Latest / UK Column Radio / The AI Bubble Is Bursting and China Lit the Fuse | Silk and Steel

Transcript

The AI Bubble Is Bursting and China Lit the Fuse | Silk and Steel · UK Column Radio

0:01
Speaker 1: Okay, we are live.
0:03
Speaker 2: Welcome to another episode of the Silken Steel Podcast. I'm
0:07
Speaker 2: your host Carloza. Today we have returning guests, expert on
0:13
Speaker 2: Chinese AI and tech development, TP want joining us. TP
0:19
Speaker 2: wished to remain somewhat anonymous, so we'll just have him
0:24
Speaker 2: his audio. TP welcome back to the show.
0:28
Speaker 3: I'm really happy to be back.
0:31
Speaker 1: We're very happy to have you.
0:34
Speaker 2: And you know, there's recently there's a couple of development
0:38
Speaker 2: that really shook the the AI world in the United States,
0:44
Speaker 2: but also the financial world because in US there's an
0:48
Speaker 2: incipient AI bubble that's building a lot. There's a lot
0:51
Speaker 2: of money on the line. But the two new Chinese
0:57
Speaker 2: models that came out is one is fund Ai they
1:01
Speaker 2: I think they rebranded to z a I. And the
1:04
Speaker 2: other one is from a very unlikely source, a Chinese
1:09
Speaker 2: delivery made one. Can you just tell us in layman's terms,
1:14
Speaker 2: why there are you know, why they're such a big deal,
1:18
Speaker 2: because some some somebody is calling this another deep seek moment.
1:23
Speaker 2: What's so different about them?
1:25
Speaker 3: Yeah, so let me just talk a little bit about
1:28
Speaker 3: the background in like I guess, in the AI following circles,
1:34
Speaker 3: if you look at the Chinese labs. There are the
1:38
Speaker 3: classic you know, hyper scalers that do AI like Ali Baba,
1:44
Speaker 3: Bike Dance and Tensen, and they are uh you know,
1:49
Speaker 3: they have products built around these things, and you know
1:54
Speaker 3: they they tend to do UH stuff that fit different
1:59
Speaker 3: kind of mark kids and like you know bite dance
2:02
Speaker 3: for example with the Doboll app, it's very good for
2:07
Speaker 3: video editing, video creation, images and and everything. And then
2:12
Speaker 3: there are the startups. They're the ones that are very
2:19
Speaker 3: focused on just achieving the best large language model possible,
2:24
Speaker 3: tech spased ones and those are you know deep Sea
2:27
Speaker 3: because people would most people know, and also Jupu or
2:31
Speaker 3: z Ai as they like to refer themselves as now
2:35
Speaker 3: and also Mushot with Kimmy. So for the when when
2:42
Speaker 3: deep Sea came out last year, they were really a
2:45
Speaker 3: big surprise because this was the first time that the
2:47
Speaker 3: Chinese UH players had come out with what we call
2:52
Speaker 3: reasoning model. So before this, in the first stage when
2:56
Speaker 3: Chackie came out, AI was just like chatbots. You ask
3:00
Speaker 3: questions and it spits out whatever the most likely token is,
3:07
Speaker 3: and sometime it can be really dumb, sometime it can
3:10
Speaker 3: hallucinate a lot, and then a reasoning model came out.
3:15
Speaker 3: This actually changed the dynamic where it will first have
3:22
Speaker 3: a lot of what they call thinking tookens, so I
3:25
Speaker 3: think a little bit and figure out what the right
3:27
Speaker 3: answer before actually telling you what the answer is. And
3:32
Speaker 3: so Deep Sequel was the first one in China to
3:35
Speaker 3: really do this, and over the past year, you know,
3:41
Speaker 3: most all the Chinese players have migrated to sinking model,
3:44
Speaker 3: just like you know all the Americans have gone onto
3:48
Speaker 3: the stage. And over time it became I guess in
3:55
Speaker 3: China the three leading ones or Deep Sea, GLM and
4:00
Speaker 3: uh Kimmy the mooshot model, and I think over the
4:06
Speaker 3: past few months it's become quite obvious that the z
4:14
Speaker 3: Ai and Kimmy or so GLM and Kimmy were the
4:18
Speaker 3: one doing the best in terms of like the best
4:21
Speaker 3: you know, reasoning model, as if you ask it to
4:24
Speaker 3: do some kind of research topic or you asked it
4:26
Speaker 3: to do coding, it does a really good job of that.
4:30
Speaker 3: Deep c seems to be very much into the research
4:33
Speaker 3: part of things, like their architecture is the best one.
4:37
Speaker 3: But because it takes so long to do the architecture
4:40
Speaker 3: and also the getting their stuff to work with as
4:46
Speaker 3: end platform for training as such, they kind of actually,
4:51
Speaker 3: I think in terms of model quality itself has fallen behind.
4:54
Speaker 3: And GLM five point two, which came out like a
4:59
Speaker 3: month ago, well, it's got a lot of good reviews,
5:03
Speaker 3: I would say, because it is the first Chinese model
5:08
Speaker 3: to seem like achieve what we call RSI recurse your
5:14
Speaker 3: self improvement kind of quality. So I think amongst American
5:20
Speaker 3: model is the first model to really obviously move to
5:24
Speaker 3: a stage where it can really do really great code
5:27
Speaker 3: I think for a long time and make right decisions.
5:30
Speaker 3: Probably opens four point five or four point six by anthropic,
5:34
Speaker 3: and this is like the first model where people using
5:38
Speaker 3: it like, oh, this feels like have at least that
5:40
Speaker 3: opens four point six with me, you know, depending on
5:43
Speaker 3: the task, some people would say it's like more like
5:45
Speaker 3: a four point five equality. Other people will say this
5:48
Speaker 3: feels like opens four point eight. And so this really
5:53
Speaker 3: impressed a lot of people. If you actually try Zi's
5:58
Speaker 3: latest model to do programming, I mean, you probably have
6:02
Speaker 3: a lot of software development experience, like also, but it
6:06
Speaker 3: works very well. Like I just get it to do
6:10
Speaker 3: all my work. Now I don't program anymore. I just
6:15
Speaker 3: give it commands and it does most of the work
6:18
Speaker 3: and then it's nice. Yeah, it's it's so that's like
6:21
Speaker 3: the big deal here, Like this is the first Chinese model.
6:25
Speaker 3: If you give it a problem, it can just do
6:28
Speaker 3: it on one shot. Even stuff that you don't you
6:33
Speaker 3: don't need to do before, you know, stuff before like
6:36
Speaker 3: you might need to ask it prompted to like Kimmy,
6:39
Speaker 3: I need to prompt it two or three times before
6:42
Speaker 3: it gives me the full solutions. A GM five point
6:46
Speaker 3: two might just do it with one prompt.
6:50
Speaker 2: Wow, I this actually this corresponds to a video I
6:56
Speaker 2: just recently saw on YouTube. One who there's a software
7:01
Speaker 2: engineering in the United States who have a lot of
7:04
Speaker 2: experience working with off sourcing, and he's mentioning one phenomena
7:09
Speaker 2: recently is AI is ab absolutely decimating the off sourcing
7:16
Speaker 2: pipeline to India because he said before, especially a lot
7:22
Speaker 2: of low level programming stuff now can be done with AI.
