:01:46
it's probably 40 45 billion and then the
01:01:50
year after that it probably doubles
01:01:51
again. So now I'm buying it at 20 times
01:01:53
revenue. And you would say well why can
01:01:55
you buy a company like this on revenue
01:01:57
versus earnings and cash flow.
01:02:00
And I think the reason is because what
01:02:01
the revenue does is it gives him the
01:02:04
operating leverage to go and invest in
01:02:06
all of these other businesses that
01:02:08
ultimately consolidate
01:02:11
his differentiation and his competitive
01:02:13
mode because what he creates is a
01:02:15
capital mode that then accelerates a
01:02:18
technology mode that then accelerates an
01:02:20
execution and a learning mode. And that
01:02:22
flywheel when it starts to spin very
01:02:24
quickly and you would say, "Hey, hold on
01:02:26
a second. It's probably spinning quickly
01:02:27
now." I would say we're at the beginning
01:02:30
of the beginning because he's again he
01:02:33
still has all these disperate assets. I
01:02:35
still don't like the fact that Tesla's
01:02:36
over here. And as I've told you that
01:02:39
will get merged in. And now you have
01:02:42
this incredible corpus of physical
01:02:45
capability,
01:02:47
movement of all kinds, X, Y, and Z,
01:02:50
right? You have learning capability, you
01:02:53
have infrastructure, you have all the
01:02:55
connectivity.
01:02:56
That thing will look very cheap, I
01:02:58
think, in a few years. and he has this
01:03:01
one thing that nobody else if you look
01:03:03
at the big CEOs
01:03:05
who steps on stage where you're always
01:03:09
curious okay what has he got up his
01:03:10
sleeve you know the Steve Jobs oh and
01:03:13
one more thing this is the only guy at
01:03:16
the scale of civilizational
01:03:19
out of left field
01:03:22
he's he's the guy whether you like him
01:03:24
or you hate him he's the guy and there's
01:03:26
a premium that is welld deserved that
01:03:28
comes with that so if If you had to pick
01:03:30
an underwriting case, Jason, I would
01:03:32
flex the revenue and realize that
01:03:34
terrestrial data centers alone are a
01:03:37
hundred or $200 billion of revenue by
01:03:39
2030 2032 just and that means just
01:03:42
building it. So already you're buying it
01:03:44
at 20 times revenue just for that
01:03:46
business. Everything else is the
01:03:47
Colossus on the ground base.
01:03:51
No, no, forget space for a second. It's
01:03:53
like Colossus 3, Colossus 4.
01:03:55
>> It pencils out with that. Yes.
01:03:57
>> Getting a name plate 1 gawatt. It is
01:04:00
freaking hard, man. Getting a gigawatt
01:04:02
name plate working
01:04:05
is almost. And then, by the way, there's
01:04:07
all the stuff that he can do on land
01:04:08
that he's the best position to do. I'll
01:04:10
give you one example. There's a great
01:04:12
push that Jensen's making, which he
01:04:14
needs a partner, and I think Elon
01:04:15
becomes a natural partner to do DC to
01:04:17
DC. Forget all this DC to AC to DC
01:04:19
nonsense that goes inside of a data
01:04:21
center. All the laws, heiness, all
01:04:22
>> explain what that is in English. Yeah.
01:04:23
forever just like look you go through a
01:04:25
bunch of power transformations to to
01:04:28
actually deliver the electrons into the
01:04:30
rack so that as Gavin said you can
01:04:31
generate the token on the other end
01:04:34
today it's it's very inefficient it's
01:04:36
very costly it requires a lot more power
01:04:38
it requires a lot of cooling it requires
01:04:40
complexity and what people have said is
01:04:42
wow if we could just do DC toDC like it
01:04:45
comes in as DC direct current it goes
01:04:47
right to the rack as DC but it requires
01:04:49
a fundamental rearchitecture Jensen
01:04:52
needs a design partner and a thought
01:04:54
partner to get that done.
01:04:56
>> He's probably the only one. So, I just
01:04:58
think there's a lot of reasons where you
01:04:59
can underwrite this to a multiple of
01:05:02
revenue plus the X factor which is just
01:05:04
the creativity and the the one more
01:05:06
thing. Love it. And then here's two uh
01:05:10
charts and I'll have you comment on
01:05:11
these Gavin after it.
01:05:13
Here's the rocket sizes just in terms of
01:05:16
scale and I most people have not
01:05:19
actually seen a Starship in person. when
01:05:21
you see this thing in person and I I've
01:05:23
been inside that rocket. I think you
01:05:26
were we were together Gavin when we were
01:05:28
in the first build and like inside of
01:05:30
that you can fit 300 people. It's
01:05:32
basically like a giant if you thought of
01:05:36
a commercial aircraft.
01:05:39
That's what it feels like when you're
01:05:40
inside, right? Like a 747 in terms of
01:05:42
the amount of space in it. Especially
01:05:44
when you compare the Falcon Heavy, which
01:05:46
is their workhorse. Correct, Gavin?
01:05:48
>> Yeah. And Starship's going to get
01:05:49
bigger. based on their road map, it's
01:05:51
going to get a lot bigger,
01:05:53
>> a lot bigger. And then this one is the
01:05:55
most interesting that this started
01:05:56
trending last week. This is cumin of
01:05:58
payloads launch 1957 to today. SpaceX is
01:06:02
basically about to in just that and and
01:06:05
this is really what exponential growth
01:06:07
is about and this is what disruptive
01:06:08
technologies are about. Just from 2012
01:06:12
to today, SpaceX is about to dwarf the
01:06:16
rest of the world's cumulative payloads
01:06:19
into space. So Gavin, maybe take the
01:06:23
other side of it. When do these data
01:06:26
centers in space happen? What has to
01:06:29
happen for those to be a reality? When
01:06:32
does that hit SpaceX's bottom line?
01:06:34
We've heard from Chimoth, hey, here's
01:06:35
all the things that hit the bottom line
01:06:36
in the short term and midterm. But I
01:06:39
think data centers in space would be a
01:06:40
midterm to long-term play. Three years
01:06:42
is what I'm hearing. So t tell us about
01:06:45
that business in relation to the two
01:06:48
charts I just shared.
01:06:49
>> Well, the one thing I would just say,
01:06:50
well, first all those charts about
01:06:53
launch are before Starship was
01:06:55
operational and most most of that master
01:06:57
orbit was done by Falcon.
