00:00:00
one of the really cool things about this
00:00:01
job is just that when something like
00:00:03
this happens I get to kind of talk to
00:00:05
everyone and everyone wants to talk and
00:00:08
i' I feel like I've talked to maybe not
00:00:10
everyone in like all the top people in
00:00:12
AI but it feels like most of them and
00:00:15
there's definitely a lot of takes all
00:00:16
over the map on deep seek but I feel
00:00:18
like I've started to put together a
00:00:20
synthesis based on hearing from the top
00:00:23
people in the field it was a bit of a
00:00:25
freakout I mean it's rare that a model
00:00:27
release is going to be a global news
00:00:28
story or called a trillion dollars of
00:00:31
market cap decline in in one day and so
00:00:33
it is interesting to think about like
00:00:34
why was this such a potent news story
00:00:37
and I think it's because there's two
00:00:40
things about that company that that are
00:00:42
different one is that obviously it's a
00:00:43
Chinese company rather than an American
00:00:45
company and so you have the whole China
00:00:46
versus
00:00:48
US competition and then the other is
00:00:51
it's O an open source company or at
00:00:53
least an open source the the R1 model
00:00:56
and so you've kind of got the whole open
00:00:57
source versus closed Source debate and
00:01:00
if you take either one of those things
00:01:02
out it probably wouldn't have been such
00:01:03
a a big story but I think the synthesis
00:01:05
of these things got a lot of people's
00:01:07
attention a huge part of Tik tok's
00:01:09
audience for example is international
00:01:12
some of them like the idea that that the
00:01:14
US may not win the AI race that the US
00:01:17
is kind of getting a come up and here
00:01:19
and I think that fueled some of the
00:01:20
early attention on Tik Tok similarly
00:01:23
there's a lot of people who are rooting
00:01:24
for open source or they have animosity
00:01:27
towards open AI
00:01:30
and so they were kind of rooting for
00:01:32
this idea that oh there's this open
00:01:33
source model that's going to give away
00:01:35
what open AI has done at 120th a cost so
00:01:38
I think all of these things provided
00:01:39
fuel for the story now I think the
00:01:42
question is okay what should we make of
00:01:44
this I mean I think there are things
00:01:46
that are true about the story and then
00:01:47
things that are not true or should be
00:01:50
debunked I think the let's call it true
00:01:53
thing here is that if you had said to
00:01:56
people a few weeks ago that the second
00:02:00
company to release a reasoning model
00:02:04
along the lines of 01 would be a Chinese
00:02:07
company I think people would have been
00:02:09
surprised by that so I think there was a
00:02:11
surprise and just to kind of back up for
00:02:12
people you know there's there's two
00:02:14
major kinds of AI models now there's
00:02:16
kind of the Bas llm model like Chachi
00:02:18
P40 or the Deep seek equivalent was V3
00:02:21
which they launched a month ago and
00:02:23
that's basically like a smart PhD you
00:02:25
ask a question gives you an answer then
00:02:27
there's the new reasoning models which
00:02:30
are based on reinforcement learning sort
00:02:32
of a separate process as opposed to
00:02:35
pre-training and 01 was the first Model
00:02:38
released along those lines and you can
00:02:41
think of a reasoning model as like a
00:02:43
smart PhD who doesn't give you a snap
00:02:45
answer but actually goes off and does
00:02:47
the work you can give it a much more
00:02:48
complicated question and it'll break
00:02:51
that complicated problem into a subset
00:02:53
of of smaller problems and then it'll go
00:02:56
step by step to solve the problem and
00:02:59
that's that's called Chain of Thought
00:03:01
right and so the new generation of
00:03:03
agents that are coming are based on this
00:03:05
type of idea of of chain of thought that
00:03:07
that an AI model can sequentially
00:03:09
perform tasks figure out much more
00:03:11
complicated problems so open AI was the
00:03:14
first to release this type of reasoning
00:03:16
model Google has a similar model they're
00:03:18
working on called Gemini 2.0 flinking
00:03:21
they've released kind of an early
00:03:22
prototype of this called Deep research
00:03:25
1.5 anthropic has something but I don't
00:03:28
think they've released it yet so other
00:03:30
companies have similar models to 01
00:03:34
either in the works or in some sort of
00:03:36
private beta but deep seek was really
00:03:39
the next one after open AI to release
00:03:41
you know the full public version of it
00:03:44
and moreover they open sourced it and so
00:03:46
this created a pretty big splash and I
00:03:49
think it was legitimately surprising to
00:03:51
people that the next big company to put
00:03:55
out a reasoning model like this would be
00:03:57
a a Chinese company and moreover they
00:04:00
would open source it give it away for