7:27
Speaker 2: You know, you then you no longer have to deal
7:29
Speaker 2: with a lot of inefficiency that was built in into
7:33
Speaker 2: the offshoing model because you know, like India is several
7:37
Speaker 2: times zone away from North America, for example, and there's
7:41
Speaker 2: a lot of problems dealing with coordination issue before because
7:46
Speaker 2: the labor arbitrage was so great, you know, the cheap labor,
7:52
Speaker 2: it kind of overrides all other concerns. But now when
7:56
Speaker 2: you can just have AI do all this basically I
8:00
Speaker 2: will program me, you no longer need to deal with
8:03
Speaker 2: all that inefficiency. I was building into the offshoing model,
8:07
Speaker 2: so that uh yeah, okay, that that that definitely makes sense.
8:11
Speaker 3: Yeah, I think AI is gonna kill India, like but
8:16
Speaker 3: like a more broader sense. I feel like the entire
8:18
Speaker 3: software engineering industry is getting decimated right now. I I
8:22
Speaker 3: was talking to a friend recently coming in from Silicon Valley,
8:26
Speaker 3: and you know, you just hear like a lot of
8:29
Speaker 3: horror stories about how people are thinking about this. You know,
8:34
Speaker 3: there's a lot more firing is going on, especially with
8:38
Speaker 3: matters like the worst one. But I would say all
8:42
Speaker 3: the firms are just hiring a lot less people, especially
8:45
Speaker 3: new new young people coming out of college. It's gotten
8:50
Speaker 3: so bad to the point where I would personally not
8:52
Speaker 3: want my kids going software engineering in the future because
8:55
Speaker 3: I think there's no hope. But and for me personally,
9:01
Speaker 3: I'm I'm you know, I'm I have my own firm
9:04
Speaker 3: that I'm starting up at the moment, and the main
9:09
Speaker 3: thrust of AI is that I don't need to hire
9:12
Speaker 3: anybody now because I just asked AI to do all
9:15
Speaker 3: the work. Before, if I wanted to do some shit work,
9:19
Speaker 3: I would just have to talk to someone who's like
9:21
Speaker 3: more junior dev. I would explain to them what I
9:24
Speaker 3: need them to do, and a lot of time I
9:26
Speaker 3: felt like explaining things took longer than it would take
9:30
Speaker 3: me to do it personally. Now I can just ask
9:33
Speaker 3: AI to do it and they can do it really quickly.
9:35
Speaker 3: And you know, you, as long as you make the
9:38
Speaker 3: point to it, it can do the work really well,
9:41
Speaker 3: and a lot of time you can do it better
9:43
Speaker 3: then you would personally do it. And the reason for
9:45
Speaker 3: that is because if you think about how AI improves itself,
9:50
Speaker 3: right we use people start using Anthropic a lot when
9:56
Speaker 3: they were using clock cold, and every time they they
10:00
Speaker 3: used Anthropic, they start giving Anthropic more training data because
10:05
Speaker 3: they show Anthropic what the code looks like, and then
10:09
Speaker 3: Anthropical will ask them questions and people will make choices
10:13
Speaker 3: based on those questions, and over time the model have
10:16
Speaker 3: all this training data, they learn to be better at
10:19
Speaker 3: software engineering. And this is why over time Anthropic got
10:25
Speaker 3: good first because people were using it to do a
10:28
Speaker 3: lot of the programming side of things, and that's how
10:31
Speaker 3: they they were the first to you know, get good
10:35
Speaker 3: at like cyber warfares and things like that with missiles
10:38
Speaker 3: because their thing got really good at just looking at
10:44
Speaker 3: code and and being able. Because for for all the
10:49
Speaker 3: things you say about AI, one thing you can do
10:52
Speaker 3: is if it can verify something digitally virtually, then it
10:57
Speaker 3: can keep improving itself. If it's some seeing the real
11:00
Speaker 3: world that can't verify, then it's then just can't do it.
11:04
Speaker 3: But if it's something that it can verify virtually, then
11:10
Speaker 3: it can solve a problem and get smarter at it.
11:13
Speaker 3: So over time, all these software engineering problems that has
11:17
Speaker 3: an answer, if it can verify it, then it can
11:22
Speaker 3: keep improving itself. So so that's why, like you know,
11:29
Speaker 3: Opus got good, and that is also why I think
11:34
Speaker 3: that AI got the good the first amounts of Chinese
11:38
Speaker 3: models because it was being used it kind of pushed
11:42
Speaker 3: itself as more of like a coding kind of AI
11:46
Speaker 3: for for business users. So it got a lot of
11:50
Speaker 3: training data that way, and now it is the first
11:55
Speaker 3: amounts of Chinese models to get really good at this
11:58
Speaker 3: and also like z Ai has the advantage that other
12:01
Speaker 3: firms in China don't have. It is it's professor. It's
12:05
Speaker 3: like its founder is a Team five professor. So what
12:08
Speaker 3: a lot of people don't realize is it gets a
12:10
Speaker 3: lot of free labor or really cheap labor out of
12:13
Speaker 3: Team pive university, and those students can do their interns
12:17
Speaker 3: at Zai for and they can do like stuff like
12:20
Speaker 3: data labeling, verifying model results, and you know a lot
12:25
Speaker 3: of stuff that they have access to that other firms
12:28
Speaker 3: just don't have. So like this has kind of built
12:32
Speaker 3: up z AI over time.
12:36
Speaker 2: I when I hear a lot of conversation about these
12:40
Speaker 2: new AI development I heard the term agen tick ai
12:43
Speaker 2: getting throw around. What does that mean in the layman's term?
12:49
Speaker 3: I In general, like when you ask Ai to do
12:53
Speaker 3: a task, right, it has like agents or like work
12:57
Speaker 3: think about as like virtual workers. So you tell it
13:01
Speaker 3: to create a new front end picture, and then you
13:05
Speaker 3: have some requirements and then it kicks off a bunch
13:08
Speaker 3: of agent to do calls for you. So each agent
13:14
Speaker 3: might okay first for example, UH, you say I want
13:19
Speaker 3: to use this library, so that one agent might go
13:21
Speaker 3: to the UH read the you know the API for
13:24
Speaker 3: this for this library and figure it out and record it.
13:29
Speaker 3: And then another agent might start by designing what the
13:33
Speaker 3: interface looks like and then ask you questions. You know,
13:36
Speaker 3: another one might create test data for to run test later.
13:43
Speaker 3: So like you know, each agent could be in charge
13:46
Speaker 3: of doing different different things for you during this time.
13:50
Speaker 3: If you use z ai to like z code, what
13:56
Speaker 3: you'll see is like a step by step process. It
13:59
Speaker 3: creates us app about how One it plans out how
14:02
Speaker 3: it's going to design the code, and two it does
14:04
Speaker 3: a research to try to figure out what it needs
14:07
Speaker 3: to be done. And then it has these agents to
14:12
Speaker 3: read through your code base and then it will ask
14:15
Speaker 3: your permission for it obviously, so then you can give
14:17
Speaker 3: it permission to look through your code base, and maybe
14:21
Speaker 3: if you give it access to your database, it will
14:22
Speaker 3: look through your databases also, and then and then it
14:27
Speaker 3: might write some test codes just to see what it
14:30
Speaker 3: what it thinks the uh the API that say it's
14:34
Speaker 3: going to use works as it expects, and then it
14:38
Speaker 3: will test. They will create these like test code and
14:41
Speaker 3: try it out and if it works, they will put
14:43
Speaker 3: it in your code. They'll ask, okay, do you want
14:45
Speaker 3: to change your code down and then you have to
14:47
Speaker 3: give it permission to change your code, and then after
14:51
Speaker 3: that it will then write the code, and then it
14:54
Speaker 3: will test the cold and then you know, there's there's
14:58
Speaker 3: like multiple steps where IF verifies tests and then generates
15:02
Speaker 3: the CULDE for you. So that is how it's able
15:04
Speaker 3: to do it so well, whereas some of the other
15:08
Speaker 3: Chinese models, they really haven't got to the stage yet.