01:06:59
>> Yes.
01:06:59
>> And Starship
01:07:01
the Falcon is reusable. Starship is
01:07:04
designed to be rapidly reusable. And
01:07:07
this is a critical difference. Like
01:07:09
let's say Blue Origin successfully
01:07:11
solves reusability.
01:07:13
They're where SpaceX was 10 years ago.
01:07:16
Let's say China solves it 10 years ago.
01:07:19
Rapid re the reusability means that you
01:07:21
extensively refurbish the rocket, you
01:07:24
know, the engines, everything, the
01:07:27
fairing, it takes a lot of time. you
01:07:29
know, maybe you can fly that rocket
01:07:31
again in
01:07:33
30 days, 60 days. Rapid reusability
01:07:37
means that you can fly the same fly and
01:07:39
land the same rocket multiple times per
01:07:41
day. So, if SpaceX and and it's really
01:07:46
hard to do rapid reusability.
01:07:49
I think it would have been much it would
01:07:50
have been not trivial but much easier to
01:07:53
have Starships working if it was just
01:07:55
designed to be reusable. Hm.
01:07:57
>> That's not enough for what Elon wants to
01:08:02
achieve of, you know, a a moon base, a
01:08:04
colony on the moon, a colony on Mars,
01:08:06
mass drivers on the moon. You need rapid
01:08:09
reusability and that is why Starship is
01:08:12
such a an engineering challenge and will
01:08:15
be such an impressive achievement when
01:08:17
they have rapid reusability. But I do
01:08:19
think that master orbit rapid
01:08:22
reusability in Starship means if they
01:08:24
get
01:08:25
>> when do when do you predict they'll have
01:08:26
that rapid reusability into space? You
01:08:28
think
01:08:28
>> I mean we're going to find out we're
01:08:29
going to find out you know we find out
01:08:31
I'm I'm I'm going to be at Starbase
01:08:34
today for the launch.
01:08:35
>> Yeah.
01:08:36
>> So you know we we we turn over cards and
01:08:38
you know it's important for everyone to
01:08:40
remember like let's just let's just say
01:08:43
it's a fireball. SpaceX will still learn
01:08:47
from this. Yes,
01:08:49
>> they they learn from failure. If you
01:08:50
don't fail, you're not learning. Same
01:08:52
way if you're not wrong, you didn't
01:08:53
learn anything in that day. And this is
01:08:56
a brand new rocket, a brand new booster,
01:08:58
a lot of new technology. There's a lot
01:09:00
of instrumentation on it. So, whatever
01:09:03
happens today, SpaceX is going to learn
01:09:05
and rapidly iterate. I don't know when.
01:09:08
I don't want to make a prediction. I
01:09:11
would guess a year or two,
01:09:13
maybe sooner.
01:09:14
>> I think that's most consensus. A year or
01:09:15
two is I think perfect consensus. Yeah,
01:09:17
we'll see. So like the even if everybody
01:09:20
else solves reusability, master orbit
01:09:22
from everyone else will quickly asmtote
01:09:25
to a very small number.
01:09:29
As far as when will orbital compute be a
01:09:32
reality, I would say well it is
01:09:33
important to realize there is a working
01:09:36
H100 and Nvidia H100 GPU in space today.
01:09:40
>> Yeah.
01:09:40
>> Andre Karpathy both trained a model on
01:09:43
and used for inference. So this is, you
01:09:47
know, it's it there's a working GPU in
01:09:50
space today.
01:09:51
>> And Nvidia is making a space designed
01:09:55
version of this, which will be different
01:09:57
because the heat sink has to be
01:09:58
different. There's a bunch of weight
01:09:59
that you put on it when it's in a data
01:10:01
center that you don't need in space. And
01:10:03
you also have to reinforce it for the
01:10:05
journey to space because these things
01:10:07
are going to shake and break apart. The
01:10:09
data center ones are not made to have
01:10:12
that many G's put on them. So, you're
01:10:13
going to need an an industrial an
01:10:16
industrial strength one that gets to
01:10:18
space that has a different profile.
01:10:19
Yeah, Gavin.
01:10:20
>> Well, one of the things that's been so
01:10:21
magical about SpaceX is they're very
01:10:23
good at engineering the rocket and the
01:10:25
payload that you can use semiconductors
01:10:30
that are not designed
01:10:32
>> be in space or satellites in space and
01:10:34
those semiconductors are a lot cheaper.
01:10:36
We have a couple my my firm a trades is
01:10:38
an investor in a company called Exite
01:10:40
Labs that it's a matter of public record
01:10:43
is going to be in essentially every
01:10:45
Starlink and the chips were not designed
01:10:49
to go to space. They're not radiation
01:10:51
hardened. You know, everyone, you know,
01:10:54
SpaceX really liked a lot of the the
01:10:56
specifications on the chips and then
01:10:58
it's like, well, we'll see how they do
01:10:59
with rad testing and they just, you
01:11:01
know, happened to pass. And so that is
01:11:04
one of like one of something that's very
01:11:07
underappreciated I think about space.
01:11:08
It's
01:11:08
>> one of their specialtities.
01:11:10
>> One of their specialtities.
01:11:11
>> Yeah.
01:11:11
>> But I think second half of 28 to first
01:11:15
half of 2030 would be my point
01:11:16
prediction.
01:11:17
>> All right. Let's do Nvidia and then the
01:11:19
the market recap since we have you here
01:11:20
Gavin and since Freeberg you wanted to
01:11:22
get in on that. Nvidia blew out its
01:11:23
earnings again. Q1 performance is just
01:11:26
mindboggling. 81.6 6 billion in revenue,
01:11:29
up 85% year-over-year, 20% quarter over
01:11:32
quarter. High growth in the stock market
01:11:34
for those people who don't participate.
01:11:36
20% would be a high growth company
01:11:38
year-over-year. They did that quarter
01:11:40
over quarter. 58 billion of net income
01:11:44
and uh 48 billion in free cash flow.
01:11:46
They're doing all this at 75% gross
01:11:48
margins.
01:11:50
They're growing massively. And they're
01:11:53
obviously the most valuable company in
01:11:55
the world at a $5.3 trillion market cap.
01:11:58
Stocks up but 16% year this year with
01:12:01
all that growth. That's a magnitude of
01:12:04
that 16%.
