00:00:00
all right everybody welcome back to the
00:00:01
Allin podcast we've got an incredible
00:00:04
crew today don't forget to go to our
00:00:06
YouTube blah blah blah
00:00:08
subscribe and make sure you check out
00:00:10
freeberg surprise drop with his hero Ray
00:00:15
Dalia live on all platforms today how
00:00:17
did that come about for a b little
00:00:19
surprise drop just great was talking
00:00:21
with Ray about his new book which he
00:00:24
just published on how countries go broke
00:00:27
obviously which country going broke now
00:00:32
America I think he talks a lot about the
00:00:34
historical context of what's G on with
00:00:36
the death Cycles in different
00:00:38
countries and basically at the end of
00:00:41
the book he has a pretty uh I think
00:00:43
important
00:00:45
recommendation to try and get the US to
00:00:48
roughly 3% of GDP as our net deficit net
00:00:52
of all expense including interest
00:00:53
expense so that's the recommendation to
00:00:55
the administration I think it's pretty
00:00:56
timely with the change in administration
00:00:59
anyway great topics to talk through and
00:01:02
really important book awesome well done
00:01:05
and we are super delighted to have in
00:01:08
The Red Throne Travis Ken he is the
00:01:11
co-founder and CEO of cloud kitchens he
00:01:14
also uh worked in the cab business for a
00:01:16
little bit co-founder and former CEO of
00:01:18
uber and uh yeah we had a great
00:01:21
interview at the all- in Summit last
00:01:22
year and he's back up from his media
00:01:25
Hiatus he's been in the lab working on
00:01:26
cloud kitchens how you doing brother I'm
00:01:28
doing really well I I got a say just
00:01:30
like at the summit Jason I'm I'm it's an
00:01:33
honor to be in the presence of such a
00:01:37
prominent Uber investor absolutely
00:01:39
absolutely I mean finally somebody has
00:01:43
recognized my contribution the greatness
00:01:46
of JAL Absolutely I'll mention it three
00:01:48
or four times we'll be all good I'll
00:01:50
give you the props you don't have to do
00:01:51
it for yourself anymore thank you
00:01:53
appreciate it
00:01:54
appreciate let your winners
00:01:57
ride Rainman David
00:02:02
and instead we open source it to the
00:02:04
fans and they've just gone crazy with it
00:02:06
love
00:02:10
Queen give everybody a little overview
00:02:12
of cloud kitchens and the business and
00:02:15
how it's going because people are
00:02:17
obviously
00:02:18
addicted to ordering food at home and uh
00:02:22
it's it's quite a trend yeah I mean the
00:02:23
the high level for it the way to think
00:02:25
about is it's it's about the future of
00:02:27
food what does the future of food look
00:02:29
like you go well in a hundred years
00:02:32
we'll start way out there in a 100 years
00:02:34
you're going to have very high quality
00:02:37
food very low
00:02:39
cost that's incredibly convenient and
00:02:42
they going to be machines that make it
00:02:43
they going to be machines that get it to
00:02:45
you and it's going to be exactly to your
00:02:47
dietary preferences your food
00:02:50
preferences Etc and it just comes to you
00:02:52
and it's so inexpensive that it
00:02:54
approaches or has surpassed the cost of
00:02:56
going to the grocery store that's more
00:02:58
of a like a today analog
00:03:00
so you go 100 Years of course that's the
00:03:03
thing nobody's going to be making food
00:03:05
what about 20 what about 10 and so the
00:03:10
company is real estate software and
00:03:14
Robotics that's all about the future
00:03:16
food and if you can get the quality
00:03:19
there and you can get that cost down to
00:03:22
start approaching the cost of going to
00:03:24
the grocery store you do to the kitchen
00:03:26
what Uber did to the
00:03:27
car and that's the thing
00:03:31
a grind it's like lot you know bits and
00:03:33
atams in the Uber
00:03:34
world this is
00:03:37
like five times more atoms per bit this
00:03:40
is like heavyduty industrial stuff
00:03:44
probably more along the lines of like
00:03:45
you know where Elon goes and some of his
00:03:48
companies like they're super interesting
00:03:50
Tech but you got to grind out those
00:03:51
atoms do you see people actually cooking
00:03:54
in the future or does it become a
00:03:56
centralized service and is it optimized
00:03:57
to People's Health and what do you think
00:04:00
the implications to the food supply are
00:04:01
if your vision holds how do you think
00:04:03
about all those things look uh people
00:04:06
will cook in the future as a hobby I
00:04:09
sort I make a joke at the office I'm
00:04:11
like I like horses I love horses but I
00:04:13
don't ride a horse to work right and
00:04:15
it's going to be a little bit like that
00:04:17
whereas you can cook it's It's A Soulful
00:04:20
thing to do it's just very
00:04:22
human but you know it's late you know
00:04:26
Mom gets home late from the office needs
00:04:29
to get the kids you know a nutritious
00:04:31
meal she doesn't have to cook it now and
00:04:33
she doesn't she won't have to cook it
00:04:36
and she won't have to go to McDonald's
00:04:38
either it will be high quality and
00:04:41
convenient and low cost all at the same
00:04:43
time and yes dietary preference
00:04:45
everything because it'll be hyper
00:04:47
personalized like the way the internet
00:04:49
is in
00:04:50
content plus plus plus in terms of your
00:04:54
specific preferences for what you
00:04:57
want You' got these computers rocking oh
00:05:00
these robots rocking I think in Philly
00:05:02
somewhere uh in the lab where they're
00:05:04
making bows yeah I mean we're out of the
00:05:06
lab at this point we have our machine so
00:05:09
we have a machine called a bowl Builder
00:05:11
that basically makes different Cuisine
00:05:13
types with bows so like think of like
00:05:15
like sweet greens like what they yeah
00:05:18
we're not working with these Brands
00:05:19
specifically but I'll I'll just sort of
00:05:21
it's a good analogy like think of
00:05:23
Chipotle or Cava or sweet green or you
00:05:27
get the idea
00:05:29
we created test brands that were like
00:05:31
those
00:05:32
things and built the machine at the same
00:05:36
time as we were building an actual
00:05:37
restaurant and we built that restaurant
00:05:39
to prove that the machine
00:05:42
works then we have our customers now
00:05:47
touring checking out we're rolling out
00:05:49
with five customers in April that are
00:05:52
using the machine and the way it will
00:05:53
the way it's going to go down is they
00:05:56
will come into and we of course we have
00:05:59
the real estate so we have kitchens you
00:06:00
know tens of thousands of kitchens
00:06:01
around the world they will come into one
00:06:04
of our kitchens in a facility it's a
00:06:06
delivery only
00:06:07
restaurant they'll prep the food in the
00:06:09
morning and then they will leave and the
00:06:12
Machine will if you will order online
00:06:14
door dashu breats Etc they'll order
00:06:16
online the way they do build your own
00:06:19
Bowl exactly as you want
00:06:22
and the bull gets all the ingredients
00:06:25
dispensed hot or cold sauce Etc gets
00:06:30
sled the bowl goes into a bag the
00:06:33
utensils go into the bag the bag is
00:06:35
sealed and then it comes out on a
00:06:36
conveyor
00:06:37
belt and machine gets the bag it goes to
00:06:41
the front of the facility gets put into
00:06:43
a locker that Locker then is sitting
00:06:46
there door Dash re driver comes waves
00:06:48
their waves their phone with an app in
00:06:51
front of a camera and it pops open the
00:06:52
locker that has the food that you're
00:06:53
supposed to get that's so cool so like
00:06:55
if you're if you're a restaurant tour
00:06:57
you're the grind of the on man Meal
00:07:00
which is the restaurant world goes away
00:07:03
you you basically prep and that's
00:07:05
asynchronous from when people order food
00:07:08
the machine does the final assembly or
00:07:12
what's known as plating essentially like
00:07:14
do you think there's a service in the
00:07:15
future where my
00:07:18
physiology I can share that with you
00:07:21
with Cloud
00:07:22
kitchens and you guys just can always be
00:07:25
optimizing my food based on what I know
00:07:29
is good or bad for me so first what we
00:07:32
do is we serve the restaurants so what
00:07:35
happen so chamat you'll be sharing your
00:07:37
dietary preferences with uberit or door
00:07:40
Dash or sweet green or
00:07:43
somebody we like our like our customer
00:07:47
promise at our company we serve those
00:07:48
who serve others or put in another way
00:07:51
is infrastructure for better food so we
00:07:54
are the either the AWS or the Nvidia or
00:07:57
whatever you want to call it but for
00:07:58
food if that makes sense we're behind
00:08:00
the scenes we're the infrastructure and
00:08:03
so you'll give your preferences right it
00:08:05
should be a brand like then sweet greens
00:08:06
or whomever Chipotle that says hey guys
00:08:08
share with me like a yes an encrypted
00:08:11
hash of your dietary restrictions needs
00:08:15
whatever your lipid panel and I and I'll
00:08:18
customize this thing and then you enable
00:08:19
that on the back it's pretty close jth
00:08:21
right you can do authenticate your Apple
00:08:23
Health just authenticate Apple Health
00:08:25
when these bowls come off the line and
00:08:27
see how I talk it's like an assembly
00:08:28
line bows come off the line on the label
00:08:33
on the bow is how manys of every
00:08:35
ingredient is in it plus a picture of
00:08:38
what it was before we put the lid in
00:08:41
that can be sent to the person while the
00:08:43
Bull's on its way via a courier right
00:08:45
what do you think Travis about this
00:08:46
whole Maha movement and just the food
00:08:48
supply itself so then what how does that
00:08:50
change do restaurants Embrace more Farm
00:08:53
to Table stuff I think look I think what
00:08:55
like what we've SE with Supply chains in
00:08:57
a bunch of different Industries it's
00:08:59
just going to get super wired up so
00:09:02
right now we're at the point of
00:09:03
manufacturing but what happens so you go
00:09:06
okay we're doing assembly then you go
00:09:07
okay what about prep then you go further
00:09:09
upstream and you're like what about
00:09:11
supply chain like Cisco US Foods and
00:09:13
then you go further up and you're like
00:09:14
well how does how does the mechanization
00:09:16
occur on farms and in agriculture and
00:09:19
then how does that all get wired up to
00:09:21
serve the customer and sort of what
00:09:24
they're looking for so like you really
00:09:26
can know exactly what kind of wheat was
00:09:30
put into that
00:09:32
food whether it was organic for real or
00:09:35
not like what was the actual field it
00:09:37
came from things like this you could
00:09:40
imagine like really getting tied about
00:09:42
supply chain as it relates to dietary
00:09:44
stuff and as it relates to like Maha
00:09:45
like hell to the yes I mean I ordered a
00:09:50
couple different I went to the I went to
00:09:52
RFK Junior's website and they have like
00:09:55
the he has merch he has Maha merch I
00:09:58
have I have the green Maha merch hat I
00:10:00
should have worn it today I'm all about
00:10:02
it get the
00:10:03
onesie yeah that's amazing the onesie
00:10:06
was crazy your bowl Builder Friedberg
00:10:08
you tried to do this right your and we
00:10:12
had a bowl Builder 10 years ago or
00:10:14
2015 yeah
00:10:16
2016 Diego saw it he actually visited it
00:10:19
when we built it
00:10:21
and we designed the system around a
00:10:24
canister mechanism so all the food prep
00:10:26
was done in a similar sort of like com
00:10:29
AR model and then it was loaded in bulk
00:10:32
and then put into little canisters and
00:10:34
there were 30 slots in the canister
00:10:36
dispenser and then the canisters would
00:10:38
move down the device open up and and you
00:10:40
could assemble bowls with rice and beans
00:10:43
and all sort stuff the whole thing was
00:10:44
automated and we were in the process of
00:10:46
building out our first automated store
00:10:49
when I actually took a medical leave of
00:10:51
absence from it and ultimately the
00:10:54
company did not get it into production
00:10:56
but it we had great working demos and it
00:10:58
was a very
00:11:00
yeah I mean it was just definitely a
00:11:01
no-brainer that this love this you love
00:11:03
this yeah and at the time we were we
00:11:05
actually had I'll tell you guys this we
00:11:07
actually had a term sheet with
00:11:10
Chipotle this was nine years ago to
00:11:12
actually put this into Chipotle stores
00:11:14
and then we were in the early
00:11:15
conversations with sweet green at the
00:11:17
time as well and obviously Jonathan and
00:11:18
team have gone on to develop their own
00:11:20
system but you know basically you can
00:11:22
reduce so much of like qsr down to this
00:11:25
bull based system and automated as
00:11:28
Travis is doing so it's just a
00:11:29
no-brainer and it's it's certainly
00:11:31
necessary in a time when there's either
00:11:33
labor shortage or labor price inflation
00:11:36
that's causing a real issue with the
00:11:37
ability and yeah this is the original
00:11:39
automats in in New York yeah in in the
00:11:42
early 20th century I love this but yeah
00:11:44
they had a commissary behind that wall
00:11:46
and they made like plates of food you
00:11:47
put in there you put a quarter in you
00:11:48
turn the knob and you get your meal out
00:11:50
it's the classic that's the classic
00:11:52
artificial artificial intelligence right
00:11:53
this is like the Mechanical Turk thing I
00:11:56
mean look here's the thing here's the a
00:11:57
little Nuance that's super interesting
00:12:00
about Automation in qsr restaurants is
00:12:04
that they have an existing brick and
00:12:06
mortar that's built a certain way that
00:12:08
layout is meant for humans and to for
00:12:11
those humans to work in certain
00:12:12
processes an exact and very specific way
00:12:15
every square inch of that kitchen and
00:12:17
that space is is
00:12:20
dialed when you go and put a machine
00:12:22
like this in it changes the whole thing
00:12:25
and so just to get going you've got to
00:12:27
like do you've got to like if you're to
00:12:29
replace the front line in Chipotle you
00:12:31
got to you got to take out that front
00:12:32
line you got to demo it yeah you got to
00:12:34
put in a new
00:12:36
machine that's the challenge that they
00:12:38
all had and so now it's like a huge
00:12:39
amount of capex my storees down for two
00:12:41
to three months and the economics start
00:12:44
to not work and by the way I still have
00:12:46
to have humans in that brick and mortar
00:12:48
and so you know look we have a different
00:12:50
take we're in that delivery only model
00:12:53
so these are it's true infrastructure
00:12:57
for making food behind the scenes for
00:12:59
delivery so you don't have these issues
00:13:01
and of course our setup our
00:13:03
infrastructure these kitchens are
00:13:05
designed for these kinds of machines to
00:13:08
be in them and vice versa we've designed
00:13:09
the machine to be in them when we did
00:13:11
this early at ITA it was like food
00:13:14
delivery was very early we built these e
00:13:17
restaurants that were smaller frint we
00:13:18
had an 800t restaurant that was doing 3
00:13:20
million a year in revenue and it had a
00:13:23
handful of people working in it but we
00:13:24
were putting about 800 people an hour
00:13:26
during the lunch rush through that
00:13:27
restaurant ordering custom B this was by
00:13:30
one market right so one market exactly
00:13:32
and so by the way did you guys notice
00:13:34
that jcal was plugging his product there
00:13:36
in the background even though it has
00:13:38
absolutely nothing to do with what
00:13:39
Travis was saying oh welcome back to the
00:13:42
show nothing's
00:13:44
changed sax is here no one else even
00:13:46
noticed that I just heard this voice
00:13:48
from above it was the Zar of AI and
00:13:51
crypto I was like Wow send it that's all
00:13:54
sit back and listen the Zar is back ZX
00:13:57
any anecdotes you want to share about
00:13:59
life in DC how how exciting it's been in
00:14:02
the administration the first week it's
00:14:04
been amazing I mean it's hard to believe
