WEBVTT

00:00:00.200 --> 00:00:03.220
Hello listeners, welcome back to Ascent. So today

00:00:03.220 --> 00:00:04.780
we'll be talking about another really interesting

00:00:04.780 --> 00:00:07.540
company in Asia. What if I told you one of the

00:00:07.540 --> 00:00:09.380
most exciting new players in the competitive

00:00:09.380 --> 00:00:11.839
field of large -language models wasn't founded

00:00:11.839 --> 00:00:14.820
by a Stanford or MIT AI PhD straight out of the

00:00:14.820 --> 00:00:18.719
top lat, but by a quant trader? Was this quant

00:00:18.719 --> 00:00:21.339
trader, well, I'll tell you the name, Liang Wenfeng,

00:00:21.440 --> 00:00:24.539
the man who solved the Chinese market? To what

00:00:24.539 --> 00:00:26.820
extent was his quant trading business developed

00:00:26.820 --> 00:00:30.089
in China? Why did Liang Wenfeng pivot it from

00:00:30.089 --> 00:00:34.170
quant trading to LLM development? And the most

00:00:34.170 --> 00:00:36.350
mysterious of all, how was Liang Wenfeng, the

00:00:36.350 --> 00:00:38.390
founder of DeepSeek, able to compete among all

00:00:38.390 --> 00:00:40.789
the giants in and outside of China, while all

00:00:40.789 --> 00:00:43.130
the other LLM giants are much better equipped

00:00:43.130 --> 00:00:46.570
than him? In this episode of Ascent, we're diving

00:00:46.570 --> 00:00:50.289
into DeepSeek. It is not just... any one LLM

00:00:50.289 --> 00:00:52.530
story, but we'll be going through the founder's

00:00:52.530 --> 00:00:55.350
journey, how he embarked on the first half of

00:00:55.350 --> 00:00:57.750
the career as a quant trader and created a quant

00:00:57.750 --> 00:01:00.390
trading firm, and how did that firm later transition

00:01:00.390 --> 00:01:03.090
into an LLM company, which led to the product

00:01:03.090 --> 00:01:06.209
that we all know today, DeepSeek. Join us as

00:01:06.209 --> 00:01:22.109
we unpack this fascinating transition. Linda,

00:01:22.129 --> 00:01:24.090
I have a question for you. So if you were to

00:01:24.090 --> 00:01:27.189
have 80K RMB, that's roughly a little over 10K

00:01:27.189 --> 00:01:29.810
US dollar at the age of 22, when you just graduated

00:01:29.810 --> 00:01:32.150
from the college, what will you be doing? What

00:01:32.150 --> 00:01:34.609
will you be using it for? Well, first of all,

00:01:34.629 --> 00:01:37.609
I would love to have 80K when I graduate from

00:01:37.609 --> 00:01:41.069
college. I didn't immediately have this, but

00:01:41.069 --> 00:01:44.629
I remember exactly what I've done with... kind

00:01:44.629 --> 00:01:48.989
of my first major bonus paycheck, which is to

00:01:48.989 --> 00:01:52.049
spend it all on traveling in a, I think, one

00:01:52.049 --> 00:01:56.310
or two week holiday to Peru. So definitely spend

00:01:56.310 --> 00:02:00.629
it. What about you, Pan? Invested on the personal

00:02:00.629 --> 00:02:07.430
discovery. Well, I think like yourself, the first

00:02:07.430 --> 00:02:09.550
amount of money that I... get right out of the

00:02:09.550 --> 00:02:12.469
college. First, wasn't a lot. And second, I also

00:02:12.469 --> 00:02:14.689
did a lot of fair share of exploration of the

00:02:14.689 --> 00:02:18.389
world. So, but I think in hindsight, I should

00:02:18.389 --> 00:02:21.810
have invested in something, which I guess brings

00:02:21.810 --> 00:02:25.650
us to the company, the founder that we're talking

00:02:25.650 --> 00:02:28.150
about today, Liang Wenfeng. Well, not only is

00:02:28.150 --> 00:02:31.210
he a really good investor himself, he also made

00:02:31.210 --> 00:02:33.889
crazy amount of money. In between the age of

00:02:33.889 --> 00:02:36.789
22 to 30. And I guess I'll give you a little

00:02:36.789 --> 00:02:39.669
spoiler alert. Quick question for you. How many

00:02:39.669 --> 00:02:42.409
times did your income increase from the age of

00:02:42.409 --> 00:02:45.050
22 to 30? It's kind of a painful question. All

00:02:45.050 --> 00:02:48.370
right. So listeners, please think about your

00:02:48.370 --> 00:02:54.689
own income multiplication. Let's see. I think.

00:02:56.360 --> 00:02:59.879
I'm not, not that much over the past 10 years.

00:02:59.919 --> 00:03:03.819
I think probably a little bit, I would say a

00:03:03.819 --> 00:03:05.979
little bit more than inflation, right? But not

00:03:05.979 --> 00:03:08.659
that much. Yeah, definitely. It is a factor of

00:03:08.659 --> 00:03:12.000
a few times, but it wasn't a factor of like tens

00:03:12.000 --> 00:03:14.740
of times or a hundred times. For sure. Right.

00:03:14.800 --> 00:03:18.319
And I think. To your earlier point, later on,

00:03:18.319 --> 00:03:21.280
after spending my expenses on earth traveling,

00:03:21.460 --> 00:03:24.759
later on, we eventually learn about investment.

00:03:25.099 --> 00:03:29.000
So I would actually say my investment return

00:03:29.000 --> 00:03:32.939
at some point, cumulatively, probably was more

00:03:32.939 --> 00:03:40.080
than my day job income. So that kind of calls

00:03:40.080 --> 00:03:42.120
into the importance of investing, especially

00:03:42.120 --> 00:03:43.900
in the last 10 years when the market is just

00:03:43.900 --> 00:03:47.539
doing so well. Right. Yeah. Which brings us to

00:03:47.539 --> 00:03:50.780
the answer to your question earlier. How much

00:03:50.780 --> 00:03:53.800
did Liang Wenfeng make in the last 10 years?

00:03:54.139 --> 00:03:56.460
I guess apart from all the news coverage about

00:03:56.460 --> 00:04:00.060
DeepSeek and how fast it expanded to different

00:04:00.060 --> 00:04:02.039
parts of the world, Liang Wenfeng's personal

00:04:02.039 --> 00:04:06.620
income from the age of 22 to 30 was 100x. We

00:04:06.620 --> 00:04:09.659
will cover that later on in the stories, but

00:04:09.659 --> 00:04:12.139
just think about that. And he did that purely

00:04:12.139 --> 00:04:15.300
through his capability in mathematics and computer

00:04:15.300 --> 00:04:18.459
science. Very interesting. Yeah. So how did he

00:04:18.459 --> 00:04:21.699
start? So let's dive in. Chapter one, Liang Wenfeng's

00:04:21.699 --> 00:04:25.339
early ages. Small village mass genius. So he

00:04:25.339 --> 00:04:28.399
was born in 1985 and he was born in a very small

00:04:28.399 --> 00:04:33.180
and peaceful and chill village by the ocean in

00:04:33.180 --> 00:04:35.579
Guangdong province. And if you don't know where

00:04:35.579 --> 00:04:38.459
Guangdong province is at, it is the province

00:04:38.459 --> 00:04:40.699
that is basically attached to Hong Kong. It is

00:04:40.699 --> 00:04:44.420
the northern part of Hong Kong. Both of Liang

00:04:44.420 --> 00:04:46.600
Wenfeng's parents are elementary school teachers.

00:04:47.579 --> 00:04:50.199
The city, or rather should I say the town that

00:04:50.199 --> 00:04:52.720
he was in, had less than 1 million people, which

00:04:52.720 --> 00:04:55.620
is considered very small in China. The city is

00:04:55.620 --> 00:04:59.160
called Wuchuan. It is a hub for manufacturing

00:04:59.160 --> 00:05:02.860
plastic slippers. which has nothing to do with

00:05:02.860 --> 00:05:05.939
investing or, you know, like not even tourism.

00:05:06.019 --> 00:05:07.839
I don't think anyone knows about that place.