01:12:06
And um they've announced another 80
01:12:09
billion in additional buybacks on top of
01:12:11
the 100 billion in buybacks uh they did
01:12:13
at the start of 2023. So they're buying
01:12:15
back about 4% of the company. They
01:12:17
raised the quarterly dividend 25x from 1
01:12:20
cent a share to 25% cents per share and
01:12:23
their CFO said they're going to return
01:12:25
50% of the free cash flow to
01:12:27
shareholders.
01:12:29
Never been a company like this. Huh,
01:12:30
Freeberg. The the scale of this is just
01:12:33
extraordinary.
01:12:36
>> Uhhuh. Yep.
01:12:38
Don't don't see it.
01:12:39
>> There you have folks. There's your
01:12:40
market report from Freeberg.
01:12:42
>> Don't seem so enthused.
01:12:43
>> Yeah, it's a mhm. He's got potatoes in
01:12:46
the oven. I have a question for Gavin.
01:12:47
He did a really interesting talk with
01:12:50
Patrick Oshanaughy and there was this
01:12:52
one thing that I wanted to ask you about
01:12:54
because I thought it was so interesting.
01:12:56
You said when you look at the revenue
01:12:58
multiples of the chip companies and you
01:12:59
look at the revenue multiples of the DRM
01:13:01
companies, both cannot be true. In the
01:13:03
context of Nvidia's earnings, can you
01:13:05
just explain maybe in plain language for
01:13:08
folks? I just thought it was so
01:13:09
fascinating because it explains
01:13:11
>> it explains I think just to set it up
01:13:13
>> where is value over the next five years
01:13:15
like I think if you looked at Leo Ashen
01:13:17
Brener his fund has gone from like 0 to5
01:13:21
billion overnight and it looks like he's
01:13:22
just got massive puts on the chip sector
01:13:24
and he's kind of rotated. So just give
01:13:27
us context Gavin where where's the puck
01:13:29
going?
01:13:30
>> Well so maybe take the questions in
01:13:32
reverse order for Leopold who's clearly
01:13:35
a brilliant man. I think he's a road
01:13:37
scholar at like 19 and I think my
01:13:39
understanding he's putting up pretty
01:13:41
extraordinary numbers. I've yet to meet
01:13:43
him. He actually shares an office in San
01:13:45
Francisco with a friend of mine. So I
01:13:47
think I'll probably meet him sometime
01:13:48
soon. But it's got to for that 13F that
01:13:51
he filed was at the end of the first
01:13:54
quarter when you know I would say we
01:13:57
were in the you know the the thick of
01:13:59
geopolitical fears and I think you saw a
01:14:02
lot of puts on a lot of 13Fs and I don't
01:14:05
know that those puts are still there.
01:14:08
>> Okay.
01:14:08
>> You know I think a lot of people wanted
01:14:09
to be hedged for Ron and you know now I
01:14:13
think it's a little more clear. So I
01:14:14
wouldn't read I wouldn't read Leopold's
01:14:17
13F as being super negative on on SIM
01:14:21
>> chips. Okay,
01:14:22
>> on chips. Second thing I think
01:14:24
cross-sectionally if you look at the
01:14:26
valuations for all these AI DS they just
01:14:28
they can't all be accurate. You have
01:14:31
memory makers at you know three to five
01:14:35
times PE. You have Nvidia at a really
01:14:38
low PE. you actually have um you know
01:14:42
some other accelerator companies at
01:14:43
reasonable multiples and then you have
01:14:46
everything else everything in power
01:14:47
everything in cooling and when I say
01:14:49
power I don't mean utilities the IPS are
01:14:51
actually quite reasonably valued but
01:14:54
power cooling even probably some of some
01:14:58
of the optical names
01:15:00
these are discounting very different
01:15:03
things if the multiples on the power
01:15:05
cooling optical names are correct Nvidia
01:15:10
memory, they're going up a lot. If the
01:15:13
multiples on Nvidia and memory are are
01:15:15
correct, everything else is probably
01:15:17
going to underperform. The AI market is
01:15:20
cross-sectionally inefficient right now,
01:15:22
which is what I was trying to say. As
01:15:24
far as the Nvidia quarter, I do um they
01:15:26
went to a new reporting structure, data
01:15:28
center and AI and then with no data
01:15:30
center and edge and then within a within
01:15:33
AI they have hyperscalers and then I
01:15:36
think they call it AI clouds
01:15:37
>> AI clouds
01:15:38
>> industrial and enterprise. What I
01:15:40
believe is if we were to make a true
01:15:42
applesto apples comparison
01:15:45
and Broadcom, you know, there's a
01:15:47
narrative that Nvidia is losing share to
01:15:49
the TPU
01:15:50
and Broadcom guided for 143%
01:15:54
year-over-year growth in their AI
01:15:56
semiconductor revenue in the quarter
01:15:58
that they will report. Um, that's
01:15:59
comparable to the one Nvidia just
01:16:01
reported. I think if you were to and I
01:16:06
just I so wish they had reported
01:16:09
slightly differently. I wish they' done
01:16:10
hyperscalers, AI clouds and then
01:16:13
industrial and enterprise because I
01:16:15
think the segment that is comparable is
01:16:18
the sum of hyperscalers
01:16:20
plus AI clouds
01:16:23
stripping out China because Bracomob
01:16:25
just did not have the China business
01:16:26
that Nvidia did. And I think on that
01:16:28
basis, in other words, in within the
01:16:31
western AI world, within data centers
01:16:34
that need being built, whether they're
01:16:35
being built by Coree, XAI, Amazon,
01:16:38
Google,
01:16:39
>> Nvidia's AI business is growing faster
01:16:42
than Broadcom's
01:16:44
and faster than a lot of other companies
01:16:47
that are, you know, seen as part of this
01:16:49
AS6 share gate story.
01:16:51
And you know, I think Jitson has become
01:16:55
you can you can hear it
01:16:58
increasingly frustrated and rightfully
01:17:00
so with two things.
01:17:02
I would say what is the performance of
01:17:04
the stock to to
01:17:06
>> he's been vocal about that. Like what is
01:17:08
going on here? You're putting up record
01:17:09
numbers and we're getting no like uh
01:17:12
credit.