00:14:07
it's only been a week right so you in
00:14:08
the white house or that building next to
00:14:10
it do you have an office building you
00:14:11
mean the treasury building I know
00:14:14
somebody was talking about there was a
00:14:15
building next to it or what something I
00:14:16
don't know I have an office in the uh
00:14:18
old executive office building otherwise
00:14:20
known as the Eisenhower Building and
00:14:22
then I have a pass where I can just walk
00:14:24
over to the westwing if I want to walk
00:14:26
over to it there's kind of a whole White
00:14:28
House complex
00:14:29
behind the Gates that with the West
00:14:31
Wings part of it and the Eisenhower
00:14:32
Building and there a couple other
00:14:34
buildings in that complex it's really
00:14:37
cool it is really neat to to show up for
00:14:40
work at the White House to say that's
00:14:42
awesome it's like being in a movie or
00:14:44
something or a TV show it is really cool
00:14:47
you know it's awesome any interesting
00:14:49
meetings you can talk about and I mean I
00:14:51
know we are here to today to talk about
00:14:53
deep seek but any interesting meetings
00:14:55
or anecdotes from just the Vibes and
00:14:57
walking around what's the coffee l is
00:14:59
there like a commissary you run into
00:15:01
anybody
00:15:02
interesting there is a commissary
00:15:04
actually in the White House called the
00:15:06
Navy mess and I think they're just
00:15:09
opening up for for business now that is
00:15:12
one of the cooler things you could do is
00:15:13
you can take people to to lunch at the
00:15:15
Navy mess oh look forward to it J Cal
00:15:19
just invited himself look forward to it
00:15:21
I look forward to taking jamath and
00:15:22
freeberg
00:15:25
there I'll wear my Maga hat all right
00:15:27
well let's get started you're here
00:15:29
because you uh we have a very specific
00:15:32
he's here because the world is ending
00:15:33
Jason the the Western world is ending
00:15:35
okay the Western world is ending and
00:15:37
David Sachs is going to save it but we
00:15:39
had a little bit of a freak out the last
00:15:40
week regarding this deep seek if you
00:15:43
don't know that's a Chinese AI startup
00:15:45
they released a new language model it's
00:15:47
called R1 and it's on par basically with
00:15:50
some of the best models in production in
00:15:53
the west like open AI 01 model but they
00:15:56
claim and listen you can trust CLA
00:15:59
coming out of
00:16:00
China you know for what it's worth uh
00:16:02
they claim to have done this all for $6
00:16:03
million on only 2,000 gpus for
00:16:06
comparison open AI spent reportedly 8
00:16:10
100 million to train GPT 4 which you're
00:16:12
all using now and uh Sam claims they're
00:16:14
going to spend a billion dollars
00:16:16
training GPT 5 and so that's about 7% of
00:16:19
the cost of GPT
00:16:22
4 obviously there are export
00:16:24
restrictions on Nvidia h100s to China so
00:16:28
there's a big big debate as to if they
00:16:30
actually have H1s or not and uh Monday
00:16:34
was a blood bath in the stock market
00:16:36
Nvidia had the worst day in the history
00:16:38
of the stock market in terms of total
00:16:40
dollar amount of market cap lost it was
00:16:42
down 177% which is $600 billion tsmc was
00:16:46
down arm was down Brom was down so I
00:16:48
guess everybody's asking the question
00:16:50
how did they do this did they do it and
00:16:52
then there's a bunch of debate on
00:16:55
whether they stole which is kind of Rich
00:16:57
coming from open AI which got caught
00:16:59
red-handed stealing everybody else's
00:17:01
content and now they're crying foul that
00:17:03
the Chinese stole or
00:17:05
trained did what's called distillation
00:17:08
of their model in order to build theirs
00:17:11
Sachs obviously you are the Zar of AI
00:17:16
I'm curious what your take on all this
00:17:18
is and thanks for coming well I think
00:17:21
one of the really cool things about this
00:17:22
job is just that when something like
00:17:24
this happens I get to kind of talk to
00:17:27
everyone and everyone to talk and i' I
00:17:30
feel like I've talked to maybe not
00:17:32
everyone in like all the top people in
00:17:34
AI but it feels like most of them and
00:17:37
there's definitely a lot of takes all
00:17:38
over the map on deep seek but I feel
00:17:40
like I've started to put together a
00:17:41
synthesis based on hearing from the top
00:17:44
people in the field it was a bit of a
00:17:46
freakout I mean it's rare that a model
00:17:48
release is going to be a global news
00:17:50
story or cause a trillion dollars of
00:17:52
market cap decline in in one day and so
00:17:54
it is interesting to think about like
00:17:56
why was this such a potent news story
00:17:59
and I think it's because there's two
00:18:02
things about that company that that are
00:18:03
different one is that obviously it's a
00:18:05
Chinese company rather than an American
00:18:06
company and so you have the whole China
00:18:08
versus
00:18:09
US competition and then the other is
00:18:13
it's an open- Source company or at least
00:18:15
it open source the the R1 model and so
00:18:17
you've kind of got the whole open source
00:18:19
versus closed Source debate and if you
00:18:22
take either one of those things out it
00:18:23
probably wouldn't have been such a a big
00:18:25
story but I think the synthesis of these
00:18:27
things got a lot of people's attention a
00:18:29
huge part of Tik tok's audience for
00:18:31
example is international some of them
00:18:34
like the idea that that the US may not
00:18:36
win the AI race that the US is kind of
00:18:39
getting a comeuppance here and I think
00:18:41
that fueled some of the early attention
00:18:43
on Tik Tok similarly there's a lot of
00:18:45
people who are rooting for open source
00:18:46
or they have animosity towards open
00:18:50
Ai and so they were kind of rooting for
00:18:53
this idea that oh there's this open
00:18:54
source model that's going to give away
00:18:56
what open AI has done at 12 cost so I
00:18:59
think all of these things provided fuel
00:19:01
for the story now I think the question
00:19:04
is okay what should we make of this I
00:19:06
mean I think there are things that are
00:19:07
true about the story and then things
00:19:09
that are not true or should be
00:19:12
debunked I think that let's call it true
00:19:14
thing here is
00:19:16
that if you had said to people a few
00:19:19
weeks ago that the second company to
00:19:23
release a reasoning model along the
00:19:26
lines of 01 would be a Chinese company
00:19:29
I think people would have been surprised
00:19:30
by that so I think there was a surprise
00:19:32
and just to kind of back up for people
00:19:34
you know there's there's two major kinds
00:19:36
of AI models now there's kind of the
00:19:37
base llm model like Chach D40 or the
00:19:41
Deep seek equivalent was V3 which they
00:19:43
launched a month ago and that's
00:19:44
basically like a smart PhD you ask a
00:19:47
question gives you an answer then
00:19:49
there's the new reasoning models which
00:19:51
are based on reinforcement learning sort
00:19:53
of a separate process as opposed to
00:19:56
pre-training and 01 was the first Model
00:19:59
released along those lines and you can
00:20:02
think of a reasoning model as like a
00:20:05
smart PhD who doesn't give you a snap
00:20:07
answer but actually goes off and does
00:20:08
the work you can give it a much more
00:20:10
complicated question and it'll break
00:20:12
that complicated problem into a subset
00:20:15
of smaller problems and then it'll go
00:20:18
step by step to solve the problem and
00:20:20
that's that's called Chain of Thought
00:20:22
right and so the new generation of
00:20:25
agents that are coming are based on this
00:20:26
type of idea of of Chain of Thought that
00:20:29
that an AI model can sequentially
00:20:31
perform tasks figure out much more
00:20:32
complicated problems so open AI was the
00:20:36
first to release this type of reasoning
00:20:38
model Google has a similar model they're
00:20:39
working on called Gemini 2.0 flash
00:20:42
thinking they've released kind of an
00:20:43
early prototype of this called Deep
00:20:45
research
00:20:47
1.5 anthropic has something but I don't
00:20:49
think they've released it yet so other
00:20:51
companies have similar models to 01
00:20:56
either in the works or in some sort of
00:20:57
private beta but deep seek was really
00:21:00
the next one after open AI to release
00:21:03
you know the full public version of it
00:21:06
and moreover they open sourced it and so
00:21:08
this created a pretty big splash and I
00:21:10
think it was legitimately surprising to
00:21:12
people that the next big company to put
00:21:16
out a reasoning model like this would be
00:21:18
a a Chinese company and moreover that
00:21:21
they would open source it give it away
00:21:22
for free and I think the API access is
00:21:24
something like 12th the the cost so all
00:21:28
of these really did drive the the news
00:21:30
cycle and I think for good reason
00:21:32
because I think that if you had asked
00:21:34
most people in the industry a few weeks
00:21:36
ago how far behind is China on AI models
00:21:40
they would say six to 12 months and now
00:21:44
I think they might say something more
00:21:45
like three to six months right because
00:21:47
01 was released about four months ago
00:21:50
and R1 is comparable to that so I think
00:21:53
it's definitely moved up people's time
00:21:55
frames for how close China is on on AI
00:22:00
now let's take the um we should take the
00:22:04
claim that they only did this for $6
00:22:05
million on this one I'm with Palmer
00:22:09
lucky and Brad gersner and others and I
00:22:12
think this has been pretty much
00:22:13
corroborated by everyone I've talked to
00:22:14
that that that number should be
00:22:17
debunked so first of all it's very hard
00:22:20
to
00:22:21
validate a claim about how much money
00:22:23
went into the training of this model
00:22:26
it's not something that we can imper
00:22:28
Ally discover but even if you accept it
00:22:30
at face value that $6 million was for
00:22:33
the final training run so when the media
00:22:36
is hyping up these stories saying that
00:22:39
this Chinese company did it for six
00:22:40
million and and these dumb American
00:22:42
companies did it for a billion it's not
00:22:44
an Apples to Apples comparison right I
00:22:46
mean if you were to make the Apples to
00:22:47
Apples comparison you would need to
00:22:50
compare the final training run cost by
00:22:53
Deep seek to that of open AI or
00:22:55
anthropic and what the founder of
00:22:59
anthropics said and what I think Brad
00:23:01
has said being an investor in open Ai
00:23:03
and having talked to them is that the
00:23:05
final training run cost was more in the
00:23:09
tens of millions of dollars about nine
00:23:12
or 10 months ago and so you know it's
00:23:14
not six million versus a billion okay
00:23:17
it's the billion doll number might
00:23:19
include all the hardware they've bought
00:23:20
the years of putting into it a holistic
00:23:22
number as opposed to the training number
00:23:24
yeah it's running it it's not fair to
00:23:27
compare let's call it a nuts number a
00:23:29
fully loaded number by American AI
00:23:32
companies to the final training run by
00:23:34
the Chinese company but real quick sax
00:23:37
you've got you've got an open source
00:23:40
model and they've the the white paper
00:23:42
they put out there is very specific
00:23:45
about what they did to make it and uh
00:23:48
sort of the results they got out of it I
00:23:51
don't think they give the training data
00:23:52
but you could start to stress test what
00:23:56
they've already put out there and see if
00:23:57
you can do it cheap
00:23:59
essentially like I said I think it is
00:24:00
hard to validate the number I think that
00:24:02
if let's just assume that we give them
00:24:04
credit for the 6 million number my point
00:24:05
is less that they couldn't have done it
00:24:08
but just that we need to be comparing
00:24:11
likes to likes yeah so if for example
00:24:13
you're going to look at the fully loaded
00:24:15
cost of what it took deep seek to get to
00:24:18
this point then you would need to look
00:24:21
at what has been the R&D cost to date of
00:24:25
all the models and all the experiments
00:24:27
and all the training runs they've done
00:24:29
right and the compute cluster that they
00:24:31
surely have so Dylan Patel who's leading
00:24:35
semiconductor analyst has estimated that
00:24:37
deep seek has about 50,000 Hoppers and
00:24:40
specifically he said they have about
00:24:42
10,000 h100s they have 10,000 H 800s and
00:24:47
30,000
00:24:48
h20s now the cost s sorry is they deep
00:24:51
seek or it's deep seek plus the hedge
00:24:53
fund deep seek plus the hedge fund but
00:24:55
it's the same founder right and by the
00:24:57
way that doesn't mean they did anything
00:24:58
illegal right because the h100s were
00:25:02
banned under export controls in 2022
00:25:05
than they did the h800 in 2023 but this
00:25:08
founder was very farsighted he was very
00:25:09
ahead of the curve and he was through
00:25:11
his hedge fund he was using AI to
00:25:13
basically do algorithmic trading so he
00:25:17
bought these chips a while ago in any
00:25:19
event you add up the the cost of a
00:25:22
compute cluster with 50,000 plus Hoppers
00:25:25
and it's going to be over a billion
00:25:26
dollars so this idea a that you've got
00:25:29
this Scrappy company that did it for
00:25:30
only 6 million just not true they have a
00:25:33
substantial compute
00:25:35
cluster that they use to to train their
00:25:40
models and frankly that doesn't count
00:25:42
any chips that they might have beyond
00:25:45
the 50,000 you know that they might have
00:25:47
obtained in violation of export
00:25:51
restrictions that obviously they're not
00:25:53
going to admit to and we just don't know
00:25:55
we don't really know the full extent of
00:25:57
of what they
00:25:58
have so I just think it's like worth
00:26:01
pointing that out that I think that part
00:26:03
of the story got overhyped it's hard to
00:26:05
know what's fact and what's fiction
00:26:07
everybody who's on the outside guessing
00:26:11
has their own incentive right like so if
00:26:13
you're a semi-conductor analyst that
00:26:15
effectively is massively bullish Nvidia
00:26:19
you want it to be true that it wasn't
00:26:22
possible to train on $6 million
00:26:25
obviously if you're the person that
00:26:26
makes an alternative that's that disrupt
00:26:28
you want it to be true that it was
00:26:30
trained on $6 million all of that I
00:26:33
think is all speculation the thing that
00:26:35
struck me
00:26:37
was how different their approach was and
00:26:41
TK just mentioned this but if you dig
00:26:43
into not just the original white paper
00:26:45
of deep seek but they've also published
00:26:47
some subsequent papers that have refined
00:26:49
some of the
00:26:51
details I do think that this is a case
00:26:53
and Sak you can tell me if you disagree
00:26:54
but this is a case where necessity was
00:26:56
the mother of invention so I'll give you
00:26:59
two examples where I just read these
00:27:00
things and I was like man these guys are
00:27:02
like really clever the first is as you
00:27:05
said let's let's put in a pin on whether
00:27:07
they distilled 01 which we can talk
00:27:09
about in a second but at the end of the
00:27:12
day these guys were like well how am I
00:27:14
going to do this reinforcement learning
00:27:15
thing they invented a totally different
00:27:17
algorithm there was the the Orthodoxy
00:27:20
right this thing called po that
00:27:22
everybody used and they were like no
00:27:24
we're going to use something else called
00:27:26
I think it's called grpo or something it
00:27:28
uses a lot less computer memory and it's
00:27:31
highly
00:27:32
performant so maybe they were con strain
00:27:35
sacks practically speaking by some
00:27:37
amount of compute that caused them to
00:27:39
find this which you may not have found
00:27:41
if you had just a total surplus of
00:27:44
compute availability and then the second
00:27:45
thing that was crazy is everybody is
00:27:48
used to building models and compiling
00:27:50
through Cuda which is nvidia's