00:05:08.040 --> 00:05:10.519
So very, very modest beginnings. Very modest

00:05:10.519 --> 00:05:13.500
beginnings. He is very talented. Léon Pond is

00:05:13.500 --> 00:05:15.819
very talented in his elementary school, exhibiting

00:05:15.819 --> 00:05:21.339
mathematics, talents. When he did his college

00:05:21.339 --> 00:05:24.860
entrance examination, it was... It was no surprise

00:05:24.860 --> 00:05:28.339
that he got number one in the town of Wuchuan.

00:05:28.500 --> 00:05:31.139
And also with that, he made it into Zhejiang

00:05:31.139 --> 00:05:33.720
University and he started studying electronic

00:05:33.720 --> 00:05:36.660
information engineering, which sounds very much

00:05:36.660 --> 00:05:41.199
like double E studies. Yeah. And for those who,

00:05:41.300 --> 00:05:44.220
I guess, are less familiar with Chinese universities,

00:05:44.439 --> 00:05:47.740
Zhejiang University is, I think no one would

00:05:47.740 --> 00:05:51.139
disagree if I say among the top five. Definitely

00:05:51.139 --> 00:05:53.920
tier one. Definitely tier one. Yeah. It's extremely

00:05:53.920 --> 00:05:56.959
hard to get in. I don't have the exact numbers,

00:05:57.040 --> 00:05:59.740
but I would say like one in a thousand or more.

00:05:59.959 --> 00:06:04.009
Yeah. Yeah. In fact, my Google. His referral

00:06:04.009 --> 00:06:06.850
was a graduate from Zhejiang University, and

00:06:06.850 --> 00:06:09.329
she is a computer scientist as well. Do they

00:06:09.329 --> 00:06:12.250
know each other? Different times, different times.

00:06:12.410 --> 00:06:16.149
But, you know, from a tier one university is

00:06:16.149 --> 00:06:19.790
a very good start for the journey that he got

00:06:19.790 --> 00:06:21.649
himself into, like in terms of quant trading,

00:06:21.810 --> 00:06:24.050
because that's when the time that not only did

00:06:24.050 --> 00:06:27.689
he study, he also made his later partners in

00:06:27.689 --> 00:06:31.550
business for quant trading. So throughout his

00:06:31.550 --> 00:06:35.470
undergrad. Besides studying, he did also run

00:06:35.470 --> 00:06:39.370
into a few interesting people. So in 2006, that

00:06:39.370 --> 00:06:42.930
was the time when he's a senior of college. He

00:06:42.930 --> 00:06:46.470
ran into someone. who you may have heard of his

00:06:46.470 --> 00:06:50.550
brand, but the name Wu Tao was the person that

00:06:50.550 --> 00:06:53.009
Liang Wanfeng had encountered. What's Wu Tao's

00:06:53.009 --> 00:06:55.589
brand? Wu Tao's brand is the number one drone

00:06:55.589 --> 00:06:57.949
brand in the world. Currently, it's called DJI.

00:06:58.209 --> 00:07:01.290
So back in the days, Wu Tao is from Zhejiang,

00:07:01.389 --> 00:07:04.769
and Wu Tao had invited Liang Wanfeng to join

00:07:04.769 --> 00:07:09.189
the company. Had he started to join the company,

00:07:09.310 --> 00:07:12.310
the DJI, back then, Liang Wanfeng would have

00:07:12.310 --> 00:07:16.439
also been... equally financially free or had

00:07:16.439 --> 00:07:19.699
achieved his own wealth accumulation by now,

00:07:19.819 --> 00:07:25.689
but he did not. He has his own thing, I guess.

00:07:25.870 --> 00:07:28.910
Yeah. Hardware may be not his thing as much.

00:07:29.089 --> 00:07:31.829
For sure. So Liang Wenfeng seemed to have very

00:07:31.829 --> 00:07:35.470
early and strong conviction to artificial intelligence.

00:07:35.810 --> 00:07:38.709
And he believes that this will eventually be

00:07:38.709 --> 00:07:41.050
the force that changes the world, not so much

00:07:41.050 --> 00:07:46.009
as a drone, I guess. Even back in 20... With

00:07:46.009 --> 00:07:50.350
his 2006. In 2006, he said no to Wu Tao. And

00:07:50.350 --> 00:07:52.569
a few years later, the actual partner that he

00:07:52.569 --> 00:07:56.069
had picked was a machine learning major. So I

00:07:56.069 --> 00:07:57.930
think maybe something was in his mind that he

00:07:57.930 --> 00:07:59.930
wanted someone who is, you know, doing mathematics

00:07:59.930 --> 00:08:01.589
or in the computer science or in the machine

00:08:01.589 --> 00:08:03.949
learning field. So that seed was planted early

00:08:03.949 --> 00:08:06.209
on. Very interesting. Yeah. Yeah. He has his

00:08:06.209 --> 00:08:10.189
ideas. Yeah. Yeah. So, OK, so at the end of his

00:08:10.189 --> 00:08:12.910
undergrad, you know, like like any other student

00:08:12.910 --> 00:08:14.850
in China, you would think of where do I go to

00:08:14.850 --> 00:08:17.410
graduate school? There's no way out directly

00:08:17.410 --> 00:08:21.149
from undergrad to job force. Yeah. So starting

00:08:21.149 --> 00:08:25.009
in 2007, he once again enrolled in Zhejiang University

00:08:25.009 --> 00:08:28.009
for his graduate school. And at that time, he

00:08:28.009 --> 00:08:31.110
enrolled in information and communications engineering.

00:08:31.509 --> 00:08:34.250
Yeah. But think about this. The second year of

00:08:34.250 --> 00:08:38.730
grad school, which is 2008, the world global

00:08:38.730 --> 00:08:41.750
financial crisis happened. Right. Financial crisis

00:08:41.750 --> 00:08:44.909
happened. While he was doing grad school, he

00:08:44.909 --> 00:08:49.129
became very interested in the financial system.

00:08:49.769 --> 00:08:52.009
itself and also the crisis, he was thinking,

00:08:52.049 --> 00:08:55.190
how do I leverage this into something that could

00:08:55.190 --> 00:08:57.789
really benefit me? So rumor has it that he was

00:08:57.789 --> 00:09:01.090
hiding in Chengdu and not so much spending time

00:09:01.090 --> 00:09:05.049
in Zhejiang. And he had spent time in Chengdu

00:09:05.049 --> 00:09:08.970
and learning how to code. And he had also written

00:09:08.970 --> 00:09:13.269
a automatic trading system while he was in Chengdu.

00:09:13.370 --> 00:09:16.909
And this program later on became the foundation

00:09:16.909 --> 00:09:19.919
for the trading algorithm. that he had started

00:09:19.919 --> 00:09:24.159
but this journey started as early as 2008 wait

00:09:24.159 --> 00:09:28.360
so are you saying he skipped school for his own

00:09:28.360 --> 00:09:30.799
startup you know what from all of the research

00:09:30.799 --> 00:09:33.320
it is impossible to actually know whether he

00:09:33.320 --> 00:09:36.259
attended school but but it but he was just so

00:09:36.259 --> 00:09:40.460
intrigued and he kind of started his own company

00:09:40.460 --> 00:09:43.799
like sort of the seed round of his company the

00:09:43.799 --> 00:09:47.210
the very early stage of it in 2008 he just he

00:09:47.210 --> 00:09:50.350
just couldn't contain the curiosity of you know

00:09:50.350 --> 00:09:52.350
doing more in trading and doing more in writing

00:09:52.350 --> 00:09:54.629
algorithm right that's very interesting i mean

00:09:54.629 --> 00:09:58.509
without revealing my age i started college in

00:09:58.509 --> 00:10:08.730
2008 and i remember do the math guys um so in

00:10:08.730 --> 00:10:12.929
my first year of college all everyone could talk

00:10:12.929 --> 00:10:16.000
about What's the financial crisis? I mean, spoiler,

00:10:16.179 --> 00:10:19.080
I study economics. Obviously, my classes are

00:10:19.080 --> 00:10:21.980
all about like, hey, if you graduated this year,

00:10:22.039 --> 00:10:23.799
you might not be able to get a job. But great

00:10:23.799 --> 00:10:28.360
that you're still a freshman. So I kind of. You

00:10:28.360 --> 00:10:30.539
know the mindset. Why he, you know, all of a

00:10:30.539 --> 00:10:33.779
sudden in 2008, like is trying to figure out

00:10:33.779 --> 00:10:36.659
a way to combine what he's studying with the

00:10:36.659 --> 00:10:39.080
financial market. Yeah. Also rumor has it that

00:10:39.080 --> 00:10:42.139
in 2008 was the time that he had the initial

00:10:42.139 --> 00:10:46.139
funding or the small principle of a couple 15K

00:10:46.139 --> 00:10:49.980
US dollar. It's like 80K RMB. It's not a lot.