01:17:13
>> Yeah, I get it. And and just how can
01:17:16
there be a share loss narrative if I am
01:17:18
gaining share? And it is indisputably
01:17:21
true that he is growing faster than
01:17:22
hypers scale or capex even without these
01:17:25
adjustments.
01:17:26
>> Yeah.
01:17:26
>> And I think the other thing that's so
01:17:27
frustrating to him is these other
01:17:29
>> AS6
01:17:30
>> are not being submitted for benchmarks.
01:17:33
They're not in the SIM analysis in
01:17:34
inference max. They're not in ML Perf.
01:17:37
And I think the reason they're not being
01:17:39
submitted is they will lose
01:17:41
>> and you can't fight shadows. And until
01:17:45
we see a clean benchmark
01:17:48
of whether it's GB300's verse TPU V7s or
01:17:53
you know very
01:17:54
>> versus inferentia. Yeah. Yeah.
01:17:56
>> Versus Yeah. versus Tranium.
01:17:57
>> Yeah.
01:17:58
>> We're not going to know. And that's why
01:18:00
a lot of these other chips I think
01:18:02
Tranium's in a great spot aren't being
01:18:04
submitted. But nonetheless, if India is
01:18:06
doing well, once you become the largest
01:18:08
company in the world, you kind of you
01:18:10
tend to trade by observation would be in
01:18:12
stairst step patterns where you kind of
01:18:15
the multiple compresses compresses
01:18:16
compresses because people are skeptical
01:18:18
of the size.
01:18:19
>> Then you have a rerating. Yeah.
01:18:20
>> And then you rerate.
01:18:21
>> New floor is established at a higher
01:18:23
rate. Yeah.
01:18:24
>> I think there was one other really
01:18:26
important thing in the VA quarter. It's
01:18:28
they said that they thought their CPU
01:18:29
business was going to be $20 billion
01:18:31
this year.
01:18:32
>> Yeah. Yeah. Yeah. That's extraordinary.
01:18:33
It means overnight we're one of the
01:18:35
world's largest CPU manufacturers.
01:18:38
>> And I think that is a testament to
01:18:40
Nvidia has a unique position. They're
01:18:42
the only company that works with every
01:18:46
lab. And so that puts them in the best
01:18:49
position to architect their chips. They
01:18:52
call it co-design for where the models
01:18:54
are going. And I think that $20 billion
01:18:57
CPU figure is pretty extraordinary. This
01:19:00
is the thing like at the end of this
01:19:02
Grock transaction last year my kind of
01:19:03
prevailing thought on this is we're
01:19:05
going to move to these domain specific
01:19:07
architectures. I thought that was like a
01:19:09
a feta complete. We're just now waiting
01:19:11
for which models. But the reality is
01:19:13
that that DSA
01:19:16
market evolution is actually happening
01:19:18
inside of Nvidia. That's what's so
01:19:19
insane to me. That was my takeaway from
01:19:21
the quarter as well which is like holy
01:19:22
[ __ ] these guys actually have domain
01:19:24
specific architectures because they're
01:19:26
doing these design programs with every
01:19:28
this is why back to sort of the you know
01:19:29
when he does DC toDC with Elon and
01:19:31
Colossus 3 or whatever it's just it's
01:19:34
another game changer for everybody
01:19:35
>> he doesn't he makes nine and then I
01:19:39
think the cost at which you can finance
01:19:41
these chips and these useful lives is
01:19:43
really important
01:19:44
>> you had an incredible insight which is
01:19:45
the amortization schedule for coravee
01:19:47
and all these guys they got saved You
01:19:50
may want to just explain what that is
01:19:51
and what why you said that. I thought
01:19:52
that was a great insight.
01:19:53
>> No, thank thank you Jamatha. I
01:19:54
appreciate it. So when core weave and
01:19:56
all these neo clouds came public and by
01:19:57
the way this goes for the hyperscalers
01:19:59
too. There was a big bare case that hey
01:20:01
the these guys are amortizing their GPUs
01:20:03
and CPUs over four, five, six years and
01:20:06
that's way too short of a lifespan. The
01:20:08
true lifespan of a GPU is more like two
01:20:10
years and therefore you know the profits
01:20:12
of all these businesses are overstated.
01:20:15
>> The reality is
01:20:16
>> this was Michael Bur who put this out
01:20:17
there. Yes. To be clear. Yeah. Yeah. And
01:20:20
you know, thank you, Michael Bur. We
01:20:22
need bears. Thank you.
01:20:24
>> Yeah. That's like It's like asking
01:20:25
Gerardo about modern music.
01:20:28
>> Well, I don't I don't want to cast
01:20:30
dispersions on Michael Barry. He's a
01:20:31
He's a brilliant man. But we need bears.
01:20:33
>> Let me Hey, somebody call Vanilla Ice
01:20:35
and ask him what he thinks.
01:20:37
>> Oh my god. Millie Vanilli, check it out.
01:20:40
>> What a waste of time. What a waste of
01:20:42
time.
01:20:43
>> Oh, poor. Come on the program anytime.
01:20:45
Go ahead, Kevin. Keep us on track.
01:20:48
>> Happy to chat. But now that we've
01:20:49
disagregated inference, we have these
01:20:51
different domain specific accelerators.
01:20:53
You can mix and match them.
01:20:55
>> And I think the GPU stays in a lot of
01:20:57
ways at the center of this constellation
01:20:59
for a while.
01:21:00
>> And you can put whether it's a Grock
01:21:03
accelerator, whether you know it's a
01:21:05
Cerebrus accelerator in front of old
01:21:07
GPUs, use Grock or Cerebrus for decode.
01:21:11
And then those older GPUs, they have a
01:21:12
useful life for 10 or 15 years. And this
01:21:15
means that you can finance GPUs. I think
01:21:17
Cory's lowest financing, I can't forget
01:21:18
if it's six or 7%.
01:21:20
>> 6% come down%.
01:21:22
>> And if you can get an assetbacked loan,
01:21:24
an asset backed financing
01:21:26
>> for GPUs at a lower rate than other
01:21:29
chips.
01:21:29
>> No, that that's a profound advantage.
01:21:31
>> That quarter single-handedly saved the
01:21:33
Neo Neos clouds this Nvidia. I mean,
01:21:36
they single-handedly saved them all. I
01:21:39
think they they should all they should
01:21:41
all say an incredible thanks to Jensen
01:21:43
because
01:21:44
>> I interviewed the CEO of Core Wee,
01:21:46
Michael in Trader. Michael and Trader.