00:27:52
proprietary language which I've said for
00:27:55
a couple times is their biggest moat but
00:27:57
it's also
00:27:58
the biggest threat factor for lockin and
00:28:01
these guys worked totally around Cuda
00:28:03
and they did something called PTX which
00:28:05
goes right to the bare metal and it's
00:28:07
controllable and it's effectively like
00:28:09
writing assembly now the only reason I'm
00:28:11
bringing these up is we meaning the West
00:28:14
with all the money that we've had didn't
00:28:17
come up with these
00:28:18
ideas and I think part of why we didn't
00:28:20
come up is not that we're not smart
00:28:22
enough to do it but we weren't forced to
00:28:25
because the constraints didn't exist and
00:28:27
so I just wonder
00:28:29
how we make sure we learn this principle
00:28:31
meaning when the AI company wakes up and
00:28:33
rolls out of bed and some VC gives them
00:28:36
$200
00:28:37
million maybe that's not the right
00:28:40
answer for a series A or a seed and
00:28:42
maybe the right answer is 2 million so
00:28:44
that they do these deep seek like
00:28:47
Innovations constraint makes for great
00:28:49
art what do you think uh freedberg when
00:28:51
you're looking at this well I think it
00:28:54
also enables a new class of investment
00:28:56
opportunity
00:28:58
given the low cost and the speed it
00:29:02
really highlights that maybe the
00:29:03
opportunity to create value doesn't
00:29:05
really sit at that level in the value
00:29:07
chain but further Upstream apology made
00:29:09
a comment on Twitter today that was
00:29:10
pretty funny or I think ref this about
00:29:13
the ra he's like turns out the rapper
00:29:15
may be the the mo the the mo which is
00:29:19
true at the end of the day if model
00:29:21
performance continues to improve get
00:29:23
cheaper and it's so competitive that it
00:29:25
commoditized much faster than anyone
00:29:27
even
00:29:28
thought then the value is going to be
00:29:31
created somewhere else in the value
00:29:32
chain maybe it's not the rapper maybe
00:29:35
it's the user and maybe by the way
00:29:37
here's an important point maybe it's
00:29:38
further in the economy you know when
00:29:40
electricity production took off in the
00:29:42
United States it's not like the
00:29:43
companies are making a lot of money that
00:29:45
are making all the electricity it's the
00:29:46
rest of the economy that Acres a lot of
00:29:48
the value well you're about to see a big
00:29:49
test of this because if open AI raises
00:29:52
40 billion at 340 billion that just hit
00:29:55
the wire the underwriting IC at 340
00:29:59
billion exactly what you just said
00:30:00
freeberg it is the rapper meaning chat
00:30:02
GPT is the next killer app it's getting
00:30:05
to a billion plus ma hundreds of
00:30:08
millions of da it's competing for
00:30:09
Consumer usage that's that's the model
00:30:12
that's the model is like consumer usage
00:30:14
which puts them on a on a collision
00:30:16
course with meta it's the only company
00:30:19
that could really impact that because
00:30:21
the only company right now that has
00:30:24
billions of eyeballs of daus per day and
00:30:27
who and by the way Zuck said this in his
00:30:28
earnings release he's like there's only
00:30:30
going to be one company that brings AI
00:30:33
to a billion plus people and it will be
00:30:35
us some version of that quote in his
00:30:38
earnings released yesterday and
00:30:39
Microsoft showed weakness in the in
00:30:41
their cloud and then Microsoft's down
00:30:43
six% today and you know I think it's a a
00:30:47
window for open AI to say we're going to
00:30:50
go up against meta this is it we're
00:30:52
going to be the players what you
00:30:53
everyone's kind of ignoring Google at
00:30:55
what do you guys think is happening
00:30:56
right now between open Ai and Microsoft
00:30:58
cuz if it's true that this distillation
00:31:00
thing actually happened well there's
00:31:02
only one place where you could have
00:31:03
distilled the o1 model it's on aure so
00:31:06
what the hell is going on over there
00:31:08
well and there R1 is supported on
00:31:10
explain distillation real quick yeah so
00:31:13
when you have a big large parameter
00:31:15
model the way that you get to a smaller
00:31:18
more usable model along the lines of
00:31:20
what saaks mentioned is through this
00:31:22
process called distillation where the
00:31:24
big model feeds the little model so the
00:31:26
little model is asking questions of the
00:31:28
big model and you take the answers and
00:31:30
you refine and by the way you can see
00:31:33
this Nick I sent you a clip you guys can
00:31:35
see this I mean it's there's clearly
00:31:36
distillation happening Nick can you show
00:31:38
the clip of the of the deep seek run
00:31:41
where it shows the China answer and then
00:31:43
deletes it what was Winston's job in
00:31:45
1984 right and it sort of starts to go
00:31:47
through this whole
00:31:49
summary and then the person says are
00:31:51
there any actual states that currently
00:31:53
do
00:31:53
that hold on here it goes it says North
00:31:57
Korea wait goes China and then wait
00:31:58
watch this boom so the reason why this
00:32:02
is happening is like you're seeing this
00:32:03
Chain of Thought you're seeing the
00:32:04
several layers and then it's catching it
00:32:06
after the fact so we know that this is
00:32:08
distilled from some other
00:32:10
model and my only point there it's a
00:32:12
little tongue and cheek is right now
00:32:15
when you go and use open AI you're using
00:32:18
it sitting in an Azure instance
00:32:20
somewhere right so this is Microsoft's
00:32:22
Cloud infrastructure that runs it so it
00:32:23
begs the
00:32:25
question it's not that it's 's fault
00:32:28
open ai's fault that this distillation
00:32:29
happened and I'm not trying to assign
00:32:31
blame but typically if this were to
00:32:33
happen you'd look to your cloud provider
00:32:35
and say how are you letting this happen
00:32:37
and I don't think anybody's had a good
00:32:39
answer for that well and the cloud
00:32:41
provider is hosting R1 now so they're
00:32:44
literally
00:32:45
undercutting chat GPT and open AI at the
00:32:48
same time just to clean that up they're
00:32:49
they're hosting their own copy of it
00:32:52
right because r1's been open when you
00:32:54
say they who are you who are you
00:32:55
referring to saxs Microsoft yeah
00:32:57
Microsoft is hosting their version of R1
00:32:59
which means they are actively subverting
00:33:02
their partner open Ai and pushing people
00:33:06
to a cheaper model well whatever I mean
00:33:08
look Amazon's going to host their own
00:33:09
version of R1 grock has a version of
00:33:13
R1 just red out it's open source now
00:33:16
have who has R1 on his laptop you know
00:33:18
yeah exactly yeah but if it was if it
00:33:20
was stolen as um and the IP was stolen
00:33:22
as Sam is claiming that would be like
00:33:25
You' think he'd be able to call up S and
00:33:26
say hey can you not put the stolen IP on
00:33:29
your server and promote it to everybody
00:33:30
at a lower cost it just shows Microsoft
00:33:33
has no loyalty to open AI yeah and they
00:33:35
have but you think they would have
00:33:36
loyalty they have loyalty what it would
00:33:38
take to distill
00:33:40
01 like brute
00:33:42
force it wouldn't be like oh geez I
00:33:46
can't believe it was distilled it would
00:33:48
be like such a massive number of calls
00:33:50
against an API or against something
00:33:54
something that you it wouldn't be
00:33:56
unnoticed and oh they did actually came
00:33:59
out and said they blocked some
00:34:00
suspicious activity recently yeah no but
00:34:02
they're always doing that they're always
00:34:04
that's that's constant you're always
00:34:05
doing that that's like you know the old
00:34:08
school you know go go ahead s let me let
00:34:10
me address the distillation point so I I
00:34:12
mentioned this a few days ago on on Fox
00:34:15
news that I thought it was likely or
00:34:17
possible that desolation had occurred
00:34:20
and there was some evidence for this and
00:34:21
it became like a news story and I I
00:34:23
didn't even realize that saying that
00:34:25
would be news because it's kind of an
00:34:27
Open Secret Silicon Valley everyone I
00:34:28
talked to they're doing some level of
00:34:31
distillation because you need to test
00:34:32
your model against theirs anyways yeah
00:34:34
and every single person I've talked to
00:34:37
basically has agreed that there was some
00:34:39
distillation here from open
00:34:42
AI now that doesn't mean it was the only
00:34:45
thing going on here I mean to be sure
00:34:46
the Deep seek team is very smart and
00:34:49
there were some Innovations but also
00:34:51
there was some distillation and really
00:34:54
this wasn't even a fresh news story I
00:34:57
think from the point of view of Silicon
00:34:58
Valley because a month ago we had a
00:35:01
press cycle in Silicon Valley when deep
00:35:04
seeks V3 model came out that deep seek
00:35:08
V3 was self-identifying as chat GPT when
00:35:11
you would ask it who are you like what
00:35:13
model are you five out of eight
00:35:16
times V3 would tell you that it was
00:35:18
Chachi
00:35:19
T4 and there's lots of videos and
00:35:21
examples of this online that have been
00:35:23
posted right the point is that we knew a
00:35:25
month ago that V3 had been framed on a
00:35:28
substantial amount of chat GPT output
00:35:30
obviously because V3 was
00:35:32
self-identifying as chat GPT and there's
00:35:34
two ways that that could have happened
00:35:36
so the let's call it innocent
00:35:39
explanation is that deep seek had
00:35:41
crawled the web and found lots of
00:35:43
published output from chat GPT and then
00:35:46
trained on that and that wouldn't be a
00:35:48
violation of open AI terms of service or
00:35:51
their IP or the other explanation would
00:35:54
be that they used the API from open Ai
00:35:58
and basically you know went to town yeah
00:36:01
went to town and there's no way I think
00:36:04
based on what we know to prove that one
00:36:06
way or another but I know what most
00:36:08
people think happened and at the end of
00:36:11
the day open AI can probably figure it
00:36:12
out and they they've indicated that they
00:36:14
think there was
00:36:16
some improper distillation here in the
00:36:19
financial times it says opening eye says
00:36:21
it has found evidence that Chinese
00:36:23
artificial intelligence startup DC used
00:36:25
the US company's proprietary models to
00:36:26
train its own open source compan right
00:36:28
that's what I'm referring to so they say
00:36:29
they've been very clear about this by
00:36:31
the way you have to be sympathetic I
00:36:32
think to open AI in this because if
00:36:34
you're building a startup you're trying
00:36:36
to raise money we've all gone through
00:36:38
this cycle guys where it's like there's
00:36:40
momentum we celebrate internally the
00:36:42
momentum that's what gets you the energy
00:36:45
to push your team even further and
00:36:47
harder and then all of a sudden it turns
00:36:49
out that some portion of that like it
00:36:52
Travis said it well like you're there's
00:36:53
probably a chart inside of open ai's
00:36:55
offices where you're showing how many
00:36:57
times these apis are getting hit right
00:36:59
you know how many times these end points
00:37:00
are getting hit it all looks positive
00:37:02
and then you realize that some portion
00:37:04
of it was actually bad and trying to
00:37:06
undercut your
00:37:08
value it's a hard pill to swallow and
00:37:10
then you have to course correct very
00:37:12
quickly you have to lock down this is
00:37:14
one area where security exactly we have
00:37:16
not talked about this like you have to
00:37:18
lock these models down now you have to
00:37:20
lock the endpoints down look in the
00:37:23
Biden Administration if this had
00:37:24
happened the first conversation would
00:37:26
have been we need need to kyc the people
00:37:28
that use these models and it's like what
00:37:30
are you talking about we don't kyc the
00:37:32
cloud if you're trying to use like an
00:37:34
ec2 endpoint or an S3 bucket you don't
00:37:36
have to all of a sudden prove who you
00:37:38
are you just use a credit card and go
00:37:39
that's the whole point of why
00:37:41
proliferation can happen so quickly but
00:37:43
if we take the wrong takeaways from this
00:37:46
period there's going to be a bunch of
00:37:48
people that'll clamor to like lock these
00:37:50
folks down and make Innovation go much
00:37:52
slower I don't I think that that would
00:37:53
be bad here's the other side and totally
00:37:55
agree Chim but here's the other side you
00:37:57
go through the white paper you see what
00:38:00
it is they did what they innovated on
00:38:03
yeah the science behind it the
00:38:05
thoroughness and you're like these guys
00:38:07
are badass it doesn't it does not feel
00:38:11
or sound like somebody who took
00:38:13
something just when you get through it
00:38:15
it could be that it could be that open
00:38:17
AI wrote the white paper for them just
00:38:19
putting it out there but it's real
00:38:22
Innovative I agree with that real
00:38:24
Innovation strong Tech you're like this
00:38:26
is legit I agree with that but in that
00:38:29
paper they're very hazy about where the
00:38:32
data is coming from and they're they're
00:38:35
fairly transparent about everything else
00:38:37
they did but they're not really clear
00:38:38
about the data and specifically they say
00:38:41
that to get from V3 which is the base
00:38:45
model to R1 which the reasoning model
00:38:48
they had about 800,000 samples of
00:38:51
reasoning they were quite unclear about
00:38:55
where those reasoning samples came from
00:38:57
from by the way it is remarkable that
00:38:59
you can get from a base model to an R1
00:39:01
with just 800,000 samples but this is a
00:39:03
problem like we meaning like the Western
00:39:05
AI Community we've been trudging around
00:39:08
on this path where we've been very we
00:39:11
had a very Orthodox approach the only
00:39:14
way you can do reinforcement learning is
00:39:16
through po okay but is that true and it
00:39:19
turns out that if you're like a really
00:39:20
smart team that has no other choice you
00:39:24
move away and you invent your way out of
00:39:26
it and so we have to get that example
00:39:28
too I think it's technically brilliant
00:39:30
some of the things they've done but they
00:39:32
also use constraint as a very much a
00:39:35
feature not a bug and the the Western AI
00:39:37
economy has been the opposite so far I
00:39:40
think the best part of this is the fact
00:39:42
that Sam mman was supposed to be doing
00:39:43
open source he made it a Clos Source
00:39:45
company he stole everybody's data every
00:39:48
got caught red-handed he's being sued by
00:39:50
the New York Times for all that and now
00:39:52
the Chinese have come and open sourced
00:39:54
all the stuff he stole and he's got a
00:39:56
real competitor on the original Mission
00:39:58
of what opening ey was supposed to do so
00:40:00
I have zero sympathy for him or the team
00:40:02
over there I'm glad that this is all
00:40:04
going open source it should have been
00:40:05
open source and it's better for Humanity
00:40:07
and the fact that the Chinese did it to
00:40:08
Sam wman has come up and for him
00:40:10
stealing everybody else's content that's
00:40:12
my okay there you have it but I don't
00:40:15
have strong opinions on it it's
00:40:17
hilarious does nobody see the irony in
00:40:20
this he was supposed to be doing open
00:40:23
source because Jal I will say the models
00:40:25
are closed you're right yeah there was
00:40:28
there's the lawsuit with Scarlet
00:40:29
Johansson for stealing her voice even
00:40:31
when she said no there's a real question
00:40:34
and and people have asked New York Times
00:40:35
and then there's now the question about
00:40:36
YouTube data being used to train the
00:40:38
video models so there's a lot of being
00:40:40
on their on their heels a little bit so
00:40:43
I definitely I definitely see your point
00:40:44
stealing I think yeah I think the the
00:40:48
really all the pressure right now I
00:40:49
think is on meta because I think meta
00:40:52
has to show up with the next iteration