15:12
Speaker 3: So I find with deep Seak, if you asked to
15:15
Speaker 3: do something, it's really good at giving you answers really quickly,
15:18
Speaker 3: but it makes a lot of mistakes along the way.
15:20
Speaker 3: And Kimmy is a little better, but it's still not
15:24
Speaker 3: as good at avoiding mistakes as Joe m is.
15:30
Speaker 2: So why is it like what makes z ai such
15:37
Speaker 2: a big threat to say anthropics and open AI.
15:41
Speaker 3: Then I would say that AI with GEM and Kimmy
15:45
Speaker 3: with two point seven combined is a huge threat because
15:51
Speaker 3: the two of them combined con virtually do all the
15:54
Speaker 3: coding tasks you need to do. And as of now,
16:00
Speaker 3: most of the people use models for the purpose software
16:05
Speaker 3: engineering tasks and the I would say ninety nine percent
16:10
Speaker 3: of the software engineering task you can probably do it
16:12
Speaker 3: with at least the stuff I deal with, you can
16:15
Speaker 3: do it with GM and Z and GLM and Kimmy combined.
16:19
Speaker 3: Like I use both of them. I have subscription services
16:22
Speaker 3: to both of them, and the two of them combined
16:25
Speaker 3: can do ninety nine percent of stuff I need to
16:27
Speaker 3: be done. And maybe if you're doing more like high
16:30
Speaker 3: level cs C plus plus programming. Those are some of
16:34
Speaker 3: the things that it struggles with. But in reality, if
16:39
Speaker 3: you have a situation where Silicon Valley firms find that
16:43
Speaker 3: Anthropic and open AI are too expensive and the tokens
16:49
Speaker 3: cost too much money, and they just start using these
16:51
Speaker 3: Chinese models to do all their work, then that's a
16:54
Speaker 3: huge revenue loss for the Anthropic and Open AI and
17:00
Speaker 3: and it's like a you know, it's a big loss
17:04
Speaker 3: for the entire ecosystem because you don't have money coming
17:07
Speaker 3: in to pay pay the hyper scalers who are I
17:10
Speaker 3: call it the AI gravy train. It's like the government,
17:13
Speaker 3: you know, like they just spend money without any questions asked,
17:16
Speaker 3: Like this is how AI is working out. Like Micron
17:19
Speaker 3: has eighty five percent march profit margin on its memoryship sales,
17:22
Speaker 3: and the reason for that is they can discharge whatever
17:25
Speaker 3: and and the AI firms are so concerned about falling
17:28
Speaker 3: behind in their build up that they would they just
17:31
Speaker 3: asked no questions, it's just spend it. So then what
17:36
Speaker 3: happens is that, uh, if you have a problem over
17:41
Speaker 3: time where these firms eventually realize they're not going to
17:44
Speaker 3: get the return on the investment, then they make then
17:46
Speaker 3: the entire bubble just falls apart.
17:51
Speaker 2: So basically it's a thread in terms of both capability
17:56
Speaker 2: and pricing because it sounds like z ai they can
18:01
Speaker 2: do like ninety ninety what anthropics and open ai can do,
18:10
Speaker 2: but at a much much cheaper pricing, and you know
18:14
Speaker 2: it's open source.
18:15
Speaker 3: There's another advantage, right like, if you use an open
18:19
Speaker 3: source model, you can host it on your own service
18:22
Speaker 3: if you're worried about data security, And aside from that,
18:26
Speaker 3: you can train, you can fine tune the model with
18:28
Speaker 3: your own data, and you can change it a little bit,
18:33
Speaker 3: and then you just you're entirely safe that way. And
18:39
Speaker 3: when you find tune with your own internal data, then
18:42
Speaker 3: you can then you can make it suit your own application.
18:45
Speaker 3: So that's something that I think people don't nearly appreciate
18:51
Speaker 3: enough because a lot of big firms they have been
18:55
Speaker 3: just giving away their data to Anthropic for like a
18:59
Speaker 3: couple years now, and Dariel has the audacity of attacking
19:04
Speaker 3: other people for distilling his models. I mean, he's been
19:07
Speaker 3: taking our data for like three years. What is this
19:10
Speaker 3: guy talking about? This guy, yeah, stealing everyone's data for
19:17
Speaker 3: three years, and nobody even knows like what Anthropic is
19:21
Speaker 3: doing with all this stuff, right, Like they're probably just
19:25
Speaker 3: stealing everyone's intellectual property.
19:31
Speaker 1: And he's our.
19:32
Speaker 2: Record of saying open source is a threat because you know,
19:36
Speaker 2: because we don't control it. Basically, because we can't control it,
19:40
Speaker 2: so we don't know it's gonna fall into some hands
19:43
Speaker 2: of some parents who are gonna use the open source
19:46
Speaker 2: AI to do harm.
19:47
Speaker 1: So we have to put a kabach on that which is.
19:51
Speaker 3: Becoming and being I'm telling there, yeah, like that's what
19:55
Speaker 3: threat at you know.
19:56
Speaker 4: Like everyone that goes to Silicon Valley knows these days.
19:59
Speaker 4: You know, if you want to buy a place that's
20:01
Speaker 4: more than five million dollars, it's very hard to find
20:04
Speaker 4: those places. If you want to buy a place that's
20:06
Speaker 4: under four.
20:06
Speaker 3: Million, they're they're widely available because there's they're creating a
20:11
Speaker 3: lot of like really rich people. And then killing the
20:14
Speaker 3: rest of the software engineers.
20:18
Speaker 2: Yeah, I mean, that's that's really what all these uh
20:22
Speaker 2: uh you know, this this AI race is supposedly AI
20:27
Speaker 2: race China is all about. Now, like they're they're using
20:30
Speaker 2: China as the as as an excuse, as a rationale.
20:35
Speaker 2: That why we got to continue to spend you know,
20:39
Speaker 2: gazillion dollars on AI bubble in the United States because
20:43
Speaker 2: we cannot lose to China for some reasons.
20:47
Speaker 3: But what Yeah, But like I would say, like, if
20:50
Speaker 3: you think about it, it's good that the anthropic has
20:54
Speaker 3: become the face of this because there's such easily identical
20:57
Speaker 3: villains in this matter, right that if you had like
21:01
Speaker 3: promoting this stuff, it might actually work. But because Dario
21:06
Speaker 3: is such a villain, how he's preventing this and he
21:09
Speaker 3: spent all day telling the world we're gonna kill everyone
21:11
Speaker 3: with our models.
21:13
Speaker 1: So so.
21:15
Speaker 3: Now he's got getting restricted by the US government because
21:19
Speaker 3: he's been.
21:19
Speaker 2: Telling, oh, we just lost we just lost uh uh
21:30
Speaker 2: TP for a second. Let me let me see, let
21:32
Speaker 2: me let me reach out to him. Man, Wayne just
21:36
Speaker 2: got getting good.
21:38
Speaker 1: What happened?
21:39
Speaker 2: Let me let me reach out to TP on d
21:42
Speaker 2: m okay, TP is back what happened?
21:46
Speaker 1: What happened?
