00:00:00
there's a new AI model on the scene
00:00:01
that's smart cheap and made in China
00:00:03
it's called Deep seek and it's causing a
00:00:05
panic in Silicon Valley which is paying
00:00:06
a lot of attention and also on Wall
00:00:08
Street deep seek has reportedly
00:00:10
outperformed models from open AI meta
00:00:11
and anthropic in some tests and it
00:00:13
operates at a fraction of the cost of
00:00:15
those models using fewer high-end chips
00:00:17
this is the ones that are made by Nvidia
00:00:19
and are hard to get and the incumbents
00:00:21
have been pricing them up heavily by
00:00:24
grabbing all of them the markets are not
00:00:25
reacting well to deep seek as of this
00:00:27
recording Nvidia is down 16% Oracle is
00:00:30
down 10% Microsoft is down nearly 4%
00:00:34
obviously meta is going to be affected
00:00:36
all the others so there's a lot to talk
00:00:38
about and I've seen different analysis
00:00:40
of exactly what deep seek does Yan laon
00:00:43
from meta was making an argument that it
00:00:45
isn't as what they're re they're doing
00:00:47
sort of a cheap and dirty version then
00:00:49
it's not nearly as the stuff they're
00:00:51
doing is much more advanced by the US
00:00:53
companies uh it's currently number one
00:00:55
on Apple's uh free top apps chart uh
00:00:57
again China invading it in this country
00:01:00
in a very different way so thoughts on
00:01:02
this situation because you and I have
00:01:04
talked about this quite a bit is this
00:01:05
money ill spent by us uh companies and
00:01:09
is it being relegated to the rich
00:01:11
incumbents well first you just have to
00:01:13
temper the or put some context to the I
00:01:16
mean Nvidia is down 15 or 16% it's shed
00:01:18
something like a half a trillion dollars
00:01:19
which basically if you take out Tesla
00:01:21
it's shed today the value of the entire
00:01:23
Global automobile industry sounds Tesla
00:01:25
so this is pretty dramatic but at the
00:01:27
same time that just takes it back to its
00:01:29
valuation in October
00:01:30
and when you look at market dynamics
00:01:32
when these companies have experienced
00:01:33
these type of run-ups it is like a
00:01:35
balloon inflating Beyond its natural
00:01:38
capacity and the slightest the slightest
00:01:40
touch can pop it and so in some ways the
00:01:44
market was probably looking for an
00:01:46
excuse to take these stocks down a bit
00:01:48
and it got it because what's interesting
00:01:50
is NVIDIA will have a pretty interesting
00:01:52
argument on on Capitol Hill saying when
00:01:54
you refuse to let us sell into these
00:01:57
countries they come up with workarounds
00:01:59
and in this case work around might tank
00:02:01
the US economy and everyone's excited by
00:02:04
the fact that these models open AI
00:02:06
supposedly the models their llms cost
00:02:08
100 million to train and they're
00:02:10
claiming this thing costs and they've
00:02:12
been public it's open source cost a
00:02:13
little over 5 million to train so
00:02:16
whereas the majority of lm's and U AI
00:02:19
companies have been taking sort of this
00:02:20
Brute Force strategy where it's buy as
00:02:23
many chips as possible this is saying
00:02:26
maybe you don't need as many chips the
00:02:29
thing find it equally interesting is the
00:02:31
second order effects here and that is
00:02:33
Constellation Energy and some of these
00:02:34
nuclear stocks have skyrocketed because
00:02:37
the choke point was supposedly going to
00:02:38
be energy but now with this this model
00:02:42
which appears to have chips speaking to
00:02:44
each other in a more efficient less
00:02:46
Energy consumptive Way nuclear stocks
00:02:48
are crashing electric Constellation
00:02:51
Energy all these things that have had
00:02:52
incredible run-ups are saying wait the
00:02:54
entire supply chain or the assumptions
00:02:56
we made about the supply chain in terms
00:02:58
of the the kind of the Brute Force of
00:03:00
chips that we're going to need the
00:03:01
amount of energy it's all now coming
00:03:03
into a little bit of question but to be
00:03:05
clear the correction here is like it's
00:03:07
taking them back three months and all of
00:03:09
the stocks that have crashed quote
00:03:11
unquote crashed are are only up you know
00:03:15
70% for the year now not 98 and a lot of
00:03:18
analysts the smart analysts I've read
00:03:19
have said like every Community or any
00:03:22
sector it's going to bifurcate into the
00:03:24
cheap layer and then the high-end layer
00:03:28
which will still go hard at massive
00:03:30
Computing and massive energy and do more
00:03:31
sophisticated things and this will be
00:03:33
sort of you know everything eventually
00:03:35
goes Walmart Tiffany right and they're
00:03:38
saying this might be the Walmart and
00:03:39
it's the Chinese and they'll come up
00:03:41
with cheaper models but I it's
00:03:43
fascinating to see that basically this
00:03:46
notion this this kind of conventional
00:03:49
wisdom that you would need massive gpus
00:03:50
and massive Energy may not be um kind of
00:03:54
the written in law that we thought it
00:03:56
was going to be let me read Yan Lon
00:03:58
who's the head of meta I just read