00:40:54
of llama that beats and exceeds
00:40:58
Gemini that exceeds R1 and I think that
00:41:03
that is going to be crucial for us to
00:41:05
have a counterweight to whatever China's
00:41:07
going to put out after this and but I
00:41:09
mean jamat it's open source like does it
00:41:11
not so this so this is my point Embrace
00:41:13
and extend Embrace and extend meta has
00:41:16
to embrace and extend everything that
00:41:18
these guys have shown meaning like meta
00:41:20
is buying tens of thousands of Nvidia
00:41:23
gpus great but what did this show this
00:41:26
shows that actually Cuda high level
00:41:29
languages in general I think we've all
00:41:31
known that they suck okay and so we've
00:41:33
all been going through it thinking that
00:41:35
it's like the right thing to do deep
00:41:38
seek throws it out the window they use
00:41:40
something called PTX what does meta do
00:41:41
is critical now to understand they need
00:41:43
to embrace this stuff and this is where
00:41:46
I think again apologies to the Invidia
00:41:48
bles but it's going to create a more
00:41:51
heterogeneous environment and the reason
00:41:54
is because there's too much money and
00:41:56
risk on the line to go through a single
00:41:58
point of failure a chip a highle
00:42:01
framework to get to that chip that's
00:42:03
nuts so I think like that kind of like
00:42:05
Emperor has no close moment is upon us
00:42:08
well let me ask you another question
00:42:09
let's assume that we start the world of
00:42:11
AI today so there's no Legacy of the
00:42:13
last three years and you wake up today
00:42:16
and there's this open- Source model
00:42:18
that's 670 billion parameters you can
00:42:20
run it on your desktop computer it's
00:42:22
completely available everything's
00:42:24
completely transparent and I ask you the
00:42:26
question forget about all the big
00:42:28
companies that are involved in
00:42:29
everyone's strategy historically what's
00:42:32
the model today to build value here
00:42:34
where do you
00:42:35
build equity value as a business if
00:42:38
you're going to start a company if
00:42:40
you're going to invest as a as an
00:42:41
investor where do you go the first is
00:42:43
you have to build a shim and I think the
00:42:45
reason why a shim is really critical is
00:42:47
that there's so much entropy at the
00:42:49
model level what this should show you is
00:42:51
you can't pick any model and the problem
00:42:53
is that the people that manipulate these
00:42:55
models the machine learning engine and
00:42:57
whatnot they become too oriented to
00:42:59
understanding how to get output of high
00:43:01
quality using one thing meaning it
00:43:04
shouldn't have been the case that we
00:43:05
have Engineers that can only use sonin
00:43:08
right that's the anthropic model right
00:43:10
it shouldn't be the case that people can
00:43:12
only use open AI or people can only use
00:43:14
llama right now that is kind of what we
00:43:16
have you don't have the flexibility to
00:43:18
hot swap as models change so if you're
00:43:21
starting a company today the first
00:43:23
technical problem I would want to solve
00:43:25
for is that because tomorrow if it's R2
00:43:29
or alibaba's model or llama I would want
00:43:32
to be able to rip it out and put it back
00:43:34
in and have everything work and right
00:43:36
now we can't do that the answer to your
00:43:38
question fre is the answer your question
00:43:40
hold on the answer to your question is
00:43:42
the application layer because this is
00:43:44
all going to become storage it's like
00:43:45
YouTube being built on top of storage or
00:43:47
Uber being built on top of GPS all these
00:43:50
Innovations are being commoditized and
00:43:52
this one is happening faster than all
00:43:53
the rest do you want to be in the
00:43:54
storage business or you want to be in
00:43:55
the YouTube business do you want to be
00:43:57
in the Uber business or do you want to
00:43:58
be in the GPS chip business I mean
00:44:00
they're both decent businesses but Gavin
00:44:02
Baker came on this podcast and said this
00:44:04
the fastest deprecating asset in the
00:44:05
world was a large language model he's
00:44:07
been proven right they're not worth
00:44:09
anything they're all going to be open
00:44:10
sourced they're all going to be
00:44:11
commoditized and that's for the best of
00:44:12
humanity and now we're going to be on
00:44:14
the application Level the hardware level
00:44:16
with robots and I think that's where the
00:44:17
opportunity is Travis what do you do
00:44:19
what company do you start today if you
00:44:21
start a company today given where the
00:44:23
world is at given the open- source
00:44:25
models like what do you do oh I'm
00:44:27
getting so excited um look I I think the
00:44:31
first the first degree out is it's what
00:44:35
you got is there a rapper company okay
00:44:38
so of course maybe those companies
00:44:41
already exist and then is there a tools
00:44:43
company right so in a funny way even
00:44:46
though Facebook could be the rapper they
00:44:48
have a tools
00:44:50
business that these you know that the
00:44:54
that deep seek is basically challenging
00:44:56
going full open source and like putting
00:44:58
something out there that's really good
00:44:59
and what has to happen is fa meta has to
00:45:02
decide like we are going to embrace and
00:45:05
extend this we're going to make sure
00:45:06
that all the developers come to us that
00:45:08
all the cool applications get built here
00:45:10
so I think it's like there's a tools
00:45:12
business and then there's the rapper
00:45:14
business um and then you know look when
00:45:18
AI here's the one thing on the Nvidia
00:45:19
thing that I would counter with a little
00:45:21
bit of what's been said here is like
00:45:23
when AI gets cheap you know what's going
00:45:24
to happen guys there's going to be a lot
00:45:26
more I right I don't think I think the
00:45:30
price elasticity on this one is actually
00:45:32
positive so as the price goes down
00:45:34
that's right the revenue usage
00:45:36
everything's going to go up we through
00:45:37
the RO this is a history of tech forever
00:45:40
since like Bill Gates said I don't know
00:45:42
what to do with more than 64 kilobytes
00:45:44
of memory like you know the question is
00:45:47
did we cheap oil cheap oil in the United
00:45:49
States drove the Industrial Revolution
00:45:52
right and like when we started
00:45:53
discovering oil suddenly we were able to
00:45:55
build factories and make stuff that we
00:45:57
never imagin possible and so then you're
00:45:58
like okay AI is like you know it's going
00:46:02
to get cheap it's going to be oil but
00:46:04
it's also going to be specialized for
00:46:06
different tasks like you're going to
00:46:08
start getting into nuances of like what
00:46:11
is the invest the investor AI look like
00:46:14
what does the autonomous car AI look
00:46:16
like what does the uh Google search I'm
00:46:19
trying to figure
00:46:21
some like yeah so you could go vertical
00:46:24
and Silo siloed air quotes understand
00:46:27
what I'm saying abolutely yeah so
00:46:28
there's a a thing called Jin's Paradox
00:46:30
which kind of speaks to this concept SAA
00:46:32
actually tweeted about it which is the
00:46:35
it's an economic concept where as the
00:46:38
the cost of a particular use goes down
00:46:41
the aggregate demand for all consumption
00:46:44
of that thing goes up so the basic idea
00:46:47
is that as the price of AI gets cheaper
00:46:49
and cheaper we're going to want to use
00:46:51
more and more of it so you might
00:46:52
actually get more spending on it in the
00:46:55
aggregate that's right because more and
00:46:56
more applications will become cost feas
00:47:00
economically feasible exactly that is I
00:47:02
think a powerful argument for why
00:47:04
companies are going to want to continue
00:47:06
to innovate on Frontier models you guys
00:47:09
are taking a very strong point of view
00:47:12
that open source is definitely going to
00:47:14
win that the leading model companies are
00:47:15
all going to get commoditized and
00:47:18
therefore there'll be no return on
00:47:19
Capital and basically continue to
00:47:20
innovate on on the frontier I'm I'm not
00:47:22
sure that's
00:47:23
true you for one thing the the R1 model
00:47:27
is is basic comp to 01
00:47:30
which which open AI released four months
00:47:33
ago and was training on internally call
00:47:35
it N9 or 10 months ago so open AI is on
00:47:39
03 now its Frontier is ahead of where R1
00:47:43
is anthropic and Google I think have
00:47:46
things in the work and even meta that
00:47:49
may be ahead of where R1 is so I think
00:47:53
R1 or deep seeks done a good job being a
00:47:55
fast follower here is it's not clear
00:47:57
that this is the frontier and those
00:48:00
Frontier Model companies now having seen
00:48:03
what might have happened with
00:48:04
distillation have a pretty strong
00:48:06
incentive to make sure that doesn't
00:48:07
happen again and they're going to be
00:48:09
taking countermeasures I mean there's a
00:48:11
question of like how much you can do to
00:48:13
stop it but I think it's a little
00:48:15
premature to conclude that there's no
00:48:17
reward for being at the frontier anybody
00:48:20
uh have any other questions for Sachs
00:48:22
before we drop him off to go back to
00:48:24
serving the American people before we
00:48:26
drop him off one final point on on the
00:48:28
whole open source verus closed Source
00:48:30
look I I'm not going to take sides in
00:48:32
that but I I think that it's a mistake
00:48:35
to just view what happened here as oh
00:48:39
it's this like Plucky upstart that's
00:48:41
like doing the community huge service
00:48:43
out of the goodness of its heart you
00:48:45
know it's basically open sourcing all oh
00:48:47
they stole it they stole it it's a
00:48:49
Chinese come on you still have this huge
00:48:52
geopolitical aspect to it right and deep
00:48:55
seek is a Chinese company they trying to
00:48:57
catch up and so if you're if you're
00:48:59
behind and you're trying to catch up
00:49:00
then open source is a strategy that
00:49:02
actually really makes sense for you and
00:49:05
you know they're trying to basically
00:49:06
undercut the leading American companies
00:49:09
and I I don't think they did it with $6
00:49:11
million I mean they have massive
00:49:13
resources behind them so I think some of
00:49:16
the the pro deep seek Vibes I think are
00:49:20
they're a little bit naive you know in
00:49:22
Silicon Valley it's like that's only the
00:49:24
uh people who worked for Sam previously
00:49:25
and quit who feel that way
00:49:29
I think there's a lot of like support
00:49:30
for deep
00:49:32
sea Valley because again people think
00:49:35
that they're doing this huge service for
00:49:36
the community and I think it's a little
00:49:37
bit more self-interested that than that
00:49:39
it could be both right I mean there
00:49:41
there is a theory that they're trying to
00:49:42
undercut and neuter the lead and at the
00:49:45
same time there's a b bunch of people
00:49:47
who believe in open source and nobody
00:49:48
should control this and certainly not
00:49:49
Sam wman should be the person who
00:49:51
controls it so two things could be true
00:49:52
at the same time David thank you so much
00:49:55
for coming on we appreciate it and uh
00:49:57
thank you for all coming on your podcast
00:49:59
David we appr David I know that this is
00:50:02
and now we're going to talk about a
00:50:03
bunch of other crazy stuff gentleman
00:50:05
David yes thank you all right thanks to
00:50:07
David Sachs for coming in and uh you
00:50:10
know I guess let's open up the aperture
00:50:12
here and talk a little bit about
00:50:15
relations with China we're obviously in
00:50:17
a bit of a cold war with them we have
00:50:20
tariffs we have
00:50:22
Taiwan and then we have uh the sort of
00:50:25
trade war going on here with uh exports
00:50:27
of h100s where do we want to start
00:50:30
gentlemen and you know Travis you've got
00:50:32
some uh deep you're one of probably five
00:50:35
American entrepreneurs who ran an at
00:50:37
scale business with Uber and the DD
00:50:39
relationship in China so you have a
00:50:41
unique position of understanding
00:50:43
business in this along with maybe Tim
00:50:45
Cook and Elon are the only other two
00:50:47
people who've really had an at scale
00:50:48
business there maybe Disney they have
00:50:50
Disneyland there
00:50:51
yeah what's your take on the
00:50:52
relationship and what's going here GE
00:50:55
China how's China going to operate
00:50:57
differently than the US Travis from your
00:50:59
experience your point of view tell us a
00:51:00
little bit about the culture and
00:51:01
business ethics in China particularly as
00:51:04
it relates to AI okay so okay so look we
00:51:08
I had this thing this is I'm going back
00:51:10
almost 10 years here Uber day we're
00:51:13
running Uber
00:51:14
China
00:51:17
and I mean I cannot there's no way I
00:51:20
could express the frenetic intensity of
00:51:25
copying the that they would do on
00:51:28
everything that we would roll out in
00:51:30
China and it was so
00:51:34
epically intense that I basically had a
00:51:39
a
00:51:40
massive amount of respect for their
00:51:43
ability to copy what we did I I just
00:51:46
couldn't believe it we would do real
00:51:48
hard work make it we dial it and it
00:51:51
would be epic and it would be awesome
00:51:52
we'd roll it out and then like two weeks
00:51:54
later boom
00:51:57
they've got it a week later boom they've
00:51:59
got it and of course I use that to drive
00:52:02
our team and there's so many great
00:52:04
stories I mean we had we had like
00:52:08
400 Chinese
00:52:11
Nationals in Silicon Valley at our
00:52:13
offices in San Francisco we had a whole
00:52:15
floor for the China growth team and it
00:52:17
was primarily Chinese Nationals we had
00:52:21
Billboards on the 101 in Silicon Valley
00:52:24
in Chinese
00:52:27
Uber Billboards to join our team in
00:52:29
Chinese to to serve the Homeland right
00:52:34
um it was like an allout War it was
00:52:36
really epic it was epic and by the way
00:52:38
when you went to that floor in our
00:52:40
office you were in China like they red
00:52:43
China style like the desks were
00:52:46
literally smaller like the density of
00:52:49
the space was it was China
00:52:52
okay so but what happens is when you get
00:52:55
really really good at coping and that
00:52:58
time gets Tighter and Tighter and
00:52:59
Tighter and Tighter and Tighter you
00:53:01
eventually run out of things to
00:53:03
copy and then it flips to creativity to
00:53:07
creativity and Innovation now at the
00:53:10
beginning you you know it's sort of all
00:53:12
over the place like the kind of
00:53:14
innovation when it was new was like what
00:53:17
you know you're like really but as they
00:53:19
exercise that muscle it gets better and
00:53:22
better and better so if you want to know
00:53:24
about the future of food like online
00:53:26
food delivery you don't go to New York
00:53:29
City you go to
00:53:31
Shanghai right what's an example like of
00:53:34
like something really Innovative their
00:53:35
doesn't doesn't mwan do drone delivery
00:53:37
and stuff like here's an example if you
00:53:40
went to offices like let's say shanhai
00:53:43
um Beijing any of the major cities hongo
00:53:46
Etc the office buildings have hundreds
00:53:49
of lockers around their
00:53:52
perimeter so that everything that you
00:53:56
get whether be food or anything else but
00:53:57
especially food it's just the the
00:54:00
couriers drop them off in these Lockers
00:54:02
in the at the office buildings and then
00:54:06
there are a whole other set of people
00:54:09
that are sort of like inter office
00:54:11
Runners Runners that then bring it to
00:54:13
your
00:54:15
office as an example like and when you
00:54:17
see it you're like and it's epically
00:54:19
efficient and it's you know they're
00:54:21
taking advantage of their economics on
00:54:23
labor and things like this it wouldn't
00:54:25
exactly work that way here but a lot of
00:54:28
the Innovation you will see coming out
00:54:31
on Uber Eats or door
00:54:33
Dash like the stuff that's coming out
00:54:36
now is stuff that existed three years