00:10:49.980 --> 00:10:52.620
It is a very humble beginning, but that's sort

00:10:52.620 --> 00:10:54.620
of the money that he started with. Anyways, we'll

00:10:54.620 --> 00:10:55.960
come back to that number later. You just have

00:10:55.960 --> 00:10:57.600
to remember it's a various amount of money. Small

00:10:57.600 --> 00:11:01.019
amount of money. 15K USD. Exactly, yeah. So then

00:11:01.019 --> 00:11:04.620
by 2010, time for graduation, he graduated from

00:11:04.620 --> 00:11:07.159
Zhejiang University with a degree in information

00:11:07.159 --> 00:11:09.860
and communications engineering, and specifically

00:11:09.860 --> 00:11:14.039
with a bit of a touch on machine vision or computer

00:11:14.039 --> 00:11:17.659
vision. And his thesis was on object tracking

00:11:17.659 --> 00:11:21.259
with the PTZ camera, which is the pan -tilt -zoom,

00:11:21.340 --> 00:11:23.379
which in layman's terms is just a surveillance

00:11:23.379 --> 00:11:27.480
camera, right? So that's what he did. Let me

00:11:27.480 --> 00:11:29.860
just add one more thing before we dive into a

00:11:29.860 --> 00:11:32.879
very interesting event. So in 2013, three years

00:11:32.879 --> 00:11:36.620
afterwards, he had started this company, his

00:11:36.620 --> 00:11:40.440
first ever investment company with a Zhejiang

00:11:40.440 --> 00:11:44.539
University PhD called Xu Jin in Hangzhou. And

00:11:44.539 --> 00:11:47.919
then the two of them started sort of expanding

00:11:47.919 --> 00:11:50.899
on the foundational algorithm that Liang Wenfeng

00:11:50.899 --> 00:11:54.240
started in 2008. And then they were trading primarily

00:11:54.240 --> 00:11:57.529
in the stock market within China. People would

00:11:57.529 --> 00:12:00.009
say that Liang Wenfeng and Xu Jing, what are

00:12:00.009 --> 00:12:02.830
they like? To give you a bit more context. So

00:12:02.830 --> 00:12:05.970
Liang Wenfeng is like the very introverted Ilya

00:12:05.970 --> 00:12:09.350
Soskever kind of tech bro. It's very quiet. And

00:12:09.350 --> 00:12:14.490
Xu Jing, despite being a PhD in machine learning

00:12:14.490 --> 00:12:18.169
and actual LM, he is more of the salesman like

00:12:18.169 --> 00:12:21.830
Sam Altman. So not all PhD are introverted is

00:12:21.830 --> 00:12:26.500
what you're saying? Well, I guess not. first

00:12:26.500 --> 00:12:30.159
ever investment company is called a Hangzhou

00:12:30.159 --> 00:12:35.759
Jacobi or Jacobi I don't know how exactly I can

00:12:35.759 --> 00:12:39.340
pronounce that but the name is from a German

00:12:39.340 --> 00:12:43.860
mathematician whose name Carl I guess Jacobi

00:12:45.049 --> 00:12:49.070
J -A -C -O -B -I. Exactly, yes. And he was the

00:12:49.070 --> 00:12:52.090
German mathematician who made fundamental contributions

00:12:52.090 --> 00:12:55.509
to elliptic functions, dynamics, differential

00:12:55.509 --> 00:12:59.769
equations, determinants, and number theory. Stuff

00:12:59.769 --> 00:13:03.789
you learn if you major in math. Yeah, yeah, yeah.

00:13:04.029 --> 00:13:07.679
And Jacobi's work particularly... The Hamilton

00:13:07.679 --> 00:13:10.360
-Jacobi -Bellman equation, the HJB equation,

00:13:10.500 --> 00:13:12.740
had a significant connection to quantitative

00:13:12.740 --> 00:13:16.620
trading. So this mathematician's work is really

00:13:16.620 --> 00:13:19.200
quite favored by all who are in quant trading.

00:13:19.279 --> 00:13:23.299
And if you're in quant trading circle... I guess

00:13:23.299 --> 00:13:25.639
mathematicians would be, this particular person

00:13:25.639 --> 00:13:28.059
would be very well known. So they nerded out

00:13:28.059 --> 00:13:30.720
on their company name. Yeah, nerded out on the

00:13:30.720 --> 00:13:32.659
company name. That's really cool. Yeah. Also,

00:13:32.700 --> 00:13:36.360
I guess a bit more on the partner, Xu Jin, like

00:13:36.360 --> 00:13:38.940
XJ. Yeah, this is a very Chinese name. And I

00:13:38.940 --> 00:13:40.820
do not blame you if you cannot reproduce the

00:13:40.820 --> 00:13:44.769
name. Do not worry about that. is also a graduate

00:13:44.769 --> 00:13:50.269
from Jersey University. And his major, when he

00:13:50.269 --> 00:13:53.309
was doing that PhD, is in machine learning. So

00:13:53.309 --> 00:13:56.830
he's actually a machine learning PhD and specializing

00:13:56.830 --> 00:14:02.190
in the LLM back then in 2013. Very good foresight.

00:14:02.470 --> 00:14:06.370
Very, very early on. He was also very lucky.

00:14:06.590 --> 00:14:11.549
So in 2010, he had already been working and had

00:14:11.549 --> 00:14:15.769
led one company public within China. And later

00:14:15.769 --> 00:14:19.850
on, he himself joined Huawei's Shanghai Research

00:14:19.850 --> 00:14:23.129
Institute. Very hardcore and very good at sales.

00:14:23.610 --> 00:14:27.549
He's also very talented in the strategy making

00:14:27.549 --> 00:14:30.710
in quant trading firms. And the annualized returns

00:14:30.710 --> 00:14:33.250
with the product that he created sustained a

00:14:33.250 --> 00:14:39.289
38 .63 % return on an annual basis. So that's

00:14:39.289 --> 00:14:42.409
not bad for a quant trading firm product. This

00:14:42.409 --> 00:14:48.000
is his... or this is Jacoby? This is the early

00:14:48.000 --> 00:14:50.899
days of Huanfeng, which is Jacoby. Yeah, so it

00:14:50.899 --> 00:14:54.799
was the two of them. But the IPO company is another

00:14:54.799 --> 00:14:59.159
company that he left. So he was running two companies.

00:14:59.159 --> 00:15:03.159
One IPO, one was Huanfeng, which is High Flyer.

00:15:03.159 --> 00:15:06.360
Yeah. Wow. And he's also very young. And he is

00:15:06.360 --> 00:15:09.379
in the special talent class within Zhejiang University,

00:15:09.620 --> 00:15:11.519
like a crazy talent class, like Qinghua, Beta,

00:15:11.679 --> 00:15:14.080
they all love it. Yeah, the special class. Yeah.

00:15:14.720 --> 00:15:18.159
Anyways. Genius class. Genius class. Yeah. You

00:15:18.159 --> 00:15:20.559
just have to let the genius use the talent that

00:15:20.559 --> 00:15:22.399
they have and, you know, make things happen.

00:15:22.620 --> 00:15:25.399
I think this is a sort of a deep six philosophy,

00:15:26.320 --> 00:15:28.940
which we will go into later on. Yeah, we'll go

00:15:28.940 --> 00:15:32.360
into that later on. So one thing that's really

00:15:32.360 --> 00:15:35.840
worth mentioning that happened in 2012, which

00:15:35.840 --> 00:15:40.279
is a paper that is published and it reads. ImageNet

00:15:40.279 --> 00:15:43.559
classification with deep convolutional neural

00:15:43.559 --> 00:15:46.840
network. Do you remember the name AlexNet project?

00:15:46.960 --> 00:15:51.519
Yes. But for those of you who don't, it's a machine

00:15:51.519 --> 00:15:54.100
learning competition, right? Right, exactly.