21:47
Speaker 3: Dario?
21:47
Speaker 2: Did Dario get you?
21:48
Speaker 3: TP? Also, don't dare to criticize Elon Musk because you
21:55
Speaker 3: don't know what he's gonna do.
21:58
Speaker 2: Yeah, we are live streaming on ex as well by
22:01
Speaker 2: the place, so.
22:05
Speaker 3: Great, you're great.
22:08
Speaker 2: Evil Okay, So yeah, with now with with with all
22:18
Speaker 2: that said, what like the the common reframe? Now I
22:25
Speaker 2: hear uh funk Eric Schmidt of all people, is that
22:31
Speaker 2: Chinese model open source model still about four to six
22:37
Speaker 2: months behind the frontier model like anthropics and open up AI.
22:42
Speaker 2: Do you think that's an accurate assessment?
22:45
Speaker 3: Yeah, Like I think in general, right, there's just different
22:48
Speaker 3: ways we can measure not like models are always stronger
22:52
Speaker 3: in one area and weakering another area. Like so in
22:57
Speaker 3: some areas it could be to three months behind, the
23:00
Speaker 3: other area might be seven to eight months behind. So
23:03
Speaker 3: I think four to eight six months. When you say
23:05
Speaker 3: GM is four to six months behind, I'm okay with that.
23:09
Speaker 3: Like a lot of people would say this is like
23:11
Speaker 3: the Chinese version of Opus four point six, and it
23:15
Speaker 3: came out four months after four point six, So I
23:19
Speaker 3: would say that, you know, if you want to say
23:21
Speaker 3: it's six or seven months behind, that's fine, that's what
23:25
Speaker 3: you want to say. But if you want to say
23:27
Speaker 3: it's like already for your own purpose, it's as good
23:30
Speaker 3: as Opus four point eight, that's good too. We did
23:35
Speaker 3: get to a point last week where because fable was
23:39
Speaker 3: off market not usable, and GBT five point six was
23:43
Speaker 3: there and restricted also that GM was as almost as
23:48
Speaker 3: good as the the for software engineering task, almost as
23:51
Speaker 3: good as the best American models available. So this was
23:54
Speaker 3: causing a panic amounts like a lot of the leading
23:58
Speaker 3: you know, Silicon Valley people for saying like what are
24:01
Speaker 3: you doing? And you know, people like David Sachs and
24:05
Speaker 3: Mark Andrews and that type of people, because they realized
24:08
Speaker 3: this model is good. A lot of people are going
24:10
Speaker 3: to use it if you if US government does the
24:13
Speaker 3: hard work of nerving all the US latest US models
24:17
Speaker 3: for global users, right.
24:19
Speaker 2: And you know that that's a lot of comments we're
24:23
Speaker 2: saying when and I think it was earlier today andthropics said, oh,
24:28
Speaker 2: United States government just lifted export control on US, and
24:32
Speaker 2: everybody in the comments like did you nerve it? Did
24:35
Speaker 2: you just nerve the model that you released to everybody?
24:38
Speaker 1: That would happened?
24:41
Speaker 3: They've been nursing their models for a while now, right,
24:44
Speaker 3: Like you know, they they run out of them. I
24:48
Speaker 3: guess they run along on compute, so then they they
24:51
Speaker 3: or they just don't want to give you the full
24:54
Speaker 3: compute that they have, so they give like they've been
24:56
Speaker 3: giving user shitty experiences for like a couple of months now,
24:59
Speaker 3: like and this is why it's good to have genuine
25:03
Speaker 3: competition to them. I mean, we got to the point
25:05
Speaker 3: where GLM is clearly a head of Gemini in the
25:08
Speaker 3: UH in the model race in terms of software engineering tasks,
25:12
Speaker 3: and you know, Google is falling behind, and there are
25:20
Speaker 3: people are leaving and going to anthropics, so like it's
25:24
Speaker 3: very much like very possible in the next year you're
25:27
Speaker 3: gonna have a four horse race between Open AI, Anthropic, GLM,
25:34
Speaker 3: and Kimmy as the four best labs, and none of
25:38
Speaker 3: them are hyper scalers, and Metas was already be knocked
25:42
Speaker 3: out of the picture for for a few for a
25:44
Speaker 3: couple of years for a year now, so you know,
25:46
Speaker 3: like it is very interesting housing housings are So we
25:52
Speaker 3: think that over the next couple of months, right, Kimmy
25:56
Speaker 3: Moushat is gonna come out with Kimmy three in the
25:59
Speaker 3: around in July, so next this sometime this month, and
26:04
Speaker 3: GLM is gonna come out with something pretty big in August.
26:09
Speaker 2: So and we do we know for sure that uh
26:14
Speaker 2: z Ai and Moonshot and Kimmy Moonshot, they all use
26:19
Speaker 2: Chinese ships while we chips, okay.
26:21
Speaker 3: So that's kind of the interesting part for that AI.
26:26
Speaker 3: They they use the Chinese ships mostly for inference at
26:31
Speaker 3: the moment, and the main reason for that is they've
26:34
Speaker 3: been waiting on the Atlas nine fifty UH super superpots
26:39
Speaker 3: for a while, so they haven't they haven't really been
26:44
Speaker 3: using assent stuff for h for training, and I'm not
26:48
Speaker 3: sure about Moonshot, and but Deep Seek is pretty much
26:51
Speaker 3: in the same boat that they've been using this the
26:54
Speaker 3: Chinese ships for inference, but for the actual UH training
26:59
Speaker 3: part of it, they did most of it still in
27:02
Speaker 3: video chips and the UH both dec ends that AI
27:08
Speaker 3: made it pretty obvious. I think when they get the
27:12
Speaker 3: Atlas nine fifty superpots that they're gonna do a lot
27:16
Speaker 3: more most of their training on that. And I think
27:20
Speaker 3: it's because Atlas nine fifty is actually a pretty big
27:23
Speaker 3: jump I expect to jump in terms of software support,
27:28
Speaker 3: in terms of just the computation is able to do
27:32
Speaker 3: versus the earlier ascent clusters.
27:37
Speaker 2: What do you say to anthropics and Dario's accusation that
27:42
Speaker 2: the Chinese AI got to where they are because they're
27:45
Speaker 2: running distillation attack against you know, uh, against Anthropics and
27:52
Speaker 2: open AI and and now it's leading Anthropics to kind
27:56
Speaker 2: of filtering for Chinese users from from China they there,
28:00
Speaker 2: which has been confirmed. They're they're now like targeting uh,
28:04
Speaker 2: Chinese users. So I guess they want to do that
28:07
Speaker 2: to prevent distillation from China.
28:11
Speaker 3: I mean, you've been stealing data off Chinese users for
28:15
Speaker 3: two years now like you've been. He's been distilling data
28:19
Speaker 3: off everyone else. Look, look, this distillation is a common practice.
28:24
Speaker 3: Meta admitted that they do distillation. Uh actually must admit it.
28:28
Speaker 3: He was doing in distillation on the core filings with
28:31
Speaker 3: s AI. I mean it's it's a common practice. Right.
28:34
Speaker 3: There's obviously stuff like benchmarking you can do when you
28:37
Speaker 3: use someone else's a p I, you could do some
28:41
Speaker 3: benchmark marking, and you can also do distillation to see
28:44
Speaker 3: what their stuff generates. Right, so use knowledge from from there.