00:04:01
free and I think the API access is
00:04:03
something like 12th the the cost so all
00:04:06
of these things really did drive the the
00:04:09
new cycle and I think for good reason
00:04:11
because I think that if you had asked
00:04:12
most people in the industry a few weeks
00:04:15
ago how far behind is China on AI models
00:04:19
they would say six to 12 months and now
00:04:23
I think they might say something more
00:04:24
like 3 to 6 months right because 01 was
00:04:27
released about four months ago and R1 is
00:04:30
comparable to that so I think it's
00:04:32
definitely moved up people's time frames
00:04:35
for how close China is on on AI now
00:04:40
let's take the um we should take the
00:04:42
claim that they only did this for $6
00:04:44
million on this one I'm with Palmer
00:04:48
lucky and Brad gersner and others and I
00:04:50
think this has been pretty much
00:04:51
corroborated by everyone I've talked to
00:04:53
that that that number should be
00:04:56
debunked so first of all it's very hard
00:04:59
to Val
00:05:00
validate a claim about how much money
00:05:02
went into the training of this model
00:05:05
it's not something that we can
00:05:07
empirically discover but even if you
00:05:09
accept it at face value that $6 million
00:05:12
was for the final training run so when
00:05:14
the media is hyping up these stories
00:05:17
saying that this Chinese company did it
00:05:19
for six million and and these dumb
00:05:20
American companies did it for a billion
00:05:23
it's not an Apples to Apples comparison
00:05:25
right I mean if you were to make the
00:05:26
Apples to Apples comparison you would
00:05:27
need to compare the final training run
00:05:31
cost by Deep seek to that of open AI or
00:05:34
anthropic and what the founder of
00:05:37
anthropic said and what I think Brad has
00:05:40
said being an investor in open Ai and
00:05:42
having talked to them is that the final
00:05:45
training run cost was more in the tens
00:05:48
of millions of dollars about nine or 10
00:05:51
months ago and so you know it's not 6
00:05:54
million versus a billion okay it's a
00:05:56
billion dollar number might include all
00:05:58
the hardware they've bought the the
00:05:59
years of putting into it a holistic
00:06:01
number as opposed to the training number
00:06:03
yeah it's not running it's not fair to
00:06:05
compare let's call it a Soup To Nuts
00:06:07
number a fully loaded number by American
00:06:10
AI companies to the final training run
00:06:12
by the Chinese company but real quick
00:06:16
sax you've got you've got an open source
00:06:19
model and they've the the white paper
00:06:21
they put out there is very specific
00:06:24
about what they did to make it and uh
00:06:27
sort of the results they got out of it I
00:06:30
don't think they give the training data
00:06:31
but you could start to stress test what
00:06:34
they've already put out there and see if
00:06:36
you can do it cheap essentially like I
00:06:38
said I think it is hard to validate the
00:06:40
number I think that if let's just assume
00:06:42
that we give them credit for the Six
00:06:43
Million number my point is less that
00:06:45
they couldn't have done it but just that
00:06:48
we need to be comparing likes to likes
00:06:50
yeah so if for example you're going to
00:06:53
look at the fully loaded cost of what it
00:06:55
took deep seek to get to this point then
00:06:58
you would need to look at what has been
00:07:00
the R&D cost to date of all the models
00:07:04
and all the experiments and all the
00:07:06
training runs they've done right and the
00:07:08
compute cluster that they surely have so
00:07:12
Dylan Patel who's leading semiconductor
00:07:14
analyst has estimated that deep seek has
00:07:17
about 50,000 Hoppers and specifically he
00:07:20
said they have about 10,000 h100s they
00:07:23
have 10,000 H 800s and 30,000
00:07:27
h20s now the cost of that s sorry is
00:07:30
they deep seek or it's deep seek plus
00:07:32
the hedge fund deep seek plus the hedge
00:07:33
fund but it's the same founder right and
00:07:36
by the way that doesn't mean they did
00:07:37
anything illegal right because the h100s
00:07:39
were banned under export controls in
00:07:42
2022 then they did the H 800s in 20123
00:07:46
but this founder was very farsighted he
00:07:48
was very ahead of the curve and he was
00:07:50
through his hedge fund he was using AI
00:07:51
to basically do algorithmic trading so
00:07:55
he bought these chips a while ago in any
00:07:57
event you add up the the C cost of a
00:08:01
compute cluster with 50,000 plus Hoppers
00:08:04
and it's going to be over a billion
00:08:05
dollars so this idea that you've got
00:08:07
this Scrappy company that did it for
00:08:09
only six million just not true they have
00:08:11
a substantial compute
00:08:14
cluster that they Ed to to train their
00:08:18
models and frankly that doesn't count
00:08:21
any chips that they might have beyond
00:08:24