00:15:54.419 --> 00:15:58.539
So it is a competition that is called ImageNet

00:15:58.539 --> 00:16:03.110
competition. a sort of a robotic slash machine

00:16:03.110 --> 00:16:05.830
learning scientist, Li Feifei in the Chinese

00:16:05.830 --> 00:16:09.929
term, she started this competition in 2006. And

00:16:09.929 --> 00:16:14.070
ImageNet is a project, it is a very large visual

00:16:14.070 --> 00:16:17.690
database designed for visual object recognition.

00:16:18.269 --> 00:16:21.850
And it's a software -based contest. So you basically

00:16:21.850 --> 00:16:25.269
just tag a lot of visual objects and see which

00:16:25.269 --> 00:16:28.470
one of the software can recognize in accuracy

00:16:28.470 --> 00:16:30.730
which one of the object it is. And the database

00:16:30.730 --> 00:16:33.110
itself is very large. Yeah, so I remember it

00:16:33.110 --> 00:16:36.590
was... So the ImageNet competition basically

00:16:36.590 --> 00:16:41.090
gave away a data set of images. And whoever gets

00:16:41.090 --> 00:16:46.940
the most accurate answer... That's right. So

00:16:46.940 --> 00:16:51.340
the implication of AlexNet or the ImageNet competition

00:16:51.340 --> 00:16:56.840
is it revolutionized image recognition by demonstrating

00:16:56.840 --> 00:17:00.559
the power of deep convolutional neural network,

00:17:00.679 --> 00:17:05.220
CNN, which is sort of the original deep learning

00:17:05.220 --> 00:17:08.940
neural network. Neural network. Yeah. Quick fun

00:17:08.940 --> 00:17:12.140
fact about that. Machine learning has been in

00:17:12.140 --> 00:17:14.539
existence for years, and the concept of neural

00:17:14.539 --> 00:17:17.759
network actually is not new. It's something that

00:17:17.759 --> 00:17:21.059
was actually, I think it started in the 70s when

00:17:21.059 --> 00:17:24.319
there's papers published about it. But when it

00:17:24.319 --> 00:17:28.539
was kind of first used in research purposes,

00:17:28.799 --> 00:17:33.259
I think it didn't perform so well. So different

00:17:33.259 --> 00:17:37.720
branches of machine learning kind of started

00:17:37.720 --> 00:17:41.349
off. better than neural network right and so

00:17:41.349 --> 00:17:45.150
it's been kind of in its winters for the last

00:17:45.150 --> 00:17:50.950
30 40 years and a different branch which is more

00:17:50.950 --> 00:17:56.890
based on teaching teaching the ai stuff and and

00:17:56.890 --> 00:17:59.829
making it learn from like what we feed them rather

00:17:59.829 --> 00:18:02.029
than using kind of neural network which is how

00:18:02.029 --> 00:18:06.529
our brain functions true led uh most of the innovation

00:18:06.529 --> 00:18:10.539
in the past 30, 40 years until this very moment,

00:18:10.680 --> 00:18:15.410
which is AlexNet, which proved. that neural networks

00:18:15.410 --> 00:18:19.309
are the way to go and that kind of shifted all

00:18:19.309 --> 00:18:22.029
the research going forward exactly and another

00:18:22.029 --> 00:18:24.309
fun fact which is do you know who are the three

00:18:24.309 --> 00:18:27.549
authors that published the alex net project are

00:18:27.549 --> 00:18:30.089
the three alums from your particular university

00:18:30.089 --> 00:18:33.849
university of toronto yay toronto toronto it's

00:18:33.849 --> 00:18:37.349
a alex frizhevsky if i mispronounce that name

00:18:37.349 --> 00:18:40.450
i'm so sorry and ilya suskever which later worked

00:18:40.450 --> 00:18:42.670
for open ai and jeffrey hinton which later worked

00:18:42.670 --> 00:18:45.759
for google Yeah. Yes. I vividly remember the

00:18:45.759 --> 00:18:48.720
email of Jeffrey quitting Google. Never met him

00:18:48.720 --> 00:18:51.099
in U of T, but great professor. Great professor.

00:18:53.390 --> 00:18:57.109
Exactly. So 20, all the way throughout Liang

00:18:57.109 --> 00:18:59.589
Wanfeng's graduate school time, seems pretty

00:18:59.589 --> 00:19:02.490
exciting. He finished his graduate study, found

00:19:02.490 --> 00:19:05.369
a partner, and his partner and him both have

00:19:05.369 --> 00:19:07.430
a lot of establishments and studies in machine

00:19:07.430 --> 00:19:09.690
learning in this particular sector. And in 2012,

00:19:10.029 --> 00:19:12.450
this very exciting competition happened, which

00:19:12.450 --> 00:19:15.069
I think I can only assume that if you're in this

00:19:15.069 --> 00:19:17.509
particular segment, you must know what had happened.

00:19:17.910 --> 00:19:22.809
On to chapter two, quant fund years. how fast

00:19:22.809 --> 00:19:26.950
can you really grow? Can you make money? It is

00:19:26.950 --> 00:19:31.349
really the speed that how fast it grew. And looking

00:19:31.349 --> 00:19:34.950
at the speed of a lot of different quant funds

00:19:34.950 --> 00:19:38.369
and how they grew. To be honest, many quant funds

00:19:38.369 --> 00:19:42.470
in the US, they are sort of an extension of an

00:19:42.470 --> 00:19:44.869
existing hedge fund. So they are always starting

00:19:44.869 --> 00:19:47.210
somewhere, I guess, besides Renaissance. That's

00:19:47.210 --> 00:19:48.950
because of Jim Summers also started from nowhere.

00:19:49.730 --> 00:19:53.880
But Liang Wenfeng, The speed that he was able

00:19:53.880 --> 00:19:58.180
to accumulate wealth was really quite fast. And

00:19:58.180 --> 00:20:01.059
let's take a look at what he had done between

00:20:01.059 --> 00:20:05.400
the age of 22 after he graduated and 30. Yeah,

00:20:05.400 --> 00:20:08.299
100 times, right? Yeah, 100 times. Yes, the end

00:20:08.299 --> 00:20:12.519
result is 100 times AUM. It's quite fast. So

00:20:12.519 --> 00:20:16.839
before we go into the numbers, let's do a quick

00:20:16.839 --> 00:20:21.579
definition on what is a quant trading. hedge

00:20:21.579 --> 00:20:24.460
fund yes teach us how to make money oh my god

00:20:24.460 --> 00:20:28.220
actually you should be doing this we we have

00:20:28.220 --> 00:20:30.420
about it after yeah let's let's talk about it

00:20:30.420 --> 00:20:34.380
we We actually went through a trading software

00:20:34.380 --> 00:20:37.039
and saw what's the current offering for Liang

00:20:37.039 --> 00:20:39.140
Wenfeng's company. And it seems that it's performing

00:20:39.140 --> 00:20:41.640
very well at this. Yeah, it's not a bad idea.

00:20:41.740 --> 00:20:46.279
It is percent of market. But it is listed as

00:20:46.279 --> 00:20:48.819
very high risk at this point. So we'll have to.

00:20:48.900 --> 00:20:51.579
The most risky type of investment. But most,

00:20:51.579 --> 00:20:56.400
I think, Chinese. hedge funds or chinese private

00:20:56.400 --> 00:21:00.440
funds are listed in the r5 most risky category

00:21:00.440 --> 00:21:04.240
right and you do have to have a minimum amount

00:21:04.240 --> 00:21:07.480
of sort of capital to be able to even buy that

00:21:07.480 --> 00:21:09.819
ticket it's not like you can just have 20k and

00:21:09.819 --> 00:21:11.660
then enter no it's not that kind of deal it's

00:21:11.660 --> 00:21:14.480
a game for the rich which we're not qualified

00:21:14.480 --> 00:21:20.359
yet a quantitative fund commonly known as a quantitative

00:21:20.359 --> 00:21:23.960
trading hedge fund It utilizes computer algorithms

00:21:23.960 --> 00:21:27.559
and quantitative methods to identify and trade

00:21:27.559 --> 00:21:32.019
most commonly stocks and commodities, etc. And

00:21:32.019 --> 00:21:35.559
these funds, they typically use a combination

00:21:35.559 --> 00:21:38.700
of mathematical and statistical models and along

00:21:38.700 --> 00:21:41.079
with algorithmic strategies for trading opportunities,

00:21:41.200 --> 00:21:44.470
identification and execution, a .k .a. You don't

00:21:44.470 --> 00:21:46.529
need a person to be executing a trading strategy.