28:49
Speaker 3: So it obviously helps you speed up because you know,
28:53
Speaker 3: Anthropic was the first one to actually get their clock
28:58
Speaker 3: code into everyone's hand, take wrong data, and then create
29:02
Speaker 3: their models. So people tend to use that to use
29:05
Speaker 3: them first. But you know, if it makes steal feel better,
29:09
Speaker 3: Like I think both Kimmy and Zai has gotten to
29:13
Speaker 3: the point they got so much data now that they
29:15
Speaker 3: can do their own like self improvement they don't need
29:18
Speaker 3: to use They are distilled enough actually, so they don't
29:23
Speaker 3: actually need Anthropic to do to do this anymore. I mean,
29:27
Speaker 3: they could still do it for like every lap they
29:31
Speaker 3: do benchmarking versus other labs to see like how well
29:35
Speaker 3: the other ones performed to see you know, they want
29:37
Speaker 3: to evaluate like how good their model is compared to
29:40
Speaker 3: the competitors model. Right, that's pretty legit, legitimate stuff. So
29:48
Speaker 3: you know, I think that I I don't have any
29:51
Speaker 3: sympathy for Anthropic, that they're they're evil, So.
29:56
Speaker 2: I don't think many people have sympathies for Daria and Drop.
30:00
Speaker 2: But I have a question. So you know, like listening
30:05
Speaker 2: to other US quote unquote AI expert talking about this
30:10
Speaker 2: AI race, a lot of them said are still saying, okay, well,
30:15
Speaker 2: as long as we can maintain this four to six
30:17
Speaker 2: month lead on the Chinese AI that once we reach
30:22
Speaker 2: a g I first, you know, then will have equivalent
30:26
Speaker 2: of thousands of PhDs training new PhDs, you know, except
30:33
Speaker 2: they don't need to sleep and they don't need to
30:34
Speaker 2: to take a take a dump, and and then you
30:37
Speaker 2: will just grow exponentially. That the artificial super intelligence will
30:42
Speaker 2: grow exponentially. China would never catch up. But what I
30:45
Speaker 2: hear you saying was that, you know, the four to
30:49
Speaker 2: six month gap is not a big deal. So what
30:52
Speaker 2: if you know China achieved AGI for six months later? No,
30:57
Speaker 2: like like yep, I mean there is equitability. Go ahead,
31:01
Speaker 2: go ahead, Sorry.
31:02
Speaker 3: I still don't understand what the point of a GI is.
31:05
Speaker 3: Is it just so you can do cybertext like that?
31:08
Speaker 3: That to me seems like the only two purposes I
31:10
Speaker 3: find so far for these l ms. I mean, they
31:13
Speaker 3: helped me, They help me do regular staff to make
31:16
Speaker 3: my life a little easier. Right, If I want to
31:18
Speaker 3: do something on my computer, you can ask UH questions
31:21
Speaker 3: and it helps you walk walk you through that stuff.
31:24
Speaker 3: But it seems like application wise, its main purpose for
31:29
Speaker 3: the national security focus actually gets cyber warfare from what
31:32
Speaker 3: I can see. So when when they say that it's
31:37
Speaker 3: four to six months ahead, I would say that you know, uh,
31:43
Speaker 3: that's probably gonna be can maintained. Why I said r
31:47
Speaker 3: SI is important because of self improvement is important is
31:50
Speaker 3: because once you get to a point where your own
31:55
Speaker 3: models can can operate like intelligent sources, that it can
32:03
Speaker 3: just what we call long context, so you give it
32:09
Speaker 3: a lot of information, it can just think for a
32:11
Speaker 3: long time to solve problems. Once you get to that stage,
32:14
Speaker 3: then what it could do is it could look at
32:17
Speaker 3: your current model right and then it can design improvements.
32:21
Speaker 3: It's saying like, oh, this is actually what you're doing
32:24
Speaker 3: right now is okay, but this could make your model
32:26
Speaker 3: even better if you do it this way, and then
32:28
Speaker 3: it can give you ideas of how to design new
32:31
Speaker 3: models and train do different experiments and train new models.
32:36
Speaker 3: So that's kind of like what the idea behind RSI is.
32:39
Speaker 3: And once you get to that stage, then the ideas
32:43
Speaker 3: your product releases will improvements will go a lot faster.
32:48
Speaker 3: And people might have noticed, like Opus have been moving
32:52
Speaker 3: pretty fast since there four point five, whereas like Google
32:56
Speaker 3: has basically stopped so and Die will keep on talking
33:00
Speaker 3: about RSI every time about how he can improve it
33:03
Speaker 3: their own models and that's kind of like the big
33:05
Speaker 3: deal with JLM five point two is that GM five
33:10
Speaker 3: point one was the first GM models where you could
33:16
Speaker 3: just tell it to do something and can sink for
33:18
Speaker 3: eight hours just working on it. So then you saw
33:21
Speaker 3: this huge improvement from five point one to five point
33:24
Speaker 3: two because five point one was so good at this stuff.
33:27
Speaker 3: And similarly Kimmy's two point six, well, like they were
33:31
Speaker 3: like those two models made a big deal when they
33:33
Speaker 3: when they reached that stage because they're like, oh, looks
33:36
Speaker 3: like they can just sink on their own and make improvements.
33:40
Speaker 3: So the idea is once Chinese laps hit this stage
33:44
Speaker 3: and if they have enough data from people like you
33:47
Speaker 3: and me using it, people who using this through open router,
33:50
Speaker 3: once it has enough you know, sources like this, then
33:53
Speaker 3: it can just improve itself. And that's how it's It
33:59
Speaker 3: can keep improving itself and keep you know, four to
34:03
Speaker 3: six months behind or maybe even if even catch up
34:06
Speaker 3: even sooner. The concern for a lot of people weren't
34:10
Speaker 3: that the Chinese labs actually don't have enough compute. It's
34:13
Speaker 3: more that they don't have the data pipeline.
34:16
Speaker 2: Hmmm, why why not? I mean, you you would think Uh,
34:21
Speaker 2: with such huge population, you know, China will have all
34:25
Speaker 2: the data he needs.
34:27
Speaker 3: Well, there's a lot of Chinese people using cloud.
34:30
Speaker 1: Code, right, ah, okay, okay, so people.
34:34
Speaker 3: In China also love using cloud cold so that's kind
34:36
Speaker 3: of actually the problem. The problem is Chinese people have
34:39
Speaker 3: been giving Dario so much free data, all this stuff
34:42
Speaker 3: for all the time.
34:44
Speaker 1: Yeah, yeah, that's true.
34:46
Speaker 2: I mean, like you you mentioned it, like deep Sea
34:50
Speaker 2: only gain Reggnation domestically because deep Seak was known internationally
34:54
Speaker 2: first and then it gained white reggnation.
34:59
Speaker 3: People are hate themselves too much. Like I find people
35:05
Speaker 3: online and saying, oh, you know, our Chinese models are good.
35:08
Speaker 3: So I ask them what what do you use at home?
35:09
Speaker 3: They're like, oh, we use clot. I was like, okay,
35:11
Speaker 3: so why are using clod if the Chinese model is
35:13
Speaker 3: good enough? Man, I don't use plot. I use z code.
35:18
Speaker 1: Uh that's that's that's funny.
35:21
Speaker 3: Yeah, Like give me code and z code on my machine.
35:24
Speaker 3: I don't let Cloud have any of my data, you know,
35:27
Speaker 3: And you have all these Chinese people who use cloud
35:29
Speaker 3: code all day and then and then they and then
35:32
Speaker 3: they wonder why cloud has as much data to train on,
35:35
Speaker 3: you know, it's it's quite amazing stuff.