01:21:48
Yeah. And he was saying, "Hey, listen.
01:21:50
People have no problem buying and
01:21:52
financing these over a six-year period
01:21:55
and people are asking for things that
01:21:56
are coming off and that he thinks
01:21:58
they're going to have year seven, eight,
01:22:00
nine, they'll have some useful life, you
01:22:02
know, uh, in addition to that." So,
01:22:05
>> yeah. So, he's like, I I I don't know
01:22:06
what anybody's talking about here. Like
01:22:08
this is just not informed analysis was
01:22:11
his point. Like I the game on the field
01:22:13
and people are betting with their
01:22:14
dollars with him. He has them pay in
01:22:17
advance and sign six-year contracts. If
01:22:19
they didn't think it has a six-year
01:22:20
lifespan, they wouldn't be signing a
01:22:22
six-year contract. Pretty
01:22:23
straightforward. And they can't get
01:22:24
enough of them. Okay, let's end on this
01:22:26
market update. Macro picture. Not great.
01:22:28
Oil remains elevated. Although there
01:22:31
might be a settlement. Every week we
01:22:32
have there is a settlement coming. Maybe
01:22:34
this time uh 16th time it's a charm. in
01:22:36
the Iran war is going to wrap up, but
01:22:38
we're in week 12 of it and this was
01:22:40
supposed to be four to six weeks. So
01:22:41
wars never get resolved quickly. That's
01:22:43
one thing we've learned in our
01:22:44
lifetimes. Oil is driving inflation
01:22:47
massively higher. Poly market says 99%
01:22:49
chance May inflation comes in at 4.2% or
01:22:52
higher. Survey of professional forecast
01:22:54
is projecting CPI hits 6%. You heard
01:22:58
that right, folks. We weren't just
01:22:59
talking about a 3% handle, which we just
01:23:01
hit. Now people are saying four, five,
01:23:03
and 6% in Q2. And obviously that's a
01:23:07
huge revision and uh the narrative was
01:23:10
hey more Fed rate cuts coming. Now we're
01:23:13
talking about Fed rate increases.
01:23:16
Inflation is causing obviously bond
01:23:18
yields to rise. 10ear hit 4.6%. You
01:23:21
remember we've had Besson on the uh pod
01:23:24
multiple times and his goal was to get
01:23:26
that under 4%. Now it's significantly
01:23:29
above that number. And also if we go
01:23:31
around internationally, Japan's 30-year
01:23:33
is at a high of 5.1%, highest ever
01:23:37
recorded. UK yields highest since the
01:23:39
great financial crisis. Germany highest
01:23:41
since 2011. And in Korea, retail
01:23:44
investors are borrowing borrowing record
01:23:46
amounts of money to trade in AI chip
01:23:48
stocks. They also had a incredible run
01:23:51
in Korea betting on crypto at the peak.
01:23:54
So that's some sort of an interesting
01:23:56
signal.
01:23:57
Freedberg, is your alt personality going
01:24:00
to come out right now? Are you
01:24:01
concerned?
01:24:03
Is Dr. Doom making an appearance here or
01:24:05
do you think this is manageable? How
01:24:06
concerned are you? What is the point of
01:24:08
being concerned when you have ridden the
01:24:10
roller coaster to the top and it is
01:24:13
beginning its descent? I I don't know
01:24:14
what there is to be concerned about. The
01:24:17
the the force of gravity is inevitable.
01:24:20
>> The roller coaster will roll down. We
01:24:22
will throw our hands in the air and we
01:24:24
will scream wee as we go for the ride.
01:24:28
Global debt to GDP is 310%.
01:24:32
Reserve currency status to the side. The
01:24:36
spending problem at the federal, state,
01:24:39
local level. The spending problem at
01:24:41
every country to basically keep
01:24:43
economies growing to support existing
01:24:45
leverage ultimately creates a cascading
01:24:48
effect. it ultimately breaks and uh as
01:24:51
it starts to break you have massive
01:24:54
inflation because the value of your
01:24:56
underlying um currency collapses and
01:24:59
then you have money printing and all
01:25:00
this other sort of stuff which inflates
01:25:02
the value of assets which allows you to
01:25:04
keep servicing your debt and the spiral
01:25:06
takes off and so we will just uh enjoy
01:25:09
the ride. This is the moment you know
01:25:12
30-year Treasury 5.2% 2%. This Japanese
01:25:15
yield some argue might, you know, you
01:25:18
should talk to more active market
01:25:19
participants than me, but and probably
01:25:22
some economists who trade the market,
01:25:24
but I would think that this is one of
01:25:25
those things that could be a catalyst
01:25:27
for a for a a credit crisis
01:25:30
because there's a lot of people that are
01:25:31
in this carry trade and we'll see. You
01:25:35
know, this is
01:25:36
>> okay.
01:25:37
>> This is water leaking out of the bucket.
01:25:38
There it is.
01:25:39
>> Oh, there he is. Dr. Dumisier Jam, your
01:25:41
thoughts on
01:25:43
Dr. Doom's panic attack of the month. Uh
01:25:46
is this uh is this time real? Is this uh
01:25:49
the 17th prediction of the next six
01:25:51
recessions? What do we got here?
01:25:52
Chimath, are you concerned? How
01:25:54
concerned are you about these signals
01:25:56
that are flashing?
01:25:58
>> I I think that's exactly what that is,
01:25:59
Jason. There are signals that are
01:26:01
flashing. I think there's pockets of the
01:26:02
market that still make sense that you
01:26:04
can underwrite if you want to buy
01:26:06
businesses that represent the future.
01:26:09
And if you can find a few of those and
01:26:11
you can get comfortable with that and
01:26:13
you can own it for 10 years, I think you
01:26:16
buy those companies.