00:54:38
ago four years ago in China maybe longer
00:54:41
so like eventually you cross that
00:54:43
threshold of coping and you you're
00:54:45
innovating and then you're leading and I
00:54:48
think we see that in a in in a whole
00:54:50
bunch of different places yeah here's a
00:54:52
look at these smart lockers that you can
00:54:54
see you're just available for sale when
00:54:56
when you go online but yeah these things
00:54:57
are crazy and you've experimented with
00:54:59
those as well didn't you have like a
00:55:01
commissary Concept in DTLA well look we
00:55:04
okay so we got a couple things so we
00:55:06
have in every one of our facilities and
00:55:08
we've got you know hundreds of
00:55:10
them we'll have lockers there so the The
00:55:13
Courier then waves their phone in front
00:55:15
of a camera the right Locker pops open
00:55:17
they get the food from there and they go
00:55:19
The Courier pickup is asynchronous from
00:55:22
production of food you never you don't
00:55:24
have lines anymore there's no more lines
00:55:26
which then speeds up delivery shortens
00:55:28
the amount of time shortens is reduces
00:55:30
how much money you spend on
00:55:34
couriers and we've got a whole other
00:55:37
thing this doesn't work in it probably
00:55:39
wouldn't work in China because well for
00:55:42
a lot of reasons but let me explain what
00:55:43
it is it's called picnic where if you
00:55:46
are in an office building you order food
00:55:50
you go to a website you order whatever
00:55:53
it is from a 100 different restaurants
00:55:54
those restaurants happen to be in my
00:55:55
facilities
00:55:57
and there'll be one Courier that goes to
00:55:59
one of our facilities and picks up 50
00:56:01
orders at a time brings it to an office
00:56:03
puts it there's a shelf on every floor
00:56:05
you get notified when your food arrives
00:56:08
and it arrives the same T time every day
00:56:10
and you just go to the Shelf get it on
00:56:12
your floor and dip it right back into
00:56:14
your meeting saving people time at the
00:56:17
office giving them selection on food
00:56:20
especially in food deserts but even
00:56:21
going like there's a sweet green right
00:56:23
down there in my current in my office
00:56:25
right now I could save 20 minutes by
00:56:28
just using our own service versus doing
00:56:30
that and you get at the same price
00:56:31
because the Courier economics The
00:56:33
Courier is bring is delivering 50 orders
00:56:35
at a time so Courier costs go basically
00:56:38
to zero what do we think of the um of
00:56:41
the export controls here should we chth
00:56:43
be maybe Banning more h100s or other
00:56:46
chip sets going there or is that futile
00:56:49
I don't know the answer to that and I
00:56:50
think that's I think saxs
00:56:52
and president Trump will make a good
00:56:55
decision but here here's the Curious
00:56:57
Case of the export controls Nick I sent
00:57:01
you a couple of tweets if you want to
00:57:03
just bring this up so the first thing
00:57:05
that people are claiming is that deep
00:57:07
seek is getting access to a bunch of
00:57:10
Nvidia gpus using Singapore as a back
00:57:14
door so essentially you create a
00:57:16
Singaporean shell company you place an
00:57:19
order with Nvidia Nvidia fulfills that
00:57:21
into Singapore and then the chips go
00:57:24
someplace uhoh and so there's a bunch of
00:57:27
examples where people are saying that
00:57:29
you're talking about up to a quarter of
00:57:31
all Nvidia
00:57:34
Revenue goes into
00:57:36
Singapore and the speculation right now
00:57:39
is that 100% of those then go into China
00:57:42
which is an enormous claim because
00:57:44
that's a huge amount of of nvidia's
00:57:47
Revenue now the interesting thing is if
00:57:49
you actually try to understand well
00:57:51
maybe that's not true and maybe it's
00:57:53
sitting inside of Singapore this is
00:57:55
where that kind of unravels so just to
00:57:58
be clear like Singapore is about 250 or
00:58:01
260 square miles like it's like a small
00:58:04
small place also the Tik Tok
00:58:06
headquarters
00:58:08
and I tried to find out how many data
00:58:10
centers are in Singapore and it's about
00:58:12
a 100 and so you would think that okay
00:58:14
well what does that mean 100 could mean
00:58:16
anything but then you look at the energy
00:58:19
and they publish that and all of those
00:58:21
100 data centers consume about 876 megaw
00:58:25
so these are are small data centers
00:58:28
right and the entire industry is like a
00:58:30
one and a half2 billion doll Revenue
00:58:33
business so I do think that saaks and
00:58:36
the administration are going to have to
00:58:37
dig into this and figure out what their
00:58:39
opinion should be but there is
00:58:42
clearly a ton of of these chips going
00:58:45
into
00:58:46
Singapore I don't think anybody knows
00:58:48
where they end up and the question is
00:58:51
what does America think about that and
00:58:53
why did we Implement these export
00:58:55
controls in the first place and if
00:58:56
there's a simple back door how do you
00:58:58
want to react if the US finds a path I
00:59:00
mean let's talk about like what happened
00:59:02
with sanctions in Russia and other prior
00:59:05
kind of sanctioning efforts around the
00:59:07
world but as you kind of close the the
00:59:10
floodgates and close access the
00:59:14
buyer or the receiver of those goods or
00:59:17
that Capital are going to look elsewhere
00:59:19
they're going to look to create a market
00:59:21
somewhere else and so if we do cut off
00:59:24
access to Nvidia ch we do cut off access
00:59:27
to US exports are we not kind of
00:59:30
recognizing that the second order effect
00:59:32
of that is that China will take IP that
00:59:35
they've stolen copies that they've made
00:59:36
to Travis's point and develop and build
00:59:39
out their own Fabs and they'll find ways
00:59:41
to copy the asml technology and you know
00:59:44
at the end of the day there's a lot to
00:59:45
put together and I know it's deeply
00:59:47
technically complex but if ever there
00:59:49
were a group of people in the history of
00:59:51
human civilization to pull it off it's
00:59:53
probably the modern Chinese to be able B
00:59:56
to say let's go build our own it's worse
00:59:58
than that our own infrastructure this is
00:59:59
a great point but it's worse than that
01:00:02
the models today are capable of
01:00:03
Designing chips for you that don't rely
01:00:06
on the most complicated technologies
01:00:08
that asml creates I mean look one of the
01:00:12
luckiest things that happened to Gro was
01:00:14
we designed our chip at 14 nanometer
01:00:17
which is effectively in the spectrum of
01:00:19
Technology like VHS and beta so we chose
01:00:22
a simple simple technology stack to
01:00:25
build to
01:00:27
the latest cuttingedge chips at like 2
01:00:29
nanometer that use these complicated
01:00:31
asml machines it's not clear that the
01:00:33
yield is actually that good so why would
01:00:36
you spend all that money and if China is
01:00:38
forced to engineer its way around it
01:00:40
yeah freeberg the answer to your
01:00:41
question is they'll use these models to
01:00:44
design chips that can be manufactured in
01:00:46
simple ways and they'll make simple
01:00:49
stuff so this I'm just not sure it
01:00:51
solves the problem is my point well it
01:00:53
doesn't and this is why I think like it
01:00:55
doesn't solve the real problem which is
01:00:57
how do we incentivize people in America
01:01:00
to really out engineer and out innovate
01:01:02
competition or AI I sure's in an era of
01:01:05
extraordinary abundance and that
01:01:06
abundance ultimately reduces the the
01:01:09
drive for conflict and things are better
01:01:11
off or the other version as well is that
01:01:13
China could just bear the cost as a
01:01:16
central authority of building an
01:01:19
incredibly great model right and they
01:01:22
will spend all the money and then they
01:01:24
will tell the Chinese companies you can
01:01:26
distill from this model for free because
01:01:29
we have a golden vote and a seat on your
01:01:30
board anyways which is effectively de
01:01:32
facto what happens if you get big enough
01:01:33
in China so there's that possibility as
01:01:36
well where one Central Authority Bears
01:01:38
the capex of creating something that
01:01:40
then everybody else can can draft off of
01:01:43
and let's talk a little bit about open
01:01:45
AI uh they're in Washington asking for
01:01:47
money now is that the uh is that the
01:01:49
concept now is that the our government
01:01:51
should back the rumor today was they're
01:01:53
raising 40 billion at a 340 billion doll
01:01:56
preone with MSA potentially being the
01:01:59
lead I would love Travis's read on this
01:02:01
because Travis has taken large money
01:02:03
from from Masa in the past and has been
01:02:05
through this but how does he think about
01:02:08
make this decision obviously we all know
01:02:09
and I mentioned you guys the meeting I
01:02:11
had with him last summer where he
01:02:12
basically kicked me out of the room
01:02:14
because my company's not generative AI
01:02:16
like someone said you should go meet
01:02:17
with Masa so I'm like sure I'll sit down
01:02:18
with him and start talking and he just
01:02:20
like looked at me and he's like uh this
01:02:22
is not generative AI I only do
01:02:24
generative AI I think you're company
01:02:26
will be very successful you will be very
01:02:27
successful goodbye and he was walked out
01:02:29
and that was like the end of everything
01:02:31
so great yeah that's all he's doing now
01:02:33
so this is the big BET right so okay so
01:02:36
I need to I need to bust a myth I did
01:02:38
not take money from Masa so he begged me
01:02:41
to take money for years and we did not
01:02:44
take it because he is
01:02:47
a he's uh what's the word I'm looking
01:02:50
for he's a he's a promiscuous investor
01:02:53
so once he once he invests in you you
01:02:57
should probably count on him in using
01:02:59
your information and investing in all of
01:03:00
your competitors at least that's
01:03:02
historically what he's done so I didn't
01:03:04
go there but then he just kept investing
01:03:06
in all my competitors and they kept
01:03:08
subsidizing these markets and then I'm
01:03:10
like maybe I should have just saturated
01:03:13
soaked up the money that was there so
01:03:15
the one of the things you should think
01:03:16
about like when you look at like oh is
01:03:18
open AI taking a lot of money from aasa
01:03:21
type situation is it's a little bit of
01:03:23
like a double-edged sword is if you
01:03:25
don't take that money it goes somewhere
01:03:27
else but if you do take that money just
01:03:29
know that whatever intelligence they get
01:03:32
when they go through the process of
01:03:33
giving you the money and maybe hanging
01:03:35
around the board or who knows what is
01:03:37
going to be used to do other things and
01:03:40
that is the nature of the Masa machine
01:03:42
so you're damned if you do damned if
01:03:44
you're don't but you got to pick and if
01:03:45
the money's going and it's flowing and
01:03:48
and access to Capital is a strategic
01:03:50
competitive weapon or Advantage you must
01:03:53
you must play ball now we were able to
01:03:56
we we we did stuff with the Saudis
01:03:58
before even Vision fund existed they
01:04:01
stroked a three and half billion dollar
01:04:02
check when that was like the biggest
01:04:04
thing that ever
01:04:05
happened so we were okay with not having
01:04:08
the M of money but that M of money then
01:04:10
went to all of our
01:04:11
competitors door Dash and so in this
01:04:14
open AI context Travis I mean like just
01:04:16
knowing what you know about AI is this
01:04:18
going to be a competitive Advantage for
01:04:20
Sam to raise 40 billion where does it go
01:04:24
when he's up against
01:04:26
we don't know what in China Microsoft
01:04:29
alphabet and meta well look I think this
01:04:32
goes to some of the things that shath is
01:04:34
saying which is like if if constraint is
01:04:38
the mother of invention or or whatever
01:04:40
that that euphemism is the the the
01:04:42
aphorism is if if that's the case you
01:04:45
get into a real weird spot when you get
01:04:48
over capitalized over capitalized in the
01:04:52
in the Uber model like the war was
01:04:56
subsidizing rides for market share
01:04:58
essentially being the rapper for
01:05:00
transportation and using the parlons we
01:05:01
were using earlier to in the in in this
01:05:04
discussion so it was necessary you're
01:05:07
screwed if you don't the the question is
01:05:11
do you get to this place of over
01:05:12
capitalized too big you know too
01:05:15
bureaucratic too loose too weak too soft
01:05:20
and with when you have an open source
01:05:22
model that's very smart and it's a
01:05:24
thousand flowers blooming lots of
01:05:26
innovation happening
01:05:28
everywhere could be an overwhelming
01:05:30
force uh now I think there's going to be
01:05:32
different sectors treat it different
01:05:34
ways where like going full stack in
01:05:36
certain industry sectors is going to
01:05:38
matter and then in other places having
01:05:40
like a very sort of chaotic everybody
01:05:42
does a little slice is going to be okay
01:05:44
in other places and I think we could
01:05:47
probably spend days or hundreds of
01:05:50
dozens of hours just talking about the
01:05:52
nuances there you know well it seems
01:05:54
like there's some degree of relationship
01:05:57
between the Stargate announcement with
01:05:59
Masa and Sam standing up there with
01:06:01
Larry and then Sacha showing up in the
01:06:03
conversation as well and this raise and
01:06:06
the idea that more Hardware more
01:06:08
infrastructure faster creates emote and
01:06:13
I guess that's the real thing you have
01:06:14
to believe which becomes harder to
01:06:17
believe in the context of what happened
01:06:18
in the last week I personally think that
01:06:21
these models are and I've said this for
01:06:24
a while it doesn't make sense to one
01:06:27
large do everything model this mixture
01:06:30
oferts architecture ultimately you can
01:06:32
kind think about taking a large model
01:06:34
making two copies of it and then trying
01:06:36
to shrink each model down to whatever
01:06:39
the necessary so that you run two models
01:06:41
in less frequently meaning that that
01:06:43
combination of two models uses less
01:06:46
power and takes less time and then you
01:06:49
do the same thing again and you shrink
01:06:50
it down to four and then 12 and
01:06:52
eventually you have lots of smaller
01:06:54
models some of which in some cases are
01:06:56
experts at one thing like doing
01:06:58
mathematics or reading or writing but
01:07:00
the reality is we don't know how whether
01:07:02
humans have kind of thought about the
01:07:03
world the right way that the AI May
01:07:05
resolve to having smaller expert models
01:07:07
that we don't really understand why
01:07:08
that's the expert on something but you
01:07:10
have a network of very small kind of
01:07:12
things that work together and that
01:07:13
ultimately leads to a like
01:07:15
commoditization not just in kind of
01:07:17
model cost and in development and
01:07:19
runtime but also in like what's needed
01:07:23
like do you really create much of an
01:07:24
advantage by having all these dat c key
01:07:27
this is the key point I think freeberg
01:07:28
is you're not going to get an advantage
01:07:30
by having more h100s at a certain point
01:07:32
and the actual Advantage is going to be
01:07:34
in the IP and owning content and the
01:07:36
really smart thing to do would be for
01:07:38
somebody to go buy Reddit Kora the New
01:07:40
York Times The Washington Post and
01:07:42
Disney and take all that IP and then not
01:07:45
allow other people to use it sue the
01:07:46
hell out of them every time they try I
01:07:48
take Washington Post off that list but
01:07:50
yes but I'll say New York Times comes
01:07:53
off the list too well whatever I mean
01:07:55
all those
01:07:56
are definitely going to be what would be
01:07:59
great about those is you could then like
01:08:01
a patent troll then tell anybody else
01:08:03
who's absorbed New York Times stories
01:08:05
historically or Disney and you could
01:08:07