the 50,000 you know that they might have
00:08:26
obtained in violation of export
00:08:30
restrictions that obviously they're not
00:08:31
going to admit to and we just don't know
00:08:34
we don't really know the full extent of
00:08:36
of what they have so I just think it's
00:08:39
like worth pointing that out that I
00:08:41
think that part of the story got
00:08:42
overhyped it's hard to know what's fact
00:08:45
and what's fiction everybody who's on
00:08:46
the outside guessing
00:08:49
has their own incentive right like so if
00:08:52
you're a semiconductor analyst that
00:08:54
effectively is massively bullish Nvidia
00:08:58
you want it to be true
00:09:00
that it wasn't possible to train on $6
00:09:03
million obviously if you're the person
00:09:05
that makes an alternative that's that
00:09:06
disruptive you want it to be true that
00:09:09
it was trained on $6 million all of that
00:09:11
I think is all speculation the thing
00:09:14
that struck me
00:09:16
was how different their approach was and
00:09:19
TK just mentioned this but if you dig
00:09:21
into not just the original white paper
00:09:24
of deep seek but they've also published
00:09:26
some subsequent papers that have refined
00:09:28
some of the details
00:09:30
I do think that this is a case and Saks
00:09:32
you can tell me if you disagree but this
00:09:33
is a case where necessity was the mother
00:09:35
of invention so I'll give you two
00:09:38
examples where I just read these things
00:09:39
and I was like man these guys are like
00:09:41
really clever the first is as you said
00:09:45
let's let's put in a pin on whether they
00:09:46
distilled 01 which we can talk about in
00:09:49
a second but at the end of the day these
00:09:52
guys were like well how am I going to do
00:09:53
this reinforcement learning thing they
00:09:55
invented a totally different algorithm
00:09:57
there was the the Orthodoxy right this
00:09:59
thing called po that everybody used and
00:10:02
they were like no we're going to use
00:10:03
something else called I think it's
00:10:05
called grpo or something it uses a lot
00:10:08
less computer memory and it's highly
00:10:11
performant so maybe they were con strain
00:10:14
sacks practically speaking by some
00:10:16
amount of compute that caused them to
00:10:18
find this which you may not have found
00:10:20
if you had just a total surplus of
00:10:22
compute availability and then the second
00:10:24
thing that was crazy is everybody is
00:10:26
used to building models and compiling
00:10:28
through kud
00:10:30
which is nidia's proprietary language
00:10:33
which I've said for a couple times is
00:10:34
their biggest moat but it's also the
00:10:37
biggest threat factor for lockin and
00:10:39
these guys worked totally around Cuda
00:10:41
and they did something called PTX which
00:10:43
goes right to the bare metal and it's
00:10:45
controllable and it's effectively like
00:10:48
writing assembly now the only reason I'm
00:10:50
bringing these up is we meaning the West
00:10:52
with all the money that we've had didn't
00:10:55
come up with these
00:10:57
ideas and I think part of why we didn't
00:10:59
come up is not that we're not smart
00:11:01
enough to do it but we weren't forced to
00:11:03
because the constraints didn't exist and
00:11:05
so I just wonder how we make sure we
00:11:08
learn this principle meaning when the AI
00:11:11
company wakes up and rolls out of bed
00:11:13
and some VC gives them $200
00:11:16
million maybe that's not the right
00:11:18
answer for a series A or a seed and
00:11:20
maybe the right answer is 2 million so
00:11:23
that they do these deep seek like
00:11:26
Innovations constraint makes for great
00:11:28
art what do you think uh freedberg when
00:11:30
you're looking at this well I think it
00:11:33
also enables a new class of investment
00:11:36
opportunity given the low cost and the
00:11:40
speed it really highlights that maybe
00:11:42
the opportunity to create value doesn't
00:11:44
really sit at that level in the value
00:11:45
chain but further Upstream apology made
00:11:48
a comment on Twitter today that was
00:11:49
pretty funny or I think ref this about
00:11:52
the raer he like out the rapper may be
00:11:54
the the moat the the moat which is true
00:11:58
at the end of the day if model
00:12:00
performance continues to improve get
00:12:02
cheaper and it's so competitive that it
00:12:04
commoditized much faster than anyone
00:12:06
even
00:12:07
thought then the value is going to be
00:12:10
created somewhere else in the value
00:12:11
chain maybe it's not the rapper maybe
00:12:14
it's with the user and maybe by the way
00:12:15
here's an important point maybe it's
00:12:17
further in the economy you know when
00:12:19
electricity production took off in the
00:12:21
United States it's not like the
00:12:22
companies are making a lot of money that
00:12:23
are making all the electricity it's the
00:12:25
rest of the economy that Acres a lot of
00:12:26
the value