00:21:46.789 --> 00:21:48.589
You make the strategy first and then you execute.

00:21:48.710 --> 00:21:53.369
In both phases, you use an algorithm to do that

00:21:53.369 --> 00:21:55.009
decision making and then to do that execution.

00:21:55.690 --> 00:21:58.869
And it heavily depends on quantitative analysis

00:21:58.869 --> 00:22:01.829
for investment decisions and eliminating subjective

00:22:01.829 --> 00:22:04.529
human judgments and emotional factors. I think

00:22:04.529 --> 00:22:06.490
this is probably the hardest part in trading

00:22:06.490 --> 00:22:09.130
because when you see that the price goes down,

00:22:09.250 --> 00:22:12.970
that feeling... You want to click sell. You want

00:22:12.970 --> 00:22:16.529
to... click sell but so the beauty of quant trading

00:22:16.529 --> 00:22:21.230
is um it is a complete rational uh system uh

00:22:21.230 --> 00:22:24.710
you program into the algorithm that when whatever

00:22:24.710 --> 00:22:28.609
number hits you know goes up then you sell when

00:22:28.609 --> 00:22:32.430
it when it drops again it right you you you can

00:22:32.430 --> 00:22:35.369
buy in again so it is a less emotional of course

00:22:35.369 --> 00:22:39.690
than human beings and these fund would use vast

00:22:39.690 --> 00:22:43.789
amount of data to understand the market efficiencies

00:22:43.789 --> 00:22:46.069
and understand the market trends. Typically,

00:22:46.069 --> 00:22:48.690
a quantitative trading team would be composed

00:22:48.690 --> 00:22:51.819
of... mathematicians, statisticians, computer

00:22:51.819 --> 00:22:55.440
scientists, which is very different from the

00:22:55.440 --> 00:22:58.599
traditional traders who are sort of working the

00:22:58.599 --> 00:23:01.759
finance and hedge fund segment. Most commonly,

00:23:01.759 --> 00:23:04.640
if you read any financial newspaper or, you know,

00:23:04.660 --> 00:23:07.660
just read any stories about the con trading firms,

00:23:08.019 --> 00:23:12.980
three strategies would show up very often, which

00:23:12.980 --> 00:23:15.319
is the statistical arbitrage, the high frequency

00:23:15.319 --> 00:23:17.960
trading and factory investing. Please explain.

00:23:18.259 --> 00:23:23.220
I will try my best. Statistical arbitrage seeks

00:23:23.220 --> 00:23:26.519
to profit from temporary price discrepancies

00:23:26.519 --> 00:23:29.359
between assets that have statistical relationship.

00:23:29.519 --> 00:23:32.099
This is... Kind of what I described before. So

00:23:32.099 --> 00:23:36.400
buy high, sell low, but with some kind of relationship

00:23:36.400 --> 00:23:42.059
between the stocks. Exactly. Typically and very

00:23:42.059 --> 00:23:45.339
popularly used in commodities because it has

00:23:45.339 --> 00:23:47.900
a very clear demand and supply relationship and

00:23:47.900 --> 00:23:49.680
then you have a lot of historical data to refer

00:23:49.680 --> 00:23:52.420
to when you make a decision. Therefore, you're

00:23:52.420 --> 00:23:55.900
using the past historical data to foresee what's

00:23:55.900 --> 00:23:58.529
going to happen in the future. the demand and

00:23:58.529 --> 00:24:01.450
supply would happen. So if you see a relationship

00:24:01.450 --> 00:24:05.789
between crude oil and natural gas, and one goes

00:24:05.789 --> 00:24:07.609
up and the other doesn't, there's an opportunity

00:24:07.609 --> 00:24:09.910
to profit from it. I guess in Barry Lehman's

00:24:09.910 --> 00:24:12.450
term, yes, that would be an opportunity. And

00:24:12.450 --> 00:24:14.490
high -frequency trading is just straightforward,

00:24:14.710 --> 00:24:17.170
as in you're taking advantage on how fast you

00:24:17.170 --> 00:24:20.869
can trade. Executing large volumes of order at

00:24:20.869 --> 00:24:23.470
extremely high speed, as in milliseconds. And

00:24:23.470 --> 00:24:27.180
I guess a little bit of a note here. What we

00:24:27.180 --> 00:24:29.099
will discuss later in terms of the high frequency

00:24:29.099 --> 00:24:31.819
trading in China, I don't think it will be exactly

00:24:31.819 --> 00:24:34.279
what's high frequency in the States. There might

00:24:34.279 --> 00:24:37.099
be a difference. Lower frequency, high frequency

00:24:37.099 --> 00:24:42.829
trading. Lower high frequency trading. So honestly,

00:24:42.990 --> 00:24:44.970
I'm not a trader myself, so I wouldn't know exactly

00:24:44.970 --> 00:24:48.589
how many milliseconds like in China it would

00:24:48.589 --> 00:24:51.210
be traded. But based on the like, first of all,

00:24:51.250 --> 00:24:53.309
Chinese government and Chinese stock market just

00:24:53.309 --> 00:24:55.130
doesn't like high frequency trading. If it's

00:24:55.130 --> 00:24:56.789
commodity, then you can do high frequency trading.

00:24:56.890 --> 00:24:59.029
But if it's stock market, then absolutely no

00:24:59.029 --> 00:25:01.230
go zone. So arbitrage itself, this idea is not

00:25:01.230 --> 00:25:04.109
very favored by the financial market here. So

00:25:04.109 --> 00:25:07.089
the second one, high frequency trading is more

00:25:07.089 --> 00:25:10.910
like trading frequently to benefit from increase

00:25:10.910 --> 00:25:13.880
in. decrease during the milliseconds the more

00:25:13.880 --> 00:25:17.180
volatile the better it is okay so if you were

00:25:17.180 --> 00:25:20.519
if you were to be experiencing a very high high

00:25:20.519 --> 00:25:23.380
or a low low then and if you can benefit off

00:25:23.380 --> 00:25:26.259
of speed and you can predict um you know the

00:25:26.259 --> 00:25:29.400
prices then yeah there might be a lot of opportunities

00:25:29.400 --> 00:25:33.279
um so especially during financial crisis this

00:25:33.279 --> 00:25:36.500
should be you know a relative combined with statistical

00:25:36.500 --> 00:25:38.740
arbitrage right you'll be making a lot of money

00:25:38.740 --> 00:25:40.940
this is why i think a lot of The high frequency

00:25:40.940 --> 00:25:45.859
trading firm actually earns, like has a lot higher

00:25:45.859 --> 00:25:49.859
return during the market volatility and downtimes.

00:25:49.900 --> 00:25:53.200
Exactly, exactly. Yeah. So this relies on fast

00:25:53.200 --> 00:25:56.529
technology and low latency. And finally, factor

00:25:56.529 --> 00:25:59.730
investing, which I think is what a lot of Chinese

00:25:59.730 --> 00:26:02.490
quant trading firms have applied in the past,

00:26:02.509 --> 00:26:04.349
perhaps a combination with the previous two,

00:26:04.430 --> 00:26:06.450
but this one is definitely most favored and most

00:26:06.450 --> 00:26:09.369
pointed out, which is factor investing. You have

00:26:09.369 --> 00:26:12.509
a lot of factors. You build portfolios based

00:26:12.509 --> 00:26:16.869
on different quantitative factors or characters

00:26:16.869 --> 00:26:20.950
that are historically been associated with high

00:26:20.950 --> 00:26:25.269
returns. So for example, Crude oil is associated

00:26:25.269 --> 00:26:29.190
with factor A, B, and C and you program your

00:26:29.190 --> 00:26:34.500
algorithm to specifically be be related to your

00:26:34.500 --> 00:26:36.920
factors A, B, and C? And should factor A, B,

00:26:36.940 --> 00:26:39.660
and C were hit and you buy in or you sell? Oh,

00:26:39.680 --> 00:26:43.799
so it's like building a relational diagram, like

00:26:43.799 --> 00:26:47.259
a tree diagram of all the, whether it's stocks

00:26:47.259 --> 00:26:51.460
or commodities that you want to trade and figure

00:26:51.460 --> 00:26:54.799
out what is the leading factors if A, then B.