just sue the hell out of them and then
01:08:09
you've got the best most proprietary one
01:08:12
you're describing you're describing text
01:08:14
so you're describing text content which
01:08:16
is a fraction of where this is important
01:08:18
so video I think you can recognize that
01:08:21
Google's YouTube Content Library is
01:08:23
probably 100 to 200 times larger than
01:08:25
the rest of the internet
01:08:27
combined to do it well actually Jas
01:08:31
you're such an old school copyright guy
01:08:33
you're such an old school Med guy by way
01:08:35
sorry I believe in artists and their
01:08:37
right to content yeah we've had a series
01:08:39
of conversations that I I feel very
01:08:41
confident to tell you that they do have
01:08:42
the right in in a good chunk of that
01:08:44
content not in a lot of the copyrighted
01:08:46
content that the big media companies
01:08:47
have given them but a lot of user
01:08:48
generated content they do have the right
01:08:50
and they are using it and they're
01:08:52
legally doing it and then there's the
01:08:54
separate kind of body of content which I
01:08:56
think comes for example from Tesla Tesla
01:08:58
has an extraordinary advantage that they
01:08:59
were really pressed to put cameras on
01:09:01
everything years ago and that gives them
01:09:03
this ability to build models that do
01:09:05
self-driving so I think that there's a
01:09:07
lot more data advantage that arises in
01:09:10
certain industry segments than others
01:09:11
and that's where the moat will lie and
01:09:13
that moat will allow you to actually
01:09:15
build better products that get you a
01:09:16
more persistent advantage in gathering
01:09:17
more data that's ultimately where I
01:09:19
think this resolves to it may not
01:09:21
necessarily be about who's got the
01:09:22
biggest data center Network yeah I mean
01:09:25
here the thing guys at some point the
01:09:27
amount of data becomes the long pole in
01:09:30
the tent at some point the quality of
01:09:33
the algorithms becomes a long pole in
01:09:35
the tent and more compute is not going
01:09:37
to change that we I don't think we're
01:09:40
there yet that's the one thing that
01:09:42
counters the cheap AI means more AI is
01:09:46
is there enough data and or algorithms
01:09:50
to make the more AI to make it work and
01:09:53
I do agree with the the siloing it and
01:09:55
getting expert and getting better in
01:09:57
these ways um but I think this is an
01:09:59
interesting sort of trade-off between
01:10:01
some of these these variables I got just
01:10:03
got offered 2500 bucks to put Angel my
01:10:05
book into because Harper Collins did a a
01:10:08
deal with uh Microsoft and so I'm
01:10:11
thinking 500 per year I think it's for
01:10:13
three years is the license and they just
01:10:15
did this blanket license for every book
01:10:17
they didn't look at your sales they
01:10:18
didn't look at how desirable it was it
01:10:20
was just like a blanket deal everybody
01:10:21
gets 2500 bucks per book for 3 years and
01:10:24
I think I'm going to just do it just to
01:10:26
support proper licensing so that people
01:10:28
can start going down uh this path but
01:10:30
let's get into Doge it's been a uh I
01:10:32
think we're in 10 days into this
01:10:34
Administration and um Trump formally
01:10:38
established Doge the department of
01:10:40
government efficiency in an executive
01:10:43
order apparently elon's been spending a
01:10:45
lot of time at the offices bunch of wins
01:10:49
doge is claiming on the interwebs to be
01:10:52
saving American taxpayers around a
01:10:54
billion dollars a day $3 for every
01:10:56
American every day about $1,000 a year
01:10:58
in savings for each US citizen and they
01:11:00
claim they can triple this and so for a
01:11:02
family five that would be about what
01:11:03
$155,000 a year maybe $60,000 during
01:11:06
Trump's uh second term we got $36
01:11:09
trillion in debt have fun with some
01:11:10
numbers there if you
01:11:12
like but the key announcement was a very
01:11:16
similar to the Twitter execution the
01:11:19
ability for people to resign done in a
01:11:22
very kind way eight months of severance
01:11:24
is is being offered to federal workers
01:11:28
they expect 5 to 10% of federal workers
01:11:30
to take this buyout
01:11:32
and it's
01:11:34
um I mean this could be something like a
01:11:37
hundred billion dollar in savings eight
01:11:39
months of severance um is not actually a
01:11:42
legal concept that you can do so these
01:11:44
are some sort of
01:11:46
buyouts and there's obviously some hand
01:11:48
ringing about it but I think they're off
01:11:50
to a good start they've also been
01:11:51
canceling leases as we talked about you
01:11:53
know pre-election
01:11:55
there is so much space not being used
01:12:00
that uh the federal government is
01:12:01
terminating a ton of stuff they own and
01:12:04
going to sell it and consolidating folks
01:12:06
and at the same
01:12:09
time all of this is happening everybody
01:12:13
has to return to office who wants to go
01:12:17
first here with uh you know the sort of
01:12:20
first 10 days of
01:12:21
Doge I see some eggplant emojis in the
01:12:24
group chat first 10 days how do group
01:12:27
chat what's that about I'm adding you
01:12:29
right now how are you not in the group
01:12:30
chat I'm adding you right now I'm
01:12:32
literally every time one of these hits
01:12:34
the group chat it's just hilarious
01:12:37
eggplants people are like oh my God
01:12:40
we're not burning
01:12:42
tax and and the eggplant always comes
01:12:44
from free bird first I'm outing him as
01:12:47
an egg I'm I'm a I'm a big Doge eggplant
01:12:50
guy oh so much
01:12:52
eggplant oh so freeberg tell us about
01:12:55
how much eggplant you love this there's
01:12:57
nothing that I would say is particularly
01:12:59
surprising in the first week a lot of
01:13:01
this was kind of talked about leading up
01:13:03
to the
01:13:04
inauguration VI and Elon published their
01:13:07
piece in the Wall Street Journal a
01:13:08
couple weeks ago they talked about the
01:13:11
mechanisms of action that they could
01:13:13
utilize to kind of Drive reduction in
01:13:16
cost one of which was come back to the
01:13:18
office another one of which is you know
01:13:21
giving people a buyout offer and by the
01:13:23
way the buyout offer is not new Bill
01:13:26
Clinton did the same thing during his
01:13:27
presidency yeah if you guys remember
01:13:29
when he tried to balance the budget get
01:13:31
to a surplus which he did successfully
01:13:33
and his intention was to actually reduce
01:13:35
us Death To Zero by the year 2013 and he
01:13:38
had a very specific economic and fiscal
01:13:40
plan for doing that which he put into
01:13:42
place incredible era I think we're
01:13:44
seeing them take the actions that they
01:13:46
said they would do they said they would
01:13:48
demand to employees federal employees
01:13:50
come back to the office and they assumed
01:13:52
some degree of attrition from that and
01:13:54
now the bi offer and we'll see how far
01:13:57
things go with the courts with respect
01:13:59
to their ability to stop a legislative
01:14:02
or statutorily mandated spending there's
01:14:05
a big question mark here on how much
01:14:07
Authority the executive branch has in
01:14:10
stopping spending and how much they're
01:14:13
not allowed to stop because it's
01:14:15
demanded by law it's demanded by
01:14:17
Congress and Acts or laws that have
01:14:18
passed and so that's going to be the big
01:14:21
test here over the next couple of months
01:14:23
a lot of lawsuits will fly but courts
01:14:25
will ultimately adjudicate and we'll see
01:14:28
how far the Doge intention can take
01:14:30
things and then there's a separate set
01:14:32
of efforts around legislative action
01:14:34
here there's about a $2 trillion annual
01:14:36
deficit right now in the United States
01:14:38
um federal government 2 trillion a year
01:14:40
and if you look at the the doo book on
01:14:43
why countries go broke you know there's
01:14:44
a pretty simple kind of arithmetic in
01:14:46
there which is not complicated it's just
01:14:48
at the end of the day the US needs to
01:14:51
get our federal deficit down below 3% of
01:14:54
GDP which means we've got to cut about a
01:14:56
trillion trillion one of spending if we
01:14:59
can do that then we're in kind of a more
01:15:00
economically sound place by the way a
01:15:02
really important point which is in the
01:15:04
doo interview as you cut spending
01:15:07
interest rates will come down because
01:15:09
right now there's a pretty significant
01:15:11
selloff in treasuries and a lot of risk
01:15:12
associated with the US's ability to
01:15:14
deliver um its debt obligations over the
01:15:17
next 30 Years which is why 30-year
01:15:18
treasuries are at 5% right now even
01:15:21
though the Federal Reserve is cutting
01:15:22
rates the rate on treasuries is going
01:15:25
people are still selling off treasuries
01:15:27
that also inflationary it's also
01:15:29
inflationary Dave yeah for sure and so
01:15:31
as we cut spending we also will see that
01:15:34
the intent that the there will be less
01:15:36
inflation and the US ability to pay back
01:15:39
their debt obligations over the next 30
01:15:41
Years goes up so the rates will come
01:15:42
down and so there's actually a really
01:15:44
nice kind of cyclical effect as these
01:15:46
Cuts start to come into play the rate at
01:15:48
which you can make the cuts actually
01:15:50
affects the the amount of cuts you have
01:15:51
to make the faster you make the cuts the
01:15:53
less you have to cut and that's a key
01:15:55
kind of principle going into this which
01:15:56
I think we should expect a big Whirlwind
01:15:59
of cutting in the next couple of months
01:16:00
or an attempt to the courts will
01:16:02
adjudicate what needs to be legislated
01:16:03
and then they're going to go to Congress
01:16:05
and start to try and get some of these
01:16:06
Cuts in but I will tell you once again
01:16:07
after our visit in DC last week there
01:16:10
was not a single member of Congress that
01:16:11
I spoke with who views cutting to be a
01:16:13
mandate for them in the laws that
01:16:15
they're trying to pass they all have a
01:16:17
very different kind of agenda than D
01:16:19
well look this is this is really one of
01:16:21
those interesting things where it's like
01:16:24
the difference between legislature and
01:16:26
executive branch is like doge is really
01:16:30
bringing it to life is like what powers
01:16:33
and controls does the executive branch
01:16:35
have to spend and not to spend and
01:16:39
especially to not spend when it's been
01:16:41
legislated to spend this is where the
01:16:44
action is like there's no law that says
01:16:47
you know you can give a bunch of folks
01:16:49
eight months of severance and they're
01:16:51
gone and you don't replace them there's
01:16:53
no law that says that the executive
01:16:55
branch and again I don't know the all
01:16:57
the the laws sort of the rules or laws
01:16:59
about whether you know how they go about
01:17:01
doing it but let's say presumably
01:17:03
they're doing this and there's some
01:17:05
legal backing behind it like they just
01:17:08
go and do it and now they're not
01:17:09
spending money if it was really hard to
01:17:12
hire people and they could even make it
01:17:14
harder to hire people do they do they
01:17:16
fight bureaucracy with bureaucracy that
01:17:18
it's harder to spend harder to hire
01:17:21
people harder to procure certain things
01:17:23
that you're supposed to spend money on
01:17:26
you can reduce the spend through a lot
01:17:28
of very interesting nuanced friction
01:17:31
rules that they're in control of yeah
01:17:33
some friction could slow things down
01:17:35
they're talking about putting competency
01:17:37
tests in they're talking about giving
01:17:38
people reviews and maybe they have to
01:17:40
hit some standards and the gentleman's
01:17:42
riff I mean when you force people to
01:17:43
come back to the office you're going to
01:17:44
lose 5 10% of people and 10% take the
01:17:47
buyout and now all of a sudden we're
01:17:49
saving things I mean it'll be
01:17:50
interesting to see if it's 5 or 10% on
01:17:52
RTO I mean that it could be a lot more I
01:17:55
mean what I'm hearing about these
01:17:56
buildings is that they are super super
01:17:59
empty like next level empty and
01:18:03
uh let's just say I'm really glad I
01:18:06
don't hold like I'm an owner that has a
01:18:09
bunch of leases yeah to the to the the
01:18:12
federal government right now yeah oh the
01:18:14
government and you know what the
01:18:15
interesting thing about those leases I
01:18:16
was talking to the DM at density which
01:18:18
does people counting and buildings so
01:18:19
they obviously um you know very
01:18:21
interested in that the government is
01:18:23
such a reliable
01:18:25
yeah client that they're all on one-year
01:18:27
leases so people don't you know do what
01:18:29
they do with startups we just force them
01:18:31
to do five or 10 years because they know
01:18:32
hey this company could go out of
01:18:33
business they're just like yeah yeah
01:18:35
we're just on a rolling year-over-year
01:18:36
lease so you can actually just cut these
01:18:38
it's going to flood the market Shaman
01:18:40
your thoughts on also the stopping of uh
01:18:44
because they're obviously going for it
01:18:46
they stopped all payments which is a
01:18:49
part of the Playbook I saw a Twitter up
01:18:50
close and personal which is hey let's
01:18:53
let's turn off subscriptions and see you
01:18:55
know if anybody's using these
01:18:56
subscriptions
01:18:57
basically obviously a judge got involved
01:18:59
in that but Aid going to other countries
01:19:02
you know we're just starting to look at
01:19:04
what are we actually sending to other
01:19:05
countries and for what purpose and then
01:19:07
there's a naming and shaming and maybe
01:19:10
appealing to the public through social
01:19:12
media and saying hey do you want this
01:19:13
money going here when hey we have
01:19:16
tragedies in our own country that need
01:19:17
to be solved we have healthare we have
01:19:20
houses burned down we have
01:19:21
infrastructure and so maybe you could
01:19:23
talk a little bit about heart and minds
01:19:25
and winning those and what your general
01:19:26
take is so far I think that we have to
01:19:28
remember that we're only n or 10 days
01:19:30
into Doge so the fact that we're already
01:19:32
at a billion dollars a day is really
01:19:35
incredible and there has really been no
01:19:38
discernable impact there has been a lot
01:19:41
of fissures of fake news and
01:19:43
misinformation but the real impacts have
01:19:45
been negligible To None since they
01:19:47
started making those cuts I think that
01:19:49
doge is a three layer onion so layer one
01:19:53
is the people
01:19:55
we have now given a pretty generous
01:19:58
offer to
01:20:00
folks and I think Elon said it it was
01:20:03
like basically the maximum allowed by
01:20:05
these contracts but they tried to do a
01:20:06
very good thing there the second as you
01:20:08
guys just said the second layer of the
01:20:10
onion is going to be the
01:20:12
infrastructure all the buildings all the
01:20:14
physical plants that the government owns
01:20:17
and operates that may be empty that may
01:20:19
be idle and getting them back in into
01:20:22
private hands so that they be they can
01:20:23
be repurposed that's going to save a ton
01:20:25
of money but both of them will pale in
01:20:29
comparison to the third layer of this
01:20:30
onion which is the it and the services
01:20:34
and the spend and what I mean by that is
01:20:36
when you read how the department is set
01:20:39
up at the center and nucleus of every
01:20:43
single one of these Doge teams is an
01:20:46
engineer and I think the reason is that
01:20:48
they can get into these systems of
01:20:50
record and start to trace where the
01:20:52
money is going and I think when you
01:20:54
start to on cover Through
01:20:56
forensic
01:20:58
analysis where these dollars are going
01:21:01
and how it's spent that's probably how
01:21:04
you're going to close the gap from a a
01:21:06
trillion to and I suspect to be honest
01:21:09
it could be more than two trillion
01:21:10