00:03:59
recently interviewed him and you can go
00:04:01
listen to that long interview about this
00:04:02
but he's writing to the people who see
00:04:04
the performance of deep seeds and think
00:04:06
China is surpassing the US and AI you're
00:04:07
reading this wrong the correct reading
00:04:09
is open- Source models are surpassing
00:04:11
proprietary ones deep seek has profited
00:04:13
from open research in open source for
00:04:16
example pie torch and llama from meta
00:04:18
they came up with new ideas and built on
00:04:20
top of other people's work because their
00:04:21
work is published and open source
00:04:23
everyone can profit from it this is the
00:04:25
power of open research and open source
00:04:26
obviously this is the way he's talking
00:04:28
his own book that's correct I was just
00:04:30
going to make SCE yes that's correct
00:04:32
that's what I was going to say but it's
00:04:34
interesting he's having really
00:04:35
interesting arguments and he said he's
00:04:38
having a bunch of them which is just
00:04:39
interesting and one of them that he just
00:04:41
did because Gary Marcus this guy who's
00:04:42
somewhat of a a crank a little bit um
00:04:45
was saying that Congress needs to bring
00:04:46
in Zuckerberg and Lon to discuss how
00:04:48
their unilateral open sourcing decision
00:04:49
rapidly undermined the US advantage in
00:04:51
general of AI he goes an absolutely
00:04:53
hilarious take revealing the complete
00:04:55
misunderstanding of the fact that open
00:04:56
research open source accelerates
00:04:58
progress for everyone from some repet
00:05:00
claimed that deep learning was hitting a
00:05:01
wall but one of the things he just wrote
00:05:03
again cuz he's he's he's getting in
00:05:05
there very deeply major misunderstanding
00:05:07
about AI infrastructure Investments much
00:05:09
of those billions are going into
00:05:11
infrastructure for inference not
00:05:12
training running AI assistant services
00:05:14
for billions of people requires a lot of
00:05:16
compute once you put video understanding
00:05:19
reasoning large scale memory and other
00:05:20
capabilities into AI systems inference
00:05:22
costs are going to increase the only
00:05:24
real question is whether users will be
00:05:25
willing to pay enough directly or not to
00:05:27
justify capex and Opex I think that's
00:05:30
that's probably he thinks these
00:05:32
reactions are woefully unjustified and
00:05:34
at the same times he's sort of arguing
00:05:35
that they aren't right which is
00:05:37
interesting interesting interesting it's
00:05:39
just so typical to Chinese to come up
00:05:41
the entire Chinese economy was sort of
00:05:43
built on more us yeah and my guess is
00:05:47
they had a mandate or they've said all
00:05:49
right we're not going to have access to
00:05:50
the same level of high-end ships we need
00:05:54
workarounds and it's it it appears to
00:05:57
spond really interesting Innovation and
00:05:59
using open source yeah using open source
00:06:02
the I mean the scary thing I I mean in
00:06:04
typical meta fashion their llm you can
00:06:07
download a version of llama with
00:06:09
absolutely no guard rails and you can
00:06:11
you can request information on anything
00:06:15
you know the most politically correct I
00:06:17
find of them is is
00:06:20
anthropic if I start asking questions
00:06:23
about insider trading from speaker to
00:06:25
Emer Pelosi it immediately gives me all
00:06:28
these things back we cannot endorse nor
00:06:30
promote strategies around inside of
00:06:31
trading chpt kind of goes straight into
00:06:34
it and I think I think llama will say
00:06:37
well here's what you do you call your
00:06:41
cousin I find it fascinating it be
00:06:43
interesting to see what happens to the
00:06:44
stock I mean these companies have
00:06:47
already let some air out it's already
00:06:49
gone to the energy guys it'll be
00:06:51
interesting to see how the market reacts
00:06:52
is this I mean the question is and I
00:06:54
don't know the answer is this the
00:06:56
beginning of a massive correction that
00:06:58
will infect the entire NASDAQ the entire
00:07:01
S&P and quite frankly now these
00:07:03
companies I don't say become too big to
00:07:05
fail but they fail you know if they
00:07:07
sneeze the US economy is going to catch
00:07:08
a cold right now because the stock
00:07:10
market's going to crash so is this the
00:07:12
beginning of the correction we've been
00:07:14
waiting for for 15 years I mean a real
00:07:16
correction we had a mild one in 21 or is
00:07:18
this feel a little nervous I think
00:07:20
people feel I think people feel a little
00:07:21
nervous about or or and it's also kind
00:07:24
of a in a weird way an argument for free
00:07:26
trade and that is if we had let them
00:07:27
just buy Nvidia gpus would they have
00:07:30
figured out this workaround would they
00:07:32
have felt as motivated to figure out a
00:07:34
workaround or quite frankly is today one
00:07:37
of those days we're going to look back
00:07:39
when we're going to think that was a
00:07:40
buying opportunity because they're going
00:07:41
to resume their hyperscaling so I think
00:07:45
it's fascinating