00:26:55.039 --> 00:26:58.599
Exactly. And I track A. So if A changes, then

00:26:58.599 --> 00:27:01.500
I buy or sell B. But here's the thing. I'm not

00:27:01.500 --> 00:27:03.150
a quantitor. myself but i've had a conversation

00:27:03.150 --> 00:27:05.710
with a real quant trader and what he had been

00:27:05.710 --> 00:27:08.069
telling me was you always need to be finding

00:27:08.069 --> 00:27:11.309
what's the next new factor factors actually change

00:27:11.309 --> 00:27:14.329
throughout the years every factor can only be

00:27:14.329 --> 00:27:17.190
useful for a certain period of time and if everyone

00:27:17.190 --> 00:27:19.690
is using the same factor then the factor loses

00:27:19.690 --> 00:27:25.049
its efficiency ah so it's like it's almost a

00:27:25.049 --> 00:27:30.700
competition of which firm which fund finds the

00:27:30.700 --> 00:27:33.019
newest factor because if you don't find new ones

00:27:33.019 --> 00:27:36.460
you can't profit from the old ones or it's overcrowded

00:27:36.460 --> 00:27:39.680
or your annualized return would start to decrease

00:27:39.680 --> 00:27:43.380
okay so it's a race to see who finds the best

00:27:43.380 --> 00:27:46.440
new factor fast constantly exactly constantly

00:27:46.440 --> 00:27:48.500
it's a lot of stress it's a lot of stress okay

00:27:48.500 --> 00:27:51.539
which is which is why you would need really a

00:27:51.539 --> 00:27:54.819
group of statisticians mathematicians and computer

00:27:54.819 --> 00:27:56.440
scientists to be able to do all of that you're

00:27:56.440 --> 00:27:59.180
just analyzing a bunch of data every single day

00:27:59.180 --> 00:28:01.859
and finding what's the new trend what's the new

00:28:01.859 --> 00:28:03.759
factor and how can I be arranging all of these

00:28:03.759 --> 00:28:05.720
in my trading algorithm it's very interesting

00:28:05.720 --> 00:28:08.099
oh my god it sounds exhausting to me it sounds

00:28:08.099 --> 00:28:11.960
like my day job Linda has a lot of potential

00:28:11.960 --> 00:28:14.180
going into a trading firm without all the yeah

00:28:14.180 --> 00:28:18.220
when's your birthday I'll be buying books let's

00:28:18.220 --> 00:28:21.460
understand solve the market let's do that Okay,

00:28:21.500 --> 00:28:25.779
coming back to Liang Wenfeng. So remember that

00:28:25.779 --> 00:28:29.880
he had a very small principal money by the end

00:28:29.880 --> 00:28:34.339
of his graduate school, right? And by 2015, this

00:28:34.339 --> 00:28:38.640
is when he is 30 years old, the market has already

00:28:38.640 --> 00:28:41.920
experienced seven years of bear and bull market

00:28:41.920 --> 00:28:45.160
after the global financial crisis. And with that

00:28:45.160 --> 00:28:49.119
fluctuation and volatility, oh my goodness. By

00:28:49.119 --> 00:28:54.839
he is 30, he had made 100 million RMB. Obviously,

00:28:54.920 --> 00:28:57.480
this is not only from the 80K that he started,

00:28:57.559 --> 00:29:01.509
as we were saying earlier. probably other LPs

00:29:01.509 --> 00:29:04.730
join later on. But all because of family and

00:29:04.730 --> 00:29:06.950
friends, I guess. But that's all because he has

00:29:06.950 --> 00:29:10.569
demonstrated that I am able to deliver such a

00:29:10.569 --> 00:29:13.670
great annualized return. And it can be as high

00:29:13.670 --> 00:29:18.190
as 160 % for the past seven to eight years. If

00:29:18.190 --> 00:29:22.789
I see 160%, I'd be like... Cumulatively or annualized?

00:29:23.549 --> 00:29:25.829
Annualized. Annualized basis. Annualized 160%.

00:29:25.829 --> 00:29:27.950
I wouldn't last a year. I would be like, take

00:29:27.950 --> 00:29:31.650
all of my money and I will shut up. Take my money.

00:29:31.769 --> 00:29:38.849
Take 25 % of the returns. So his personal wealth

00:29:38.849 --> 00:29:43.250
made it into the 100 million RMB bracket as of

00:29:43.250 --> 00:29:49.170
2015. And that is declared in the news. The AUM

00:29:49.170 --> 00:29:53.109
that he was managing by the year 2015 was 1 billion

00:29:53.109 --> 00:29:57.400
RMB. Wow. yeah yeah 10 times what he has that's

00:29:57.400 --> 00:30:01.640
a lot of money so by that time so 2015 he already

00:30:01.640 --> 00:30:05.220
started you know buying a few gpus just you know

00:30:05.220 --> 00:30:09.119
for fun for fun yeah for fun for so we yeah we

00:30:09.119 --> 00:30:11.220
don't know for sure if he bought the gpus because

00:30:11.220 --> 00:30:13.740
his fund required it or because his personal

00:30:13.740 --> 00:30:17.279
interest of just I don't know, wrench. He just

00:30:17.279 --> 00:30:21.279
needs it as a X informational. What's his major?

00:30:21.339 --> 00:30:23.299
Like he, he just likes exploring these stuff

00:30:23.299 --> 00:30:25.140
and he bought some GPUs and back then there was

00:30:25.140 --> 00:30:28.740
an export control. So, um, how many did he get?

00:30:28.779 --> 00:30:30.299
I think our research said different numbers.

00:30:30.420 --> 00:30:32.559
Well, research has different numbers. I found,

00:30:32.779 --> 00:30:38.299
so 10. nvidia geforce in 2015 i checked the version

00:30:38.299 --> 00:30:41.119
it seems correct geforce seems correct but i

00:30:41.119 --> 00:30:43.299
don't know whether it's 10 or more than 10 so

00:30:43.299 --> 00:30:47.460
i i got 100 from a source in china but that said

00:30:47.460 --> 00:30:52.559
we don't know a lot about this guy in 2015 yeah

00:30:52.559 --> 00:30:56.000
i was in a so somewhere between 10 to 100 right

00:30:56.000 --> 00:31:01.910
uh for 2015 it's a lot it's a lot yeah so but

00:31:01.910 --> 00:31:04.170
but we all know that for someone who is in the

00:31:04.170 --> 00:31:07.650
year of 2015 already started buying gpus very

00:31:07.650 --> 00:31:09.430
actively and knowing what they were doing with

00:31:09.430 --> 00:31:12.789
gpu that falls in a very small group of people

00:31:12.789 --> 00:31:16.670
definitely yeah yeah this is what like 2017 is

00:31:16.670 --> 00:31:18.730
when bite dance and tiktok really kicked off

00:31:18.730 --> 00:31:20.470
so this is they're only starting though yeah

00:31:21.619 --> 00:31:23.960
And Zhang Yiming, I don't think he had any idea

00:31:23.960 --> 00:31:25.880
that this would turn into... Yeah, they definitely

00:31:25.880 --> 00:31:29.180
didn't have GPUs. So he was ahead of the game.

00:31:29.339 --> 00:31:32.180
Yeah. We'll do an episode of ByteDance maybe

00:31:32.180 --> 00:31:35.000
later on. And we'll find out maybe in 2015, like

00:31:35.000 --> 00:31:37.539
how exactly how prepared Zhang Yiming was at

00:31:37.539 --> 00:31:42.200
the time. So then two years later, remember the

00:31:42.200 --> 00:31:44.079
money that he had, right? So two years later

00:31:44.079 --> 00:31:48.960
in 2017, Lan Wangfeng's personal wealth, five

00:31:48.960 --> 00:31:53.740
times, five X'd. It was 500 million RMB. And

00:31:53.740 --> 00:31:57.380
the AUM that he was managing was 2 billion RMB.