35:37
Speaker 2: Actually, So yeah, like it's interesting you say that Chinese
35:48
Speaker 2: AI model's biggest problem is not compute, So you don't
35:51
Speaker 2: think the hardware limitation is a problem because right now,
35:55
Speaker 2: you know, United States building AI centers like crazy. Right,
35:58
Speaker 2: there's lots of consing that going to build data centers
36:02
Speaker 2: in the United States, and the whole premise that these
36:05
Speaker 2: people betting on that United States will win by brute power,
36:09
Speaker 2: by by brute force, by building tons of data centers
36:13
Speaker 2: and with unlimited access to Nvidia's latest AI chips, that
36:19
Speaker 2: US will win based on boote strengths, you know, broot
36:23
Speaker 2: compute power. But you seem to think Chinese companies do
36:30
Speaker 2: not lack compute.
36:34
Speaker 3: I think they were a little short on compute, to
36:36
Speaker 3: be honest, especially with that AI. I have some contacts
36:41
Speaker 3: that AI. They told me that they are for training.
36:44
Speaker 3: They were using a cluster of ten thousand, h one hundred.
36:48
Speaker 3: But you know now that they're going to get assent
36:52
Speaker 3: and you know, as you mentioned earlier, may client just
36:55
Speaker 3: showcase that you can you can train like a one
36:59
Speaker 3: point six trilling parameter model just you know, a last
37:05
Speaker 3: generation a cent cluster. Then going forward, these labs will
37:12
Speaker 3: have a lot more compute. I mean deep c has
37:16
Speaker 3: been doing these job postings recently about their new having
37:20
Speaker 3: a new data center Intermongolia. That's they're they're saying they
37:25
Speaker 3: want upwards of one gigle of compute, right, So that's
37:29
Speaker 3: like a huge increase over what they have now. And
37:34
Speaker 3: part of the work that they've been doing all this
37:36
Speaker 3: time is to get their software stack to be able
37:41
Speaker 3: to run these models on a scent, to be able
37:46
Speaker 3: to do continuous training, because your model always need continuous
37:50
Speaker 3: training to absorb more data and keep improving, right, Like,
37:54
Speaker 3: you can't stop training if your AI lab. So yeah,
37:59
Speaker 3: so once they once they have access to the Atlas
38:03
Speaker 3: clusters and maybe that's gonna be a Q three kind
38:05
Speaker 3: of thing, then they will be able to do a
38:08
Speaker 3: lot more experimentation. Uh. The argument about the compute was
38:14
Speaker 3: that if you have more compute, you can do a
38:16
Speaker 3: lot more experimentation and you can do like different ideas
38:23
Speaker 3: that your models say you should try this this design
38:26
Speaker 3: to see if you can get better results, and then
38:30
Speaker 3: then you have to do scripts if your your if
38:33
Speaker 3: your model is good enough, right, then you can you
38:35
Speaker 3: can have models to determine if your experiment is a
38:40
Speaker 3: is a good one or not, so that you know
38:42
Speaker 3: you have your validation steps along the way. Uh. But
38:47
Speaker 3: if we get to the point where the Chinese labs
38:51
Speaker 3: have a lot more compute, then they can run a
38:53
Speaker 3: lot more of these experiments. And you could make the
38:58
Speaker 3: argument that because of US labs had all this experimentation,
39:04
Speaker 3: that's why the Chinese labs had to do some distilling
39:07
Speaker 3: of them before. Or you could also make the argument that,
39:13
Speaker 3: I guess up until now, we've had several labs in China,
39:18
Speaker 3: each running their own you know, experiment and then they
39:23
Speaker 3: put their the open source their research results. So in
39:26
Speaker 3: a way that they're all running research and they're all
39:28
Speaker 3: helping each other out. So but with more compute they
39:34
Speaker 3: should be able to do do training that they weren't
39:40
Speaker 3: able to do before. So that's kind of the hope.
39:43
Speaker 3: And I think if you ask any lab do you
39:49
Speaker 3: want more compute to do training any everyone would say yes, right.
39:55
Speaker 3: And that's where all the the the I p O,
39:58
Speaker 3: all the fundraising and you know, the domestic cluster comes in,
40:03
Speaker 3: because then they have more resource to do all this
40:05
Speaker 3: training and to uh continue to invest in their R
40:09
Speaker 3: and D and such.
40:12
Speaker 1: Well.
40:14
Speaker 2: But fundamentally on the hardware level, you know, Chinese sides
40:20
Speaker 2: still have a handicap, right that you know they're they're
40:23
Speaker 2: they're competing against you know, they're using older uh d
40:30
Speaker 2: u V machines using multi patterning and trying to uh
40:35
Speaker 2: uh get to like seven animeters while you know TSML,
40:41
Speaker 2: I mean uh TSMC that can get access to the
40:44
Speaker 2: latest e u V machine and make no three and
40:47
Speaker 2: animeter uh chips and and that I mean, like does
40:54
Speaker 2: doesn't that make a difference?
40:57
Speaker 3: It could? I think when it comes to the when
41:01
Speaker 3: it comes to a system, you could have different things.
41:04
Speaker 3: So the biggest difference between a three nanometer and seven
41:09
Speaker 3: nanometer when it comes to compute is you use up
41:12
Speaker 3: more energy for US seven nanometer chips and you know,
41:17
Speaker 3: might create more heating that you need to resolve. It
41:21
Speaker 3: might use more electricity you need to resolve. Right. So
41:24
Speaker 3: I think some of the things that while we had
41:28
Speaker 3: to do when they're designing their system around domestic supply
41:32
Speaker 3: chain is it had to make the actual chips smaller
41:35
Speaker 3: to one get higher yield on them because you know,
41:39
Speaker 3: the domestic processes have lower yields, so you have to
41:42
Speaker 3: adjust around that. You have to make the chips a
41:44
Speaker 3: little smaller so you know the amount of heat generated
41:48
Speaker 3: off them, it's easier to get them off the board
41:51
Speaker 3: and don't damage things. And like, those are the considerations
41:56
Speaker 3: you have to make. And also there are other part
41:59
Speaker 3: of the system. For example, Huawei is going to go
42:03
Speaker 3: to an all optical cluster with Atlas ninety sixty, and
42:10
Speaker 3: part of the rationale for that is so that it
42:15
Speaker 3: can make they can take advantage of its own personal
42:19
Speaker 3: advantage in telecom and in networking to achieve similar level
42:25
Speaker 3: of performances to like an Nvidia cluster, because it can
42:32
Speaker 3: link more chips together and optical networks are probably faster
42:36
Speaker 3: than copper ones. But you know you need higher expertise
42:39
Speaker 3: in networking to get to there. So when it comes
42:42
Speaker 3: to the Chinese clusters, once you get to a point
42:46
Speaker 3: where it has a full domestic supply chain and it
42:49
Speaker 3: can produce enough chips, then even if it uses a
42:55
Speaker 3: little more electricity and generates a little more heat, it's
42:59
Speaker 3: not a problem because inner Mongolia, the weather is very
43:04
Speaker 3: cold most of the year, so it has a lot
43:07
Speaker 3: of natural air cooling, and it's also next to a
43:11
Speaker 3: lot of wind turbines, which are really cheap sources of energy. Right,
43:15
Speaker 3: So the kind of problems that US firms and Chinese
43:19
Speaker 3: firms phase are different. Like even if you have to
43:22
Speaker 3: use twice amount of power to run your data centers,
43:28
Speaker 3: it's not a big deal for a Chinese company because
43:31
Speaker 3: you just get it from the grid, right, Like I
43:34
Speaker 3: was looking some of the figures recently, and the Chinese
43:38
Speaker 3: Energy Agency is saying that essentially about only about twenty
43:45
Speaker 3: percent of the expected power draw increase over the next
43:50
Speaker 3: five years comes from AI, you know. So it doesn't
43:54
Speaker 3: even like it can see all these projects in the pipeline, right,
43:58
Speaker 3: It can see all these the cloud uh service providers
44:01
Speaker 3: in China doing these projects. They can see all these
44:05
Speaker 3: different data center projects along the line, and it has
44:10
Speaker 3: no issue with supplying the power for that. So whereas
44:14
Speaker 3: the America you have GPUs just laying out in the
44:18
Speaker 3: warehouses right now, waiting for other parts of the the
44:22
Speaker 3: full supply chain to catch up, like Meta has like
44:26
Speaker 3: warehouses off GPUs just lying around because you know, the
44:30
Speaker 3: rest of the supply chain is not ready. So I
44:37
Speaker 3: think once once these Chinese labs have that at list,
44:41
Speaker 3: then they can then they can and then they can
44:45
Speaker 3: have a lot more compute and you know, the amount
44:51
Speaker 3: of compute deficit they have to American labs is likely
44:55
Speaker 3: to shrink over time, not increase, because because it's taken
45:00
Speaker 3: a while for the Chinese lab to get to a
45:03
Speaker 3: point where they can have these Atlas clusters that that's
45:07
Speaker 3: that's usable for training.