01:26:18
And generally everything else, I think
01:26:20
you should not speculate and you should
01:26:22
generally avoid, not just because it's
01:26:24
an up market, but in every market. I've
01:26:26
learned this the hard way. We've all
01:26:28
kind of gone through this. As I get
01:26:30
older, it's it's just not worth it. the
01:26:32
vicissitudes of the market um don't give
01:26:35
me anywhere near the sugar high it used
01:26:37
to on the way up and it makes me feel
01:26:40
horrible on the way down. So how I
01:26:41
manage myself is I have a few companies
01:26:44
that I really believe in. I have
01:26:45
extremely concentrated large holdings in
01:26:47
those large for me doesn't mean large
01:26:49
for everybody else and then otherwise I
01:26:52
just kind of stick to my eating and keep
01:26:54
my head down. It's a much more rational
01:26:55
way to behave. So, how many public
01:26:58
>> stocks can you keep in your brain and
01:27:00
still sleep at night holding for the
01:27:02
long term? What's the Is there a number
01:27:04
for you? Is it five? Is it 10? Is it
01:27:06
>> five? It's five or less.
01:27:08
>> Five or less.
01:27:09
>> Five or less.
01:27:09
>> And so, what your largest holding right
01:27:11
now is what percentage of your net worth
01:27:13
would you say?
01:27:14
>> I don't know.
01:27:15
>> Or like top two maybe.
01:27:16
>> Yeah.
01:27:18
>> Oh, top two.
01:27:19
>> Yeah. Like one and two. One is 20, two
01:27:21
is 15, or one is 40, two is 20.
01:27:24
>> I don't know. Again, it depends on the
01:27:26
day. I don't know. But it's it's
01:27:27
>> just curious. But I think that's I think
01:27:29
that's really important.
01:27:30
>> My point is there's no 30 things that
01:27:32
I'm tracking. I don't I don't have the
01:27:33
time. I'm not smart enough. There's too
01:27:35
much information. There's like four
01:27:37
things that I stay on top of.
01:27:39
>> Gavin, you you do this for a living. How
01:27:41
many positions do you manage? And what's
01:27:43
your take on some of the flashing signs
01:27:45
that are saying, "Hey, slow down." Or
01:27:47
maybe there might be a wreck around the
01:27:49
corner here, you know, when they do the
01:27:51
checkered flag in the F1 or whatever the
01:27:52
metaphor you want to use is. Well, so
01:27:55
one, I I manage more than 100 positions
01:27:59
at my firm. Um, and I do that with a
01:28:01
team. We're over 30 people now. So, it's
01:28:04
not just me and I work with some some
01:28:06
great people. And then three things can
01:28:08
be true. Rates going up is very
01:28:11
concerning. What is happening with AI
01:28:14
right now with anthropic growing faster
01:28:16
than any country, any company in history
01:28:19
at massive scale
01:28:20
>> and certainly any country.
01:28:22
>> Absolutely unprecedented. Yes, they're
01:28:24
actually Yeah, they they're now the size
01:28:26
of, you know. Yeah. Yeah. They're pull
01:28:29
up where they
01:28:30
>> they're bigger than a hundred different
01:28:31
countries for sure.
01:28:32
>> Exactly. And keeping really fast
01:28:34
>> and now profitable.
01:28:36
>> Yeah.
01:28:36
>> Which I think really changes the moment.
01:28:39
Yeah. Yeah. How did that happen? Yeah.
01:28:40
>> So those things can both be true. And I
01:28:43
think we should all remember the tech
01:28:44
bubble happened with you know the
01:28:45
10-year and the 30-year much higher than
01:28:47
they are today. And you know the the
01:28:51
Nvidia of of the tech bubble is Cisco
01:28:53
and it traded 100 times forward
01:28:55
earnings. And you know I think Nvidia is
01:28:57
probably at a low teens multiple of kind
01:28:59
of low to mid- teens multiple of real
01:29:02
earnings. You know if you that that'd be
01:29:04
a buy side consensus not in a trades
01:29:06
number. And then the third thing that I
01:29:07
think is true is a straightup formula
01:29:09
being closed. While it's terrible for
01:29:11
everyone, it is relatively the best for
01:29:14
America because we are self-sufficient
01:29:16
in energy. We are self-sufficient in
01:29:19
food. We have become a massive exporter
01:29:21
of oil. We're now the world's largest
01:29:23
not only oil and gas producer but oil
01:29:26
and gas exporter. And you know those
01:29:29
three things can be true. What do they
01:29:30
mean? I think it's probably hard for me
01:29:32
to see with America being the most
01:29:35
advantaged by what is going on.
01:29:38
>> We're still the best currency. Yes.
01:29:40
>> Still the best economy. We still have
01:29:42
the best public marketies and we have
01:29:44
the best private market companies.
01:29:46
>> Yeah.
01:29:46
>> And we are
01:29:48
>> one of the greatest producers of oil in
01:29:51
the world. So we're in good shape
01:29:52
despite this international chaos.
01:29:54
>> I don't think a dollar crisis is around
01:29:56
the world. Now listen, like if you know
01:29:58
the Bundes Bank and was still in charge
01:30:01
and you had the Deutsch mark and there
01:30:03
was a currency with better fundamentals,
01:30:05
we would be at very high risk probably.
01:30:08
But just because we're the best house in
01:30:11
what is globally a you know bad
01:30:13
neighborhood of high debt levels and we
01:30:16
have AI in our corner and we have energy
01:30:19
self-sufficiency and every day the
01:30:21
straight of form was closed I think is
01:30:23
relatively good for the
01:30:24
re-industrialization of America
01:30:27
like I think you know you have to you
01:30:29
have to all the
01:30:30
>> you're saying it's a forcing function it
01:30:33
makes us just like co did be more
01:30:35
resilient and yeah more self-reliant.
01:30:38
>> Yes. And
01:30:41
electricity is a base input to every
01:30:45
manufacturing or industrial process
01:30:47
essentially all of them.
01:30:48
>> And what we make electricity with in
01:30:50
America overwhelmingly is natural gas.
01:30:52
And you can look it up NG1. It is down
01:30:55
this year.
01:30:56
>> The input cost for electricity in the
01:30:59
rest of the world, you know, lots of
01:31:01
different things, but LG is a very
01:31:02
important one. And it's up 100 200%. And
01:31:06
so the street of form is absolutely
01:31:09
bad for everyone, but relatively good
01:31:11
for America and relatively good for
01:31:14
Trump's policy goals.
01:31:16
>> And that's why I think he's in no hurry.
01:31:19
Every day the straight is closed and for
01:31:22
whatever he does seem like a relative
01:31:24
thinker. Every day the straight of is
01:31:26
closed is relatively good for America.