00:31:57.460 --> 00:32:02.319
So that doubled again. Yeah, exactly. And in

00:32:02.319 --> 00:32:09.690
2019, he started... developing this special department

00:32:09.690 --> 00:32:14.170
within Huanfeng that specializes in large language

00:32:14.170 --> 00:32:16.009
model. It wasn't a large language model, but

00:32:16.009 --> 00:32:20.190
it was more of a processing cluster that he started,

00:32:20.289 --> 00:32:22.809
which is called the Firefly. And by Firefly,

00:32:22.829 --> 00:32:25.750
by 2019, that's another time where he got more

00:32:25.750 --> 00:32:31.170
cars, right? Yeah, so it was said that Huanfeng's

00:32:31.170 --> 00:32:39.180
Hi -Fi fund used... 1100 gpus for the high flyer

00:32:39.180 --> 00:32:43.099
one which is the ai the gpu cluster and supposedly

00:32:43.099 --> 00:32:50.940
this 1100 gpu is all all used for the fund itself

00:32:50.940 --> 00:32:55.380
but we don't know i mean i highly i think he's

00:32:55.380 --> 00:32:58.859
running his personal projects i don't know yeah

00:32:58.859 --> 00:33:02.519
yeah he just seems always to be working on something

00:33:02.519 --> 00:33:06.579
else right it's it's a lot of gpus for for quantfund

00:33:06.579 --> 00:33:08.920
that is the size that he was managing there's

00:33:08.920 --> 00:33:11.140
i feel like there's definitely some that are

00:33:11.140 --> 00:33:14.559
used for r d for other projects and for sure

00:33:14.559 --> 00:33:16.970
there might be not It might not be the direction

00:33:16.970 --> 00:33:19.250
where Deep Seek is heading today, but there seems

00:33:19.250 --> 00:33:23.289
to be some exploration beyond just using it to

00:33:23.289 --> 00:33:27.109
make money. Right, right, right. So yeah, so

00:33:27.109 --> 00:33:29.970
that was 2019. What also happened in 2019 is

00:33:29.970 --> 00:33:34.990
he was doing so well in quant fund that he was

00:33:34.990 --> 00:33:38.710
given an award for being a top leading quant

00:33:38.710 --> 00:33:42.670
trading firm in China. Was the firm how big is

00:33:42.670 --> 00:33:45.900
it? It's the top, one of the top. It was one

00:33:45.900 --> 00:33:49.980
of the top 10 firms in 2019. And in 2021, he

00:33:49.980 --> 00:33:54.839
became the top five firm. But the 2021 was primarily

00:33:54.839 --> 00:33:57.559
because of the size of the firm. The size of

00:33:57.559 --> 00:34:01.059
AUM grew even larger. The 2021 was the highest

00:34:01.059 --> 00:34:05.039
time. We'll come back to 2021. So 2019, sorry.

00:34:05.119 --> 00:34:08.639
2019, he got an award as part of the top 10.

00:34:08.860 --> 00:34:15.300
Right, exactly. In that award, he was managing

00:34:15.300 --> 00:34:19.820
about 10 billion RMB in 2019. That's already

00:34:19.820 --> 00:34:25.179
10x when he started in 2015, right? And the speech

00:34:25.179 --> 00:34:27.860
that he delivered was called... the outlook of

00:34:27.860 --> 00:34:29.900
China's quant trading, a perspective from a software

00:34:29.900 --> 00:34:32.280
engineer. He's such an engineer. Oh, so he's

00:34:32.280 --> 00:34:35.000
positioning himself as a software engineer. Yeah,

00:34:35.000 --> 00:34:37.719
yeah, yeah. So he likes being a, you know, a

00:34:37.719 --> 00:34:41.019
technocrat. We will get to this in a bit. Let's

00:34:41.019 --> 00:34:43.340
once again look at how much money he was accumulating

00:34:43.340 --> 00:34:46.260
in the next couple of years. So in 2020, you

00:34:46.260 --> 00:34:49.679
know, Firefly One was well put into use and Firefly

00:34:49.679 --> 00:34:54.809
One also created... In two years, 10x, he's 10

00:34:54.809 --> 00:35:00.050
billion. So from 2019, he went from 10 billion

00:35:00.050 --> 00:35:04.210
to 2021, 100 billion. When I look at the number,

00:35:04.230 --> 00:35:06.630
it was the first time for me to count the zeros.

00:35:06.829 --> 00:35:09.590
I wasn't sure, am I reading this number right?

00:35:09.670 --> 00:35:12.710
But it was 100 billion by 2021. That's how fast

00:35:12.710 --> 00:35:15.940
he grew. in one year with an already high base.

00:35:16.119 --> 00:35:19.800
Yeah, that's true. That's a high base for a quant

00:35:19.800 --> 00:35:23.500
trading firm developing at that speed. Because

00:35:23.500 --> 00:35:25.920
usually if you were to be going on to a different

00:35:25.920 --> 00:35:28.860
asset level, then you have to be shifting a lot

00:35:28.860 --> 00:35:32.320
of your strategies. So another callback to an

00:35:32.320 --> 00:35:36.880
earlier comment you made. This 10X, so 2020 is

00:35:36.880 --> 00:35:39.000
the year of COVID where a lot of people are putting

00:35:39.000 --> 00:35:42.860
money into investments. So this 10X partially

00:35:42.860 --> 00:35:45.159
is because of return, partially is because of

00:35:45.159 --> 00:35:47.679
new LPs joining. We don't have that exact breakdown,

00:35:47.820 --> 00:35:52.039
but it is still very, very significant and very

00:35:52.039 --> 00:35:54.920
impressive. Yeah, it is very impressive. You

00:35:54.920 --> 00:35:57.239
know, 10X, I think I absolutely agree with the

00:35:57.239 --> 00:35:59.780
reason that you just provided. 2021 was kind

00:35:59.780 --> 00:36:02.449
of a... No one knows what's the best way to make

00:36:02.449 --> 00:36:05.349
money. So putting money into a fund that obviously

00:36:05.349 --> 00:36:09.969
can deliver results is a good idea. So 100 billion

00:36:09.969 --> 00:36:13.110
RMB roughly translate into 14 billion US dollar,

00:36:13.210 --> 00:36:16.510
which is a reasonably sized quant trading firm.

00:36:16.650 --> 00:36:20.800
Oh, in the US? 14 billion US dollar. That's a

00:36:20.800 --> 00:36:25.599
lot. That's a lot. And how well did he perform

00:36:25.599 --> 00:36:31.219
in 2021? He was targeting the Chinese CSI 500

00:36:31.219 --> 00:36:34.619
stock index, so index fund. He outperformed the

00:36:34.619 --> 00:36:39.480
index by 50%, posting an annual return of 71%,

00:36:39.480 --> 00:36:43.420
thanks to the AI -powered prediction model, Firefly

00:36:43.420 --> 00:36:46.400
1, that forecasted which stocks would perform

00:36:46.400 --> 00:36:50.400
better. This made Liang Wenfeng temporarily one

00:36:50.400 --> 00:36:53.380
of the richest men in China, just the year 2021.

00:36:53.579 --> 00:36:57.000
But very briefly, but very briefly. With 2 %

00:36:57.000 --> 00:36:59.559
management fee, and if you're making that much

00:36:59.559 --> 00:37:02.280
money, you're generating an estimated 200 million

00:37:02.280 --> 00:37:04.380
management fee just for the firm. That's 200

00:37:04.380 --> 00:37:06.659
million. He's still charging only 2%, which is

00:37:06.659 --> 00:37:10.059
actually pretty cheap. 2021 was 2%. I don't know

00:37:10.059 --> 00:37:13.719
for now. Maybe it had gone up. Then I feel like

00:37:13.719 --> 00:37:16.440
that's pretty cheap. For this 75, if it's 71

00:37:16.440 --> 00:37:18.940
% return. Yeah, it was 2 % management fee and

00:37:18.940 --> 00:37:21.820
20 % carry. Okay. Yeah. But think about this.

00:37:21.980 --> 00:37:24.460
Maybe he will be charging more management fee

00:37:24.460 --> 00:37:26.760
for now, given that he has a deep seek that he

00:37:26.760 --> 00:37:29.579
has to, you know, foster, like being the dad

00:37:29.579 --> 00:37:31.400
of a very young kid that spends a lot of money.