dollars when it's all said and done that
01:21:12
is an enormous amount of waste and it's
01:21:16
unproductive so I'm very excited for
01:21:19
what happens over this next little while
01:21:20
just the the transparency is going to be
01:21:22
incredible guys Just for kicks check the
01:21:24
out right 2009 if we took if we took
01:21:29
2019 spend right the year before Co and
01:21:33
put it up against 2024 revenues $500
01:21:36
billion Surplus wow there go that's
01:21:39
crazy versus versus the1 A5 trillion
01:21:42
dollar deficit so a two2 trillion swing
01:21:46
on like a four four yeah on a$ four4
01:21:49
trillion doll budget that's all
01:21:52
waste well a lot of we've got a trillion
01:21:55
dollars a year of interest payments now
01:21:57
I mean this is guys this is the thing
01:21:59
like there's two deflationary things
01:22:02
that we need one is Doge and two is
01:22:06
where AI is going to take us if it
01:22:08
really does its thing and that will keep
01:22:11
us in an okay spot economically but like
01:22:13
we gota this spend has to go or we're in
01:22:16
we're in sort of like we're in Greek
01:22:18
Greek territory if that makes
01:22:20
sense yeah and I think this is um the
01:22:23
popular support for this this is pretty
01:22:26
incredible I'll just go through a couple
01:22:27
of numbers with you you know looking at
01:22:31
what people agree with that Trump's
01:22:33
doing early on and what they disagree
01:22:35
with you know obviously we talked about
01:22:37
it last week chamath pardoning the
01:22:39
January 6 protesters and you know ending
01:22:43
requirements for government employees to
01:22:45
report gifts that's sort of like the
01:22:46
Supreme Court thing these are
01:22:48
tremendously unpopular and then if you
01:22:50
go and you look at downsizing the
01:22:52
federal
01:22:53
government and imposing a hiring freeze
01:22:56
and requiring all federal employees to
01:22:58
return to an office these are incredibly
01:23:00
popular and Elon tweeted these uh these
01:23:03
graphs out as well so right now you have
01:23:06
Trump at the Apex of his political
01:23:08
popularity and you have these issues
01:23:11
specifically in a very
01:23:13
polarized time as incredibly popular
01:23:17
he's also done an incredible job with
01:23:20
the Border that's another
01:23:21
consensus-based issue so Trump now has
01:23:25
downsizing the government and
01:23:26
controlling immigration and getting rid
01:23:27
of violent immigrants as incredibly
01:23:30
popular parts of his mandate and that's
01:23:33
the big win for him if you look at his
01:23:35
popularity Trump is massively more
01:23:38
popular than he was the first time round
01:23:40
he's at 49% compared to last time 44
01:23:42
he's still the historically least
01:23:44
popular president ever so my point in
01:23:47
all of this is when you see Trump doing
01:23:49
things like his meme coin or you know
01:23:53
taking on Pete Bud Jed today all that
01:23:55
kind of trump 1.0 negativity grifting
01:23:58
that's the stuff that's going to derail
01:24:00
this but the stuff that's not going to
01:24:01
derail it is focusing on the Trump 2.0
01:24:03
agenda and that is as somebody who was a
01:24:06
never Trumper as you all know in the
01:24:08
audience and now somebody who is
01:24:09
supporting him
01:24:11
relentlessly that margin that extra 10%
01:24:14
of people who support him right now is
01:24:15
me and other folks who are looking at
01:24:17
the people who put around him he has to
01:24:20
stay with the 2.0 agenda as hard as it
01:24:22
is and stay away from the Steve Bannon
01:24:24
agenda
01:24:25
and the grifting those are the things
01:24:26
that will take this all apart so that's
01:24:28
my appeal to them I told everybody I
01:24:30
give a letter grade I give them a b so
01:24:31
far could do better but pretty good less
01:24:34
of the meme coin less of the dra you
01:24:37
know we have to make sure that we're not
01:24:38
dragging dishwashers and teachers and
01:24:41
and people who've been here 20 years out
01:24:42
of the country and it's going to be a
01:24:44
very DEA important um approach here if
01:24:48
this is going to be sustained and I
01:24:50
think it's a coin toss if he will be
01:24:52
able to maintain his popularity
01:24:54
and what he did today with this like I
01:24:56
don't know if you saw the Pete budha Jed
01:24:57
he was attacking him over this tragedy
01:25:00
that's the kind of stuff people don't
01:25:01
want less of that please more of the
01:25:03
Doge that's my little rant can we talk
01:25:06
to Travis about wh now Travis can I ask
01:25:09
have you taken a production wayo yes
01:25:11
what do you think about it and do you
01:25:13
think that's the future of
01:25:14
transportation and how does Uber play
01:25:16
into the uh self-driving car business
01:25:19
now I mean look it it's funny because as
01:25:23
you guys know back in the uh Back in the
01:25:25
Day 2015 16 17 we had our own autonomous
01:25:30
vehicles out there and I remember the
01:25:32
first one of ours that I
01:25:36
took and I got in the back and all I had
01:25:39
was a stop button a big red stop button
01:25:41
that I could push if things got weird
01:25:43
yeah and uh I remember this is in
01:25:45
Pittsburgh where we had our robotics
01:25:46
Division and autonomy division at Uber
01:25:49
and I got out of that car and literally
01:25:52
it's like I got off a roller coaster
01:25:54
like my legs were I could not stand
01:25:56
straight like I was like a little wobbly
01:25:58
cuz I was so freaked out and adrenaline
01:26:01
was
01:26:02
pumping you get in a wayo today and it's
01:26:05
like you you're not even thinking twice
01:26:07
you're just like it's all good you just
01:26:09
get in you get out now part of it just
01:26:11
the normalization it's
01:26:12
like it's just working and it that
01:26:15
normalizing matters in terms of the
01:26:18
psychology around it is we're just there
01:26:20
so it just works now is it a optimized
01:26:23
experience for ride sharing no like the
01:26:26
Cyber cab is sort of the ultra sort of
01:26:29
destination for what it means to get
01:26:32
transported across the city in a vehicle
01:26:35
that is not meant for a human to drive
01:26:37
no steering wheel you know folks
01:26:39
potentially even facing each other you
01:26:41
know just a whole bunch of different
01:26:44
formats the the technology works we know
01:26:47
that there are different ways to get to
01:26:48
the technology I think the the probably
01:26:50
the most interesting thing that we
01:26:51
should be or one of the most interesting
01:26:54
things to be thinking about maybe
01:26:55
there's a few first is a cheap
01:27:00
AI makes cheap
01:27:03
autonomy okay so if as as cheap AI gets
01:27:07
out there and proliferates and gets
01:27:08
broadly distributed we should expect
01:27:10
autonomy gets easier and easier and
01:27:12
easier and you see some of the stuff
01:27:13
that's happening with Tesla and FSD
01:27:16
their new models are like I think in a
01:27:18
three-month period he they went up like
01:27:20
10x in terms of performance meaning in
01:27:23
number miles per per human intervention
01:27:26
like they're seeing you know that's the
01:27:28
thing that elon's seeing right now
01:27:29
because cheap AI cheap good AI makes
01:27:33
cheap good autonomy and that's a thing
01:27:35
we need to connect the dots on I think
01:27:37
the thing then you go one level past
01:27:39
that you're like okay there's the
01:27:41
possibility literally that autonomy just
01:27:43
gets easy and commoditized similar to
01:27:45
what's happening to AI the next part is
01:27:49
okay you get the hardware you're like
01:27:50
okay manufacturing's hard that's
01:27:51
interesting that could be a long pull in
01:27:53
the tent I think that that could be a
01:27:55
place where Tesla of course has huge
01:27:58
Advantage you then look at who are wh's
01:28:01
partners are they getting set up to do
01:28:03
the right kind of manufacturing and get
01:28:05
scale of cars out there but then there's
01:28:07
like this dark horse that nobody's
01:28:10
talking about which is it's called
01:28:15
electricity It's called Power and all
01:28:18
these vehicles are electric
01:28:20
vehicles and if you said you know I just
01:28:22
did some like quick back of the envelope
01:28:25
Cals if all of the miles in California
01:28:30
went EV ride sharing you would need to
01:28:33
double the energy capacity of California
01:28:36
right let's not even talk about what it
01:28:38
would take to double the energy capacity
01:28:39
in the grid and things like that in
01:28:41
California like let's not even go there
01:28:43
even getting 20% more 10% more is going
01:28:46
to be a
01:28:47
gargantuan 5 to 10 year
01:28:51
exercise you know look I live in you
01:28:54
know I live in LA it's a nice area in LA
01:28:57
and we have power outages all the
01:28:58
freaking time because the grid is effed
01:29:01
up and they're sort of upgrading it as
01:29:02
things break that's literally where
01:29:04
we're at in La one of the most affluent
01:29:07
neighborhoods in La that's just where we
01:29:09
are so I think the the sort of the Dark
01:29:14
Horse kind of hot take
01:29:17
is combustion
01:29:20
engine
01:29:22
AVS because I I don't know how you can
01:29:25
go fast getting AV out there really
01:29:28
really really
01:29:30
massive with the electric grid as it is
01:29:34
what do you think about regulation in
01:29:35
this regard because
01:29:36
obviously there was the
01:29:39
cruise you know a person got hit by a
01:29:42
regular car they dragged it the whole
01:29:44
thing imploded we had uh at Uber the um
01:29:48
the tragedy in Arizona where somebody
01:29:49
was playing Candy Crush when they were a
01:29:51
safety driver you know what is what is
01:29:53
your outlook on on this stuff rolls out
01:29:56
and somebody gets hurt and then that you
01:29:59
know tens of thousands of cities that
01:30:01
you brought Uber to how receptive are
01:30:04
they going to be towards this and what
01:30:05
do you think the regulatory framework
01:30:07
will be
01:30:08
like you know I think similar to how you
01:30:10
get normalized it's like you're used to
01:30:13
getting in a car it's normalized
01:30:15
psychologically and in the sort of
01:30:16
public sphere the public mindset you get
01:30:19
used to it so like we're getting to a
01:30:22
place where these vehicles are provably
01:30:24
safer than human driven Vehicles so yes
01:30:27
there are mistakes but they are just
01:30:29
provably safer and people are just
01:30:31
getting used to it and that's a big part
01:30:34
of the cycle so I think we're getting
01:30:36
out of the hysteria and we're getting
01:30:38
into like yeah it's just great like talk
01:30:42
to people who are using it and they feel
01:30:45
safer from a of course like I I I feel
01:30:48
like we're going to get in less
01:30:50
accidents but also I feel safer because
01:30:52
there's like there's less chance of like
01:30:54
an interpersonal problem that does
01:30:56
happen especially you
01:30:57
know late at night you know when people
01:31:00
are out partying and things like this
01:31:02
there's just like there is a level of
01:31:04
safety on many different aspects
01:31:07
to for the drive no it's for the yeah
01:31:10
there's like there's there's safety
01:31:12
aspects across the board sure right what
01:31:14
do you think about byd and like you sort
01:31:16
of mentioned everybody getting to
01:31:18
autonomy at the same time obviously wh
01:31:20
Mo's got the biggest lead Tesla's behind
01:31:22
them byd and about 10 other providers
01:31:24
are out there doing this does does you
01:31:27
know do 10 players get there at the same
01:31:29
time and then it's just who can
01:31:31
incorporate these into their Network and
01:31:33
what do you think of the strategy that
01:31:34
Uber's doing of hey we've got these
01:31:35
eight Partners we'll take everybody into
01:31:37
the network and we'll manage people
01:31:40
vomiting the back of cars cleaning them
01:31:41
and charging them so look I think the
01:31:44
big issue you have with anything Chinese
01:31:46
is will you be allowed to bring it in
01:31:47
the
01:31:49
US just period like you maybe kind of
01:31:52
can now what happens with teror will
01:31:54
there be blocks and bringing this kind
01:31:55
of Technology into the US what happens
01:31:58
there I think that's a whole thing the
01:32:00
bet that Uber makes is
01:32:02
that whether consciously or
01:32:04
subconsciously it's like will AI will
01:32:07
cheap democratized AI happen and if so
01:32:11
does that make cheap democratized
01:32:13
autonomy then you've got to line up your
01:32:15
physical sort of Hardware Partners car
01:32:18
manufacturers then you've got to say
01:32:20
okay is the electricity where it's at
01:32:21
and are there other bets to make to make
01:32:23
sure
01:32:24
that I can charge my cars so like there
01:32:27
is a huge real estate play here and
01:32:29
Fleet Management play of like how do I
01:32:34
Electrify these plots of land known as
01:32:36
parking lots and also set them up so
01:32:40
that robots can clean cars in sort of a
01:32:43
very very efficient way there's like a
01:32:46
whole when you we talk that's super
01:32:48
interesting Travis that's like it's
01:32:50
almost like the idea that we all talk
01:32:52
about today is Data Centers and data
01:32:53
centers need their own power substation
01:32:56
in order to meet the Power demands but
01:32:57
if if we do see a world of Robotics
01:33:00
automation generally and we've got these
01:33:02
kind of moving robotic systems in our
01:33:05
world they need to have a similar sort
01:33:07
of like power demand met that probably
01:33:10
looks like hey they all go into their
01:33:12
their recharge building and they get
01:33:14
recharged whether they're a car or a
01:33:16
humanoid robot or a food delivery robot
01:33:18
on the sidewalk or whatever or Dr and
01:33:21
they just kind of get recharged huh
01:33:22
robots need actuators
01:33:24
you know what you need for an actuator a
01:33:26
permanent magnet you know what you need
01:33:27
for a permanent magnet rare Earths who's
01:33:30
the rare earth king X China
01:33:34
Greenland Greenland let's go so so guys
01:33:39
I think there is a there is a there's a
01:33:42
couple interesting things one of them is
01:33:44
going to be how are these companies
01:33:46
thinking about real estate electrifying
01:33:48
that real estate in urban environments
01:33:51
and robotizing that real estate so that
01:33:53
they can do the servicing maintenance
01:33:55
Etc look I guess it could be manual for
01:33:56
a while can I can I put you on the spot
01:33:58
just go one level above it because merge
01:34:00
the last two concepts together you
01:34:02
talked about we talked about the federal
01:34:04
government Doge Etc isn't there the
01:34:06
potential for just a complete surplus of
01:34:10
physical inventory that exists in
01:34:12
America oh yeah okay so big time so what
01:34:15
does that mean for commercial re like
01:34:17
how do you like navigate around that
01:34:19
because you got to evade the falling
01:34:20
knives first so okay so let's just just
01:34:23
go down ride sharing Lane it's
01:34:25
autonomous ride sharing Lane you go down
01:34:27
that lane car ownership which is already
01:34:30
dropping drops Like a Knife all the way
01:34:33
down and there's this thing in cities
01:34:36
which takes up 20 to 30% of all the land
01:34:38
it's called parking is no longer
01:34:40
necessary because cars are getting
01:34:41
utiliz the cars that exist on the roads
01:34:43
are getting utilized 15x more than they
01:34:45
were before per car so you need
01:34:50
hypothetically 115th the number of cars
01:34:52
maybe you could say fth or 11th because
01:34:54
you want to be able to Surge to like
01:34:57
rush hour and things like that it
01:34:58
depends on what kind of car pooling and
01:35:00
things like this are going on let's just
01:35:01
call it 10x fewer cars on10th the land
01:35:06
necessary for parking at least on Tenth
01:35:09
like maybe it's less than that okay so
01:35:11
now you're opening up you're opening up
01:35:15
20% of the land in a city that just goes
01:35:19
fallow but what what should we do with
01:35:21
that and is there a demand for that land
01:35:24
well look I mean maybe it's the should