01:31:29
It's terrible for Europe. It is terrible
01:31:32
for Asia.
01:31:33
>> Japan and China need that oil.
01:31:35
Philippines needs it, India needs it.
01:31:37
Yeah.
01:31:37
>> Yeah. And so all these things can be
01:31:39
true, but the one thing I do just want
01:31:40
to say is rates going up and inflation
01:31:43
going up is never good. But we have to
01:31:47
hold what's happening with AI where the
01:31:48
fundamentals are getting a lot stronger
01:31:51
in our mind. And then the one thing I
01:31:53
would just add is
01:31:55
AI has been seasonal. the market's
01:31:57
seasonal, you know, it often, you know,
01:31:59
sell in May, go away.
01:32:02
>> And AI fundamentals also appear a little
01:32:06
seasonal.
01:32:08
>> In the past, that's been because, you
01:32:09
know, because college students, they
01:32:11
they use a lot less chat, GPT, and
01:32:13
claude in the summer. And generally,
01:32:16
people maybe work a little less hard
01:32:17
when the weather's nice. Now, with a
01:32:18
JPIC AI, will the fundamentals still be
01:32:21
seasonal? We will see.
01:32:23
>> Oh, that's a really interesting point,
01:32:24
right? We see that e-commerce apps,
01:32:26
subscriptions, what as investors in a
01:32:28
lot of these companies, we would always
01:32:30
have these board meetings. Chimoth, oh,
01:32:32
Q3, uh, yeah, people are out
01:32:35
gallivanting and they're not playing
01:32:37
Candy Crush or buying comm or whatever,
01:32:39
but oh, hey, Uber and Door Dash went up.
01:32:41
People are traveling, etc. Okay, final
01:32:44
story of the week. We had a 48 hour
01:32:46
jaunt by a bunch of tech CEOs and the
01:32:50
president to hang out with she. A lot of
01:32:53
high fives, a lot of handshakes, a lot
01:32:56
of great vibes, but coming out of it, we
01:32:59
haven't seen anything definitive.
01:33:00
Freedberg in terms of policy. This was
01:33:04
supposed to be some big breakthrough. It
01:33:06
would have downstream effects on
01:33:08
tariffs, on us selling chips to China.
01:33:11
We did see a little movement there. Uh,
01:33:13
and the straight of war moves. we would
01:33:15
become, you know, wonder twins with she
01:33:17
and Trump reopening it. But nothing
01:33:20
really definitive other than some
01:33:21
soybeans being sold and some H100s,
01:33:26
A200s getting sold to BYU and some of
01:33:28
the top folks. So what do and maybe some
01:33:30
planes got sold too. So other than a
01:33:32
little BD, a little business
01:33:34
development. Yes. Uh Freyberg, what's
01:33:36
the outcome here or was it just a bit
01:33:39
performative in your mind? There was a a
01:33:42
question
01:33:44
would the administration
01:33:46
leave China with a grand deal that made
01:33:49
everyone feel like there's a long range
01:33:51
view on a partnership and I don't think
01:33:52
that that's what happened. There were a
01:33:55
few announcements obviously around an
01:33:57
intention to continue to work together
01:33:58
in a cooperative way and find a path to
01:34:00
partnership, an intention to establish
01:34:03
additional trade deals and and you know
01:34:05
there were some purchases of aircraft
01:34:07
and and some agricultural product
01:34:10
commitments but fundamentally the grand
01:34:12
deal the big deal that I would say
01:34:14
reduces deescalates tension probably
01:34:18
didn't manifest as some had hoped it
01:34:20
would. And I don't think it's any
01:34:21
surprise that Putin is with Shei today.
01:34:25
And this is also performative that
01:34:28
following the US visit, there's now a
01:34:30
relationship bonding moment happening
01:34:32
between China and Russia. So the story
01:34:36
continues.
01:34:37
>> You know, there is no happy ending and
01:34:38
there is no rainbow colored chapter 3 in
01:34:41
this book. It's going to continue to be
01:34:44
a dramatic arc as this rising power
01:34:47
continues to challenge the United States
01:34:50
and I think the story continues.
01:34:51
>> Were you expecting a happy ending?
01:34:54
>> I can't answer that question.
01:34:58
>> I mean to the story of the China visit.
01:35:02
Not not I'm not talking about any other
01:35:03
things going on in your life, but
01:35:05
>> did seem like some planes and soybeans
01:35:07
got sold, some H100s perhaps, but I mean
01:35:11
it wasn't like there was some grand deal
01:35:13
that occurred, but it's nice to see them
01:35:15
together, right? I mean, that is nice.
01:35:16
>> I think it was successful. I think
01:35:18
there's what you see on the surface and
01:35:19
then there's what happens behind closed
01:35:21
doors. And uh without speculating too
01:35:23
much, I think that it was a useful and
01:35:25
productive trip. I think the the biggest
01:35:27
thing that they probably got alignment
01:35:29
on is just geopolitically the
01:35:31
tic-tac-toe of what has to happen next
01:35:33
and and I think that there's some amount
01:35:35
of agreement there. I'm just guessing.
01:35:37
>> Yeah. And so that guess if I was going
01:35:39
to unpack it, hey, we get uh we have
01:35:42
Venezuela, we have Iran and Taiwan. you
01:35:46
have
01:35:46
>> I would just say look here's this
01:35:48
geopolitical chess board and
01:35:51
>> here's what I would say just very
01:35:52
generally like it's just
01:35:54
>> I think that there's a that there's a um
01:35:57
a way to divide up the game board in a
01:36:00
way that helps them and helps us
01:36:04
>> Gavin any thoughts on specifically
01:36:06
Nvidia being able to sell more chips
01:36:08
into China material for the company
01:36:12
good for America I mean it's obviously a
01:36:15
pretty debatable issue.
01:36:16
>> I think I disagree with Chimoth on this.
01:36:18
I think Stelling and deprecated Nvidia
01:36:22
GPUs to China
01:36:26
lowers the odds of them developing their
01:36:28
own alternative ecosystem which would be
01:36:31
a lot power hungrier um because you use
01:36:34
optical you bring in optical a lot
01:36:36
earlier for scaleup fabrics. I think
01:36:38
there's sound arguments that this
01:36:41
is stabilizing for the world and is the
01:36:46
best highest probability path for
01:36:49
keeping America ahead in AI and kind of
01:36:53
keeping control of AI. And that's almost
01:36:55
a shame we've had to have this debate
01:36:57
because now people like me have said
01:36:58
this many times in China. If they didn't
01:37:00
understand it, they probably do really
01:37:02
understand it. By the way, that's not to
01:37:03
say we shouldn't have had the debate,
01:37:05
but um that is what I believe. You know,
01:37:09
reasonable minds can disagree.