45:13
Speaker 2: So you think it's likely the that that that four
45:16
Speaker 2: to six month gap that Chinese companies will be able
45:20
Speaker 2: to close, uh, you know, to bring break, to narrow
45:24
Speaker 2: that gap in the in the in the coming immediate future.
45:30
Speaker 3: I don't want to, like just being a conservative, a
45:33
Speaker 3: little more conservative. I think it's more likely that you know,
45:35
Speaker 3: the four to six months gaps stay where they are
45:39
Speaker 3: for the next uh, for the foreseeable future. You can
45:43
Speaker 3: be optimistic and say that they don't narrow it, but
45:46
Speaker 3: I think we need to wait and see what the
45:49
Speaker 3: next couple of months looks looks like, to see to
45:53
Speaker 3: see if some of the Chinese models, you know, catch
45:58
Speaker 3: up to the American models in the you know, in
46:03
Speaker 3: the high end of things, because it is possible that
46:08
Speaker 3: they narrow it further, but we're gonna have to wait
46:11
Speaker 3: and see, Like GLM five point two was a surprise
46:13
Speaker 3: for a lot of people Selicon Valley because they're like, oh,
46:16
Speaker 3: this model is actually good. I can switch from Opus
46:19
Speaker 3: to GLM and not see any major differences, right. I
46:24
Speaker 3: think it would be a real shock to them if
46:26
Speaker 3: like something close to fable comes out in August from GLM.
46:34
Speaker 2: Yeah, that's that's fair. Now on one question, do you
46:39
Speaker 2: think the anthropics and open ai I p O is
46:43
Speaker 2: because they're trying to catch out at the top.
46:47
Speaker 3: Yeah, I think so. I mean like, and also I
46:50
Speaker 3: think part of it is maybe the Middle Eastern money
46:53
Speaker 3: stops flowing in. So, you know, when the US troops
46:57
Speaker 3: start pulling out the Middle East, the saudast aren't aren't
47:00
Speaker 3: funding the stuff. When the UAE is meeting with IRGC's,
47:04
Speaker 3: maybe they're not gonna put us much money into US tech,
47:07
Speaker 3: you know. And I think also Soft Bank, I don't
47:11
Speaker 3: know if you've seen the Soft Bank presentation about their
47:14
Speaker 3: investment Opening Eyes looking increasingly ludicrous. Okay, if Stoft Bank
47:19
Speaker 3: runs out of money, that's another place that that you know,
47:23
Speaker 3: Opening Eye is gonna lose cash from so cash input from.
47:27
Speaker 3: So it seems to me that if once you run
47:34
Speaker 3: out of money on the private market. That's when you
47:36
Speaker 3: want to I p O, right, because once you start,
47:41
Speaker 3: once you I p O, there's a lot of reporting requirement,
47:44
Speaker 3: your finance has really come out, and there's just a
47:47
Speaker 3: lot more things. You have the answer to the investor
47:51
Speaker 3: class that you don't have to do as much when
47:54
Speaker 3: you can just sell the hype to the you know,
47:57
Speaker 3: soft bank and the and and the amortis so. So
48:03
Speaker 3: there is that and they want to they want to.
48:10
Speaker 3: I see that some of them are talking about waiting
48:12
Speaker 3: until twenty twenty seven to IPO. Now maybe they think
48:15
Speaker 3: that the the market is not there. I don't really know, right,
48:20
Speaker 3: Like they need the money clearly because they keep spending
48:23
Speaker 3: all this money on the data center build out. So
48:28
Speaker 3: so we'll see what happens. But I think they definitely
48:32
Speaker 3: want to push. Well the market is still hot. Once
48:38
Speaker 3: the bubble happens, they just can't get as much money
48:40
Speaker 3: and then their model goes away.
48:43
Speaker 2: Yeah, I don't I don't know if the I don't
48:46
Speaker 2: know if the market will stay at all time high
48:49
Speaker 2: in twenty twenty seven. That's that's that's wildly optimistic. But
48:55
Speaker 2: who knows. Maybe the irrational exuberntce can can last longer
48:59
Speaker 2: than any unexpected.
49:02
Speaker 3: Well, but like that's kind of how we do things
49:05
Speaker 3: here in America by being keep being exuberant about everything, right,
49:10
Speaker 3: you know, reality check is not a thing. So so
49:16
Speaker 3: so I think it's so. I think it's interesting. But
49:19
Speaker 3: in a way, I think the Chinese labs also benefit
49:21
Speaker 3: from it because when z Ai and Moushat is getting
49:26
Speaker 3: a bigger fund they're doing better with fundraising as a
49:30
Speaker 3: result of this. So they're trying to take advantage of
49:32
Speaker 3: it by getting more money from the I p O
49:36
Speaker 3: and the private funding around. Like deep Sea got a
49:39
Speaker 3: you know a lot of money from its funding around,
49:42
Speaker 3: and that's how it's able to go into this expansion
49:45
Speaker 3: mode recently. So you know, everyone needs to cashing while
49:50
Speaker 3: this AI bubble is happening. The rest of us get
49:55
Speaker 3: gets screwed because we get higher memoryship prices and uh
49:58
Speaker 3: and we inflation. But you know, the the the memoryship
50:04
Speaker 3: companies make a lot of money from this.
50:07
Speaker 2: Well, I was hoping China would save us on the
50:10
Speaker 2: memory market, you know, like I was hoping that they
50:12
Speaker 2: will start cranking out mass production of rams to bring
50:18
Speaker 2: down the prices. You know, like it looks like it
50:21
Speaker 2: hasn't quite happened yet.