01:35:26
it be housing you know like and then
01:35:28
don't we have to re-evaluate all of the
01:35:31
city planning today because City
01:35:33
Planning today to your point Works
01:35:34
backwards from all these constraints
01:35:36
that are 1.0 constraints like here's the
01:35:38
traffic flow here the traffic patterns
01:35:41
those don't exist theoretically anymore
01:35:42
or they would exist in a totally
01:35:44
different way right yeah I mean we've
01:35:45
got a there's like a massive amount of
01:35:47
creativity to say what can I do with
01:35:49
that land at with a high
01:35:52
Roi right like some people are like
01:35:54
you're going to have Farms you know uh
01:35:58
hydroponic farms in urban environments
01:36:00
I'm like uh you know that's not a bad
01:36:02
idea if you want to have Farm to Table
01:36:04
healthy food it's literally Farm to
01:36:06
Table it's like a mile away from you
01:36:07
yeah so there's some interesting ideas
01:36:10
the land price has to really come
01:36:11
crashing down and there's interesting
01:36:13
ramifications if it were to do that you
01:36:15
could imagine that's that's what I
01:36:16
wanted you to say not to try to get you
01:36:18
there but that winess well that seems
01:36:21
like the crazy thing that nobody is
01:36:22
thinking about which is in this push
01:36:24
this physical built inventory has so
01:36:27
much value built up in the 401ks of of
01:36:31
individuals to the balance sheets of
01:36:33
huge Pension funds but that value is
01:36:36
could be very different right but the
01:36:37
crazy part is is it could just be
01:36:40
electricity production and electric
01:36:42
capacity on the grid could be the gating
01:36:45
factor that makes it a slow
01:36:47
burn potentially I'm just riffing here
01:36:50
guys right right right right right makes
01:36:51
total sense and if want to see what
01:36:54
happens when you have like unlimited
01:36:55
land if you live in Austin and you see
01:36:58
the distance between San Antonio Houston
01:37:00
and Dallas and Austin in that triangle
01:37:03
you know you get 30 minutes outside of
01:37:04
the city centers there's just unlimited
01:37:06
land and there's less regulation and you
01:37:08
know what's happened housing prices and
01:37:09
rents have come down two or three years
01:37:10
in a row so this could happen in other
01:37:12
major cities and if Doge has less
01:37:14
regulation you can build more it could
01:37:16
be amazing for Americans to actually be
01:37:18
able to afford homes again and maybe
01:37:22
convert some of this space you go energy
01:37:24
storage electric grid upgrades sort of
01:37:29
modular energy capacity upgrades like
01:37:33
and production these are this is going
01:37:35
to be very very important right now if
01:37:38
you want to I we do this all the time we
01:37:40
have of course facilities all over every
01:37:42
major city in the US and really around
01:37:45
the
01:37:46
world utility upgrades is the long pull
01:37:49
in the tent in in construction
01:37:52
development
01:37:54
in a lot of our cities not all cities
01:37:55
but in a lot of our cities the FED um
01:37:58
held rates they're getting close to the
01:38:00
goal of 2% I guess we're at 2.4
01:38:03
2.9% in terms of inflation any thoughts
01:38:06
on uh where we're at with the FED
01:38:08
deciding to not cut and just uh you put
01:38:11
it on the docket here jamat any any
01:38:13
wider thoughts there I would just say
01:38:14
that the long end of the yield curve is
01:38:17
basically telling us that there's a
01:38:19
still a chance for inflation so I think
01:38:21
that the the question is these next 30
01:38:24
or 60 days from the administration I
01:38:27
think are basically they're they're
01:38:28
critical and I think if if Doge gets to
01:38:31
the three billion a day number quicker
01:38:33
than people thought there's going to be
01:38:35
a lot of room for I think the president
01:38:39
to make a very valid argument that rates
01:38:42
are too high for where they are and that
01:38:44
we're going to be able to have a lot
01:38:46
more cost
01:38:48
control in the expenses which means that
01:38:50
they'll be less need to spend it doesn't
01:38:53
solve
01:38:54
the problem that Yellen created Yellen
01:38:56
and Biden on the way out the door the
01:38:58
biggest problem was that they put
01:39:00
America in this very difficult position
01:39:03
because they issued so much short-term
01:39:05
paper that is extremely expensive and as
01:39:07
all of that rolls off we have to go and
01:39:09
finance a ton of this debt at now 5% so
01:39:15
it's still nearly 30% of of the debt is
01:39:18
going to get refinanced this year and
01:39:20
then it's like what are these auctions
01:39:22
going to look like guys this is the
01:39:23
thing we we all got to bre the last
01:39:25
auction barely had 2x coverage and I
01:39:27
think that that could take a lot of the
01:39:28
energy out of the market watch the doo
01:39:30
interview because this this is exactly
01:39:32
the topic he covers you know as we end
01:39:34
up needing to refinance this debt the
01:39:36
rates climb the appetite isn't there and
01:39:38
it becomes a
01:39:39
spiral that's why we have to cut fast in
01:39:42
terms of the deficit to basically
01:39:45
attract the market now you know the
01:39:47
markets moved a little bit right so on
01:39:49
January
01:39:50
13th the 30-year treasury peaked at
01:39:53
exactly 5% and it's come down today it's
01:39:56
at 4.77 so a little bit of relief since
01:40:00
that that Peak as as kind of the
01:40:02
administration's gone into office and
01:40:03
actually taken action but as more of
01:40:05
this action is realized if people do
01:40:09
appreciate and doge is successful and
01:40:11
the Court's adjudication does allow
01:40:13
reduction in spending which I think is
01:40:15
the intention I think we could see this
01:40:17
rate drop from 478 much more
01:40:19
significantly than where it is and
01:40:21
that'll create a great deal of and David
01:40:23
it's like it either does that or it
01:40:26
really really
01:40:27
doesn't or ites the exact super nasty
01:40:32
really bad that's right I got a text
01:40:33
from someone who is pretty senior in
01:40:36
capital markets thinks this is going to
01:40:38
go to five and a half% before it goes
01:40:40
down so they think that there's going to
01:40:42
be a little bit more of a turbulent run
01:40:45
ahead but it's like but the thing is
01:40:46
it's like that whole thing of like it's
01:40:48
going to get to five and a half before
01:40:49
it comes down it's like it spirals on
01:40:51
itself it's like you got to print money
01:40:54
to then get to that place and then the
01:40:56
printing drives it for you know you get
01:40:58
to that spiral the problem is if we go
01:41:00
to 5 a half% that's not 80 basis points
01:41:03
what you really need to think about is
01:41:04
the total tonnage of actual dollars that
01:41:06
need to get repaid back and if you look
01:41:09
backwards that's effectively like 10%
01:41:11
rates from 2000 could you imagine what
01:41:13
the economy that's right would have done
01:41:15
if you had brought rates to 10 11% 20
01:41:18
years ago it would have crippled the
01:41:19
economy so we don't have a lot of room
01:41:22
here where you can walk rates up to 5
01:41:24
half 6% without a lot of things starting
01:41:27
to break this is why I actually think
01:41:28
Doge will be successful because as
01:41:30
people internalize all of these things
01:41:33
where every single Congress person
01:41:35
freeberg that may have wanted their own
01:41:37
benefit for their Community they'll have
01:41:39
to take a step back because the broader
01:41:41
optimization for America just needs to
01:41:44
take priorities sham it just doesn't
01:41:45
work like that man like my thing is like
01:41:47
I like I agree with the notion but I
01:41:49
just don't believe that any individual
01:41:51
Congress person will take responsibility
01:41:53
in this way no they won't they won't but
01:41:55
the question is can they block it yeah
01:41:57
but or put another way again the
01:41:59
executive
01:42:01
branch can slow roll spend in a lot of
01:42:06
different ways except you cannot with
01:42:09
Medicare and Social Security
01:42:10
discretionary spending is like 20% the
01:42:14
mandatory spending Social
01:42:16
Security Medicare Medicaid these are the
01:42:20
the larger outlay and this where where
01:42:23
we come back to the fact that this will
01:42:26
never I hear you get addressed until it
01:42:29
has to be because of the political
01:42:30
suicide that arises I just think there's
01:42:32
this is where I think elon's Fame can be
01:42:35
helpful and I mean very specifically
01:42:37
this following idea you know that famous
01:42:39
Sputnik comment
01:42:41
where NASA spent millions of dollars
01:42:43
trying to engineer a pen that could
01:42:45
write upside down and it turned out that
01:42:47
in Sputnik the Russians just took a
01:42:50
pencil that is what we need to do to the
01:42:52
US government because I suspect even
01:42:54
though there's a lot of mandated spend
01:42:56
the real question that nobody knows the
01:42:57
answer to Is is that spend useful so
01:43:00
even though it's appropriated by
01:43:02
Congress there has to be a feedback loop
01:43:05
that says you can just use a pencil you
01:43:07
don't need the upside down writing pen
01:43:09
and I think that if there's anybody that
01:43:11
can broadcast that to the world it's him
01:43:14
and this is where I think Trump gets
01:43:15
enormous leverage by having Elon being
01:43:18
the westwing that nobody else could give
01:43:19
him the rest of us would just be
01:43:21
chirping into the darkness yeah this is
01:43:23
the naming and shaming of government
01:43:24
waste that's actually going to work and
01:43:26
the Doge account on Twitter is doing it
01:43:28
they're basically saying hey we're
01:43:30
giving foreign aid for this project for
01:43:32
that project is it going to be perfect
01:43:33
every time no but you show an empty
01:43:35
office space you show people not coming
01:43:37
to work you show people wasting money
01:43:39
well yeah if that's even real you know
01:43:41
there's going to be a bunch of you know
01:43:43
back and forth here but overall if you
01:43:46
keep naming and shaming each of these
01:43:47
projects and then you know they were
01:43:49
talking about blockchain or whatever and
01:43:50
supposed is a Report Elon is at like the
01:43:52
Govern building working on leases at the
01:43:55
moment like this stuff is going to be
01:43:57
extraordinar popular because you can
01:43:59
just take the number of 330 million
01:44:02
Americans and whatever you just saved
01:44:04
you can just divide it by that number
01:44:06
and tell every American how much they
01:44:08
just paid less in taxes or how much they
01:44:10
just saved individually the naming
01:44:12
shaming and doing the back of the
01:44:14
envelope math for every American is
01:44:16
going to work do we want to wrap maybe a
01:44:18
little bit on this tragedy in DC okay
01:44:21
what are your thoughts uh we were
01:44:22
talking with our friends Dayton who is
01:44:24
very involved in aviation and um he's
01:44:26
got a lot of blog posts he's done
01:44:28
recently and he's got a company he
01:44:29
invested in to do uh pilot training I'll
01:44:33
share two things one is anonymous it's
01:44:35
from friend of mine gave it to me and
01:44:38
said I could share it who's a commercial
01:44:41
pilot and he and and I posted this so
01:44:43
I'll just read it honesty DCA is the
01:44:45
sketchiest airport we fly into I feel
01:44:48
like the controllers there play fast and
01:44:50
loose hence the periodic Runway
01:44:51
incursions I've said to every first
01:44:53
officer in my threat briefings that we
01:44:55
both need to be on red alert at all
01:44:57
times there DCA calls out Hilo traffic
01:45:00
helicopter traffic and vice versa all
01:45:03
the time but it's borderline impossible
01:45:04
to see them when you're bombing along at
01:45:06
150 mph I mean that's from a pilot that
01:45:10
is not I don't think he has any
01:45:11
incentive to sugarcoat things and then I
01:45:14
just wanted to read a message from Brian
01:45:15
uto who's the CEO of whisk who's
01:45:19
building a lot of these aut autonomous
01:45:21
systems he said said first autot
01:45:24
trffic Collision of voida systems do
01:45:27
exist right now these aircraft will not
01:45:30
take control from the pilot to save the
01:45:33
aircraft even if software and systems on
01:45:35
the aircraft know that it's going to
01:45:38
collide that's the bti flip that needs
01:45:40
to happen in aviation automation can
01:45:43
actually kick in and take over even in
01:45:47
piloted aircraft to prevent a crash
01:45:50
that's the minimum of where we need to
01:45:52
go some fighter jets have something
01:45:54
called automatic ground collision
01:45:55
avoidance systems that do exactly this
01:45:57
when fighter pilots pass out and it's
01:46:00
possible for commercial and then the
01:46:02
second he said is we need to have better
01:46:03
ATC Air Traffic Control software and
01:46:07
automation right now we use VHF radio
01:46:10
communications for safety and for
01:46:13
critical instructions and that's kind of
01:46:15
insane we should be using data links Etc
01:46:18
the whole ATC system runs on 1960s
01:46:21
technology they deserve better software
01:46:24
and Automation in the control Towers
01:46:26
it's totally ripe for change the problem
01:46:29
is that attempts at reform have
01:46:31
failed so I just wanted you guys to have
01:46:34
that one from this commercial pilot and
01:46:36
then two from Brian uto who I I think
01:46:38
understands this issue really well
01:46:39
there's so much opportunity here to make
01:46:41
this better this should have never
01:46:42
happened our other friend Sky Dayton has
01:46:44
been pushing really hard for the US
01:46:47
government to do Advanced pilot training
01:46:49
one of the things that he says
01:46:51
constantly is just that a lot of the
01:46:52
push back is just Union rhetoric around
01:46:55
what they perceive the right thing for
01:46:57
their constituency is and hopefully this
01:47:00
starts this conversation because I think
01:47:02
guys like Sky guys like Brian are
01:47:05
working on this next level of autonomous
01:47:08
solution that can just make flying
01:47:11
totally totally safe beyond what it was
01:47:13
the crazy stat is that we haven't had a
01:47:15
commercial airline disaster in the
01:47:16
United States in almost 25 years isn't
01:47:18
that inredible 15 yeah it it's looking
01:47:21
like pilot era here and it's there seems
01:47:23
to be some question of why these Apaches
01:47:26
are flying around this really crowded
01:47:28
airspace and it seems like they're
01:47:29
shuttling you know politicians around
01:47:32
and maybe that's not the best idea in
01:47:34
this really dense area as your pilot
01:47:37
friend was referring to jamath so God
01:47:40
thoughts and prayers and all that stuff
01:47:42
for the um for the families of the
01:47:44
people who died it's just terrible
01:47:45
tragedy terrible tragedy yeah it's just
01:47:48
this really just this is an area to
01:47:50
invest money and use the private private
01:47:53
sector and all this incredible
01:47:54
Innovation that's available to upgrade
01:47:56
these systems and infrastructure this
01:47:58
has been another amazing episode of the
01:47:59
all-in podcast thanks Travis for joining
01:48:01
us thankar for coming a lot of fun guys
01:48:04
first time this is my first time on a
01:48:05
podcast ever yes right in you come back
01:48:09
anytime you were great man appr it
01:48:12
apprciate very based is going to like it
01:48:14
tell us what you think and we'll see you
01:48:16
all next time love you boys
01:48:19
byebye let your winners ride
01:48:22
Rainman
01:48:26
David and instead we open source it to
01:48:28
the fans and they've just gone crazy
01:48:30
[Music]
01:48:39
with
01:48:41
besties that's my dog taking
01:48:44
[Music]
01:48:47
driveway oh
01:48:49
man we should all just get a room and
01:48:51
just have one big Georgie cuz they're
01:48:53
all this useless it's like this like
01:48:55
sexual tension that they just need to
01:48:56
release
01:49:03
somehow we need to get
01:49:07
[Music]
01:49:12
mer I'm going in