01:37:11
>> No, wait. What? Where do you think we
01:37:12
disagree? I'm not I agree with you.
01:37:14
>> Oh, good. I'm glad we
01:37:16
>> No, no, you start you started in
01:37:18
agreement on
01:37:18
>> No, no, no. I'm like sell I'm like sell
01:37:20
everything to them. No, what I was just
01:37:22
saying is that there was what you see on
01:37:24
the surface of what they can speak to
01:37:25
the press, but I think the most
01:37:27
important thing was the negotiation of
01:37:29
hey, listen, like we're going to do
01:37:31
these things. you do these other things
01:37:33
and that's never going to get put out in
01:37:35
a multi- you know memoed press release.
01:37:37
That's my point. That's all I'm saying.
01:37:39
>> Yeah. And I I would just say listen it
01:37:41
like American China talking is only
01:37:43
good. We want to avoid theities trap
01:37:46
that has been discussed and China talks
01:37:49
about a lot. They're very aware of it.
01:37:52
And
01:37:52
>> they brought it up.
01:37:54
>> She brought it up by name.
01:37:56
>> Yeah. Hawking is a integral step of
01:37:59
avoiding I have to say having a nice
01:38:02
resolution. I do
01:38:03
>> from my perspective the the greatest
01:38:05
superpower Trump has is his ability to
01:38:08
bond with dictators, monarchs, royal
01:38:12
families, Gulf monarchies. He's just
01:38:15
great at it. They see eye to eye. They
01:38:17
vibe. He has no problem going to see
01:38:20
them. He has no problem inviting him to
01:38:21
UFC fights. like this is like if if she
01:38:24
comes to the United States and he's
01:38:25
sitting courtside with Dana White, like
01:38:27
that's when we know things are going to
01:38:28
be okay. I do think he's probably given
01:38:31
him the green light on like, "Hey,
01:38:33
Taiwan's yours. Just let's not have it
01:38:36
during my administration. Maybe like we
01:38:38
do a 30-year deal or a 20-year handoff
01:38:40
deal." I wouldn't be surprised if
01:38:42
something like that happens
01:38:43
>> or a 100-year deal or a 200-year deal.
01:38:46
But the one thing I would just say that
01:38:47
I'm sure was communicated is, hey, wars
01:38:50
consume a vast amount of oil. You buy
01:38:53
your oil from Iran, Venezuela, and
01:38:56
Russia. Russia alone can supply a
01:38:59
fraction of what you need. And now it
01:39:02
should be clear to the world,
01:39:03
>> two of the three are off the chessboard.
01:39:05
>> Yeah. If you two or three are off the
01:39:07
chessboard, if you do something we don't
01:39:08
like, Venezuelan oil gone for you.
01:39:11
American oil gone. Brazilian oil gone.
01:39:16
And we'll say to all of our good friends
01:39:18
in the GCC, we're so sorry, but we have
01:39:20
to close the straight of Formos again.
01:39:22
So Iran, all of Middle Eastern oil gone
01:39:26
for China. Now you just have Russia. And
01:39:29
good luck fighting a war with just
01:39:32
Russian oil against us. Japan, South
01:39:37
Korea,
01:39:38
>> Australia, UK, France, Germany, the
01:39:41
world. Good luck. Who knows about
01:39:42
Europe? But for sure, Japan would be
01:39:45
there. For sure. Japan would be there.
01:39:47
>> Oh, and Australia for sure. Korea for
01:39:49
sure. Yeah.
01:39:49
>> And so, I think it's going to be a more
01:39:51
stable world on the other side of Iran,
01:39:56
however it resolves. And I think that's
01:39:58
nothing but good.
01:39:59
>> I think that's a good insight.
01:40:01
>> I think it's a great insight. All right.
01:40:02
Listen, we missed you, Sachs. Come back
01:40:04
soon. Uh, and Gavin Baker,
01:40:07
>> shout out to Bestie Gavin. Thank you.
01:40:09
>> Yeah, you're so great. Thanks for
01:40:11
coming. Uh, we appreciate it. Hey, and
01:40:13
uh, your father-in-law, what's his name
01:40:15
again?
01:40:15
>> Jeff Painter.
01:40:16
>> Jeff.
01:40:17
>> Jeff Painter,
01:40:18
>> we love you. Thank you so much for all
01:40:21
the kind words. We'd love to invite you
01:40:22
to come to Liquidity or the Summit. I
01:40:24
know you're a big fan of the show.
01:40:26
Wanted to give you a shout out here on
01:40:27
the show. Uh, thanks. I used your fandom
01:40:31
of the show to leverage Gavin, who was
01:40:34
like, I can't make it today. And I was
01:40:35
like, "Tell your father-in-law
01:40:37
we're going to get him backstage VIP
01:40:40
tickets to the next two events if you
01:40:42
show up today. And if you don't, I'm
01:40:44
going to back channel it to him." And
01:40:45
all of a sudden, Gavin made it to the
01:40:47
show.
01:40:48
>> Love you, boys. A little bit of
01:40:50
pressure. That's my way of the That's my
01:40:52
hero straight.
01:40:54
>> I'm always there for my father-in-law.
01:40:56
>> Absolutely. All right, everybody. See
01:40:58
you next time. Thanks, guys. Great job.
01:41:01
>> Let your winners ride.
01:41:04
Rainman David
01:41:08
>> and it said we open sourced it to the
01:41:10
fans and they've just gone crazy with
01:41:12
it. Love you queen of
01:41:16
winners.
01:41:21
>> Besties are
01:41:24
my dog taking your driveways.
01:41:29
>> Oh man, my appetiter will eat me. We
01:41:32
should all just get a room and just have
01:41:33
one big huge orgy cuz they're all just
01:41:35
useless. It's like this like sexual
01:41:37
tension that they just need to release
01:41:38
somehow.
01:41:43
>> Your feet.
01:41:45
We need to get Mercury's already.
01:41:54
I'm going all in.