50:23
Speaker 3: Oh. I mean, there's just so much demand right Like
50:27
Speaker 3: the AI gravy training is running so hyper the moment
50:31
Speaker 3: that any additional capacity coming online, it's just getting absorbed
50:36
Speaker 3: until twenty twenty eight. Unless we have a huge crashing AI,
50:39
Speaker 3: then the memoryship price is likely it won't come down
50:42
Speaker 3: til like twenty twenty eight, later, twenty twenty eight and
50:45
Speaker 3: twenty twenty nine, that's what they say. Yeah, and that's
50:49
Speaker 3: with the man suppressed already because the Chinese OEMs are
50:54
Speaker 3: selling less phones these days in a way that's bad,
50:57
Speaker 3: but also a way, you know, like these Chinese ems
51:00
Speaker 3: were making very little margin, very low margins on the phone,
51:03
Speaker 3: so that they were selling and uh, you know, c
51:07
Speaker 3: x MT and YTC are going IPO soon and you'll
51:11
Speaker 3: see how much profit they're making. They're going to be
51:13
Speaker 3: like the two largest tech companies in China, just because
51:16
Speaker 3: the memory ship prices are so high. Like the when
51:23
Speaker 3: c x MT I wrote about the c x M
51:25
Speaker 3: t went I p O, it should be a trillion
51:27
Speaker 3: dollar company and same with YNTC should be very close
51:31
Speaker 3: to that just because of how outrageous and memoryship prices
51:35
Speaker 3: are at the.
51:35
Speaker 2: Moment that that's just it just doesn't seem sustainable, like
51:43
Speaker 2: the whole AI bubble.
51:46
Speaker 1: But yeah, you know.
51:47
Speaker 3: Think about it, if you think about snick, right, was
51:50
Speaker 3: making has been making like two billion a little bit
51:52
Speaker 3: more than two billion dollars per quarter for a long time. Right,
51:56
Speaker 3: And my projection is that by later this year, YMTC
52:01
Speaker 3: and CXMT are gonna be make twenty billion a quarter.
52:05
Speaker 2: Wow.
52:06
Speaker 1: Wow, they're printing money.
52:09
Speaker 3: They are printing money. Their first quarter they were at
52:11
Speaker 3: already at seven like seven billion dollars and six and
52:15
Speaker 3: seven billion dollars. The second quarter the price went up
52:18
Speaker 3: another like forty forty fifty so and third quarter is
52:22
Speaker 3: going up another forty fifty. So it's crazy these days.
52:27
Speaker 2: Yeah, I mean, like are are they but are they
52:31
Speaker 2: finally making inroads into you know, uh elbow into the
52:37
Speaker 2: South Korean territory in terms of a ramp dram production.
52:43
Speaker 3: Yeah, I mean, like once you have DDR five and
52:48
Speaker 3: maybe the Chinese ones, you know they have a large
52:51
Speaker 3: slightly larger feature size, so then you just end up
52:55
Speaker 3: having like less. Look, you can you can can't produce
53:00
Speaker 3: many chips per wafer. That's the major difference. But for
53:06
Speaker 3: people like you and me out there, you know, twelve
53:10
Speaker 3: gigabye of RAM is twelve gigabaye of RAM. It doesn't
53:13
Speaker 3: really matter if it's like a fourteen nanometer or a
53:15
Speaker 3: sixteen nanometer.
53:18
Speaker 1: I played.
53:19
Speaker 2: I played fifteen year old games, so I'm good good.
53:24
Speaker 3: Yeah, So most people don't care about that stuff, right,
53:27
Speaker 3: and that's why, Like Apple was talking about buying Chinese
53:32
Speaker 3: memory chips. But the problem with that is, you know,
53:35
Speaker 3: like CXMP doesn't have enough production even for Chinese customers,
53:39
Speaker 3: So like I don't know, like where where suddenly thinks
53:43
Speaker 3: that all these Chinese chips are coming from, So we'll see,
53:47
Speaker 3: we'll see. I think cexm T and YMTC are going
53:51
Speaker 3: to make huge amount of money. Then people keep on
53:54
Speaker 3: talking online about like these Chinese crash. I was like,
53:56
Speaker 3: have you have you seen how much money they're making
53:59
Speaker 3: AI right now? All these Chinese producers, like they are
54:02
Speaker 3: just raking if you look at the the Chinese industry
54:06
Speaker 3: profit margins right like it's being carried right now by
54:09
Speaker 3: the AI trade, Like the AI related trade right now
54:13
Speaker 3: is making so much money in China.
54:17
Speaker 1: Wow, wait, how are they making money?
54:20
Speaker 2: Like I don't think open ai and UH andthropics really
54:24
Speaker 2: makes that much money compared to the money investment that's
54:27
Speaker 2: pouring in.
54:28
Speaker 3: Well, if you if you look at how much money
54:30
Speaker 3: and the free cash flow these uh you know cloud
54:34
Speaker 3: service provider are spending on this stuff right like that
54:37
Speaker 3: they're spending hundreds of billions of dollars. They're going deep
54:40
Speaker 3: into that to buy and video service. Right Like, just
54:44
Speaker 3: think about all the battery unit you need, the memory
54:47
Speaker 3: tip you need, the PCBs Prince circuit boards and every
54:52
Speaker 3: part of the print board, the fiber optic cables, the
54:57
Speaker 3: optical transceivers, h the supercapacitors and the battery and the
55:04
Speaker 3: memory and the power chips. There's just so much stuff
55:07
Speaker 3: involving there, and all that is concentrated in Asia, right
55:11
Speaker 3: and then you have this like major major what I
55:15
Speaker 3: call major AI supply chain war right now going between
55:18
Speaker 3: China and Japan, where China is starving Japan's natural resources
55:22
Speaker 3: so that you can destroy Japanese from this rout out
55:26
Speaker 3: the Japanese producers from this this this industry, I mean,
55:30
Speaker 3: the Takaki is going to go down in history as
55:32
Speaker 3: a person that destroyed Japan. Like just you know, before
55:36
Speaker 3: China didn't have an excuse to you know, stopping the
55:40
Speaker 3: supplies of the upstream materials to Japan. Now it does, so.
55:46
Speaker 2: Yeah, now China can claim why am I selling you
55:50
Speaker 2: critical minerals that you're just gonna use to re arm yourself.
55:54
Speaker 2: You know, we're not crazy to give you the material
55:57
Speaker 2: to make weapons against us, but you know it's just
56:01
Speaker 2: a bonus that they kill saw the Japanese.
56:04
Speaker 3: Yeah, they use the excuse of like, oh, some of
56:07
Speaker 3: the constant powders we sent to the Koreans got ready
56:10
Speaker 3: direct to the Japanese industries, so we're not gonna e
56:12
Speaker 3: support that to Koreans either. So then they have this
56:16
Speaker 3: like they're gonna start the Korean the Japanese economy of
56:20
Speaker 3: constant powders, which are critical in production of these constant
56:26
Speaker 3: based chemicals used in semiconductor production. So it's very interesting
56:32
Speaker 3: what's going on in that that entire supply chain right now. Yeah,
56:36
Speaker 3: you know the problem is that Chinese producers they're still
56:41
Speaker 3: if you look at how much money the memory producers
56:44
Speaker 3: in Korea are making off their chips. The only answer
56:47
Speaker 3: to that is the Chinese supply chain is not asking
56:50
Speaker 3: for enough money from these Koreans Korean memory trips. They
56:53
Speaker 3: should be asking one hundred, a thousand times the price
56:56
Speaker 3: they are right now, because that's how that's how much
56:59
Speaker 3: of the Koreans are making off these days.
57:03
Speaker 2: I'm just cracking up because I saw the comment that
57:05
Speaker 2: says Candace open Garlands on twenty twenty eight, sucky okay,