00:37:31.699 --> 00:37:35.639
Maybe, yeah. Maybe, yeah. How did Liang Wangfeng

00:37:35.639 --> 00:37:38.820
spend the, you know, 2 % management fee? out

00:37:38.820 --> 00:37:42.780
of the 200 million he spent 155 million to buy

00:37:42.780 --> 00:37:47.599
10 000 pieces of nvidia a100 chips for firefly

00:37:47.599 --> 00:37:51.900
2 that's going to come up he also spent 60 the

00:37:51.900 --> 00:37:56.409
research fund on its ai lab which i think later

00:37:56.409 --> 00:37:59.230
on became what's deep seek today but in 2021

00:37:59.230 --> 00:38:01.690
he's already actively spending a lot of money

00:38:01.690 --> 00:38:04.789
on a the card and b what else what other infrastructure

00:38:04.789 --> 00:38:07.429
can be put onto the research lab so the management

00:38:07.429 --> 00:38:10.969
fee and carry got reinvested into the future

00:38:10.969 --> 00:38:13.409
business exactly well fun too and future business

00:38:13.409 --> 00:38:16.550
yeah didn't take much now that i'm looking at

00:38:16.550 --> 00:38:21.500
the numbers 71 percent means he 10 times right

00:38:21.500 --> 00:38:24.579
but 71 is the return so actually a lot of people

00:38:24.579 --> 00:38:26.539
put money in the fund for sure for sure for sure

00:38:26.539 --> 00:38:29.400
a lot of lps join the fund later on yeah are

00:38:29.400 --> 00:38:31.019
you seriously thinking about putting money i'm

00:38:31.019 --> 00:38:34.559
not i can't afford it because yeah it's uh there's

00:38:34.559 --> 00:38:37.139
the minimum threshold to enter is true true true

00:38:37.139 --> 00:38:39.920
yeah this is not investment advice friends so

00:38:39.920 --> 00:38:42.679
please invest yes very rationally if you're outside

00:38:42.679 --> 00:38:44.960
of china you can't even yeah get in but yeah

00:38:44.960 --> 00:38:49.170
but just something in very Interesting to look

00:38:49.170 --> 00:38:52.289
at the historical returns for this fund. Yeah,

00:38:52.329 --> 00:38:54.949
but remember the year 2021 because we're going

00:38:54.949 --> 00:38:57.730
to come back to that. So let's do a quick summary,

00:38:57.829 --> 00:39:02.150
right? In between Liang Longfeng's 30th birthday

00:39:02.150 --> 00:39:06.429
to his 36th birthday, how much money did he make?

00:39:06.590 --> 00:39:09.469
He delivered five times AUAM in three years,

00:39:09.730 --> 00:39:13.590
10 times AUAM in four years, and finally... A

00:39:13.590 --> 00:39:17.650
hundred times AUM in six years. Let's just say

00:39:17.650 --> 00:39:21.070
this guy is constantly improving things and things

00:39:21.070 --> 00:39:24.929
were upward trending despite like COVID happened

00:39:24.929 --> 00:39:28.349
and a lot of competitors then joined the market.

00:39:28.429 --> 00:39:31.130
As you said, a lot of Wall Street bros came back

00:39:31.130 --> 00:39:34.960
and started. quant trading firms. He still is

00:39:34.960 --> 00:39:38.079
very much in the game and delivering the ground.

00:39:38.119 --> 00:39:40.699
Delivering more than most funds. Yeah, delivering

00:39:40.699 --> 00:39:44.739
a lot of funds. So, okay, remember the 2019 speech

00:39:44.739 --> 00:39:47.840
that he had very emotionally, sort of, I don't

00:39:47.840 --> 00:39:49.539
know how emotional. As a software engineer. As

00:39:49.539 --> 00:39:51.199
a software engineer that he gave on the outlook

00:39:51.199 --> 00:39:54.559
of China's quant trading future. Yeah. So, I

00:39:54.559 --> 00:39:57.639
looked through everything. A few things I can

00:39:57.639 --> 00:40:00.880
tell. First, he likes Jim Simons a lot. Oh, Jim

00:40:00.880 --> 00:40:04.710
Simons, the renter. Renaissance guy, the US equivalent

00:40:04.710 --> 00:40:07.969
of Liang Wenfeng. Exactly. Liang Wenfeng is such

00:40:07.969 --> 00:40:12.409
a fan of Jim Simons that he wrote the foreword

00:40:12.409 --> 00:40:15.389
for the Chinese translation of The Man Who Solved

00:40:15.389 --> 00:40:18.449
the Market. Jim Simons' book? Yes. Oh, wow. Well,

00:40:18.469 --> 00:40:21.110
Jim Simons, he's a... Autobiography about him.

00:40:21.230 --> 00:40:22.650
Right, right, right, right, right. That was written

00:40:22.650 --> 00:40:26.989
by Gregory Zuckerman, a Wall Street Journal writer.

00:40:27.210 --> 00:40:30.949
Oh. I can't believe you remember that. I recently

00:40:30.949 --> 00:40:36.190
finished the book. So the core insights of the

00:40:36.190 --> 00:40:40.690
speech that he had delivered in 2019, I can sum

00:40:40.690 --> 00:40:45.619
up, I guess, three points, which is... He thinks

00:40:45.619 --> 00:40:47.860
that there is going to be massive growth potential

00:40:47.860 --> 00:40:51.960
for quant trading inside of China. It's kind

00:40:51.960 --> 00:40:55.420
of true. Yeah, it's kind of true. This is only

00:40:55.420 --> 00:40:57.500
the beginning of the industry. And Liang Wenfeng

00:40:57.500 --> 00:41:01.719
highlights that the quantitative trading is becoming

00:41:01.719 --> 00:41:04.840
mainstream globally. And he was comparing the

00:41:04.840 --> 00:41:08.099
Chinese market with what's happening in the US,

00:41:08.179 --> 00:41:12.579
like Bridgewater and AQR, Rentech, managing huge

00:41:12.579 --> 00:41:16.789
assets. And he points out that in China, there

00:41:16.789 --> 00:41:19.730
is just so much headroom with the Chinese quant

00:41:19.730 --> 00:41:23.929
market. And comparing to the current AUM that

00:41:23.929 --> 00:41:26.969
is being held in the Chinese quant firms versus

00:41:26.969 --> 00:41:30.210
the trillions of dollars that is being managed

00:41:30.210 --> 00:41:33.250
by the US firms. I think the number is that half

00:41:33.250 --> 00:41:36.730
of the U .S., all trades in the U .S. are coming

00:41:36.730 --> 00:41:38.989
from hedge funds, and China is much lower. So

00:41:38.989 --> 00:41:42.650
by benchmarking, he thinks that there's a lot

00:41:42.650 --> 00:41:44.869
of room for growth. Exactly, a lot of room for

00:41:44.869 --> 00:41:47.949
growth. And second of all... a quick comment

00:41:47.949 --> 00:41:51.949
on the efficiency of the models or the algorithms

00:41:51.949 --> 00:41:54.550
that's being used. So Liang Mengfeng observed

00:41:54.550 --> 00:41:58.429
that the industry in China is improving very

00:41:58.429 --> 00:42:02.329
fast and it's roughly doubling every 18 months

00:42:02.329 --> 00:42:06.690
due to Moore's law. But the average profitability

00:42:06.690 --> 00:42:08.989
hasn't increased because of market efficiency

00:42:08.989 --> 00:42:13.619
is also rising at the same similar pace. So this

00:42:13.619 --> 00:42:16.380
means that the strategy that worked before is

00:42:16.380 --> 00:42:19.039
not going to work as well in the future or over

00:42:19.039 --> 00:42:23.340
time. So he thinks that there must be a shift

00:42:23.340 --> 00:42:25.360
that is moving from traditional multi -factor

00:42:25.360 --> 00:42:26.800
models, which is the one that we talk about,

00:42:26.940 --> 00:42:29.380
to AI and then to more integrated frameworks.

00:42:30.199 --> 00:42:33.179
So he thinks that innovation in this sector with

00:42:33.179 --> 00:42:36.139
the algorithm will continue to go on and it will

00:42:36.139 --> 00:42:38.619
be more complex and more robust to support a

00:42:38.619 --> 00:42:40.440
more complicated and more efficient trading in

00:42:40.440 --> 00:42:42.760
the market. So like more competition among his

00:42:42.760 --> 00:42:45.480
peers are actually driving the market to more

00:42:45.480 --> 00:42:49.449
sophistication. Going forward. Exactly. Exactly.

00:42:49.489 --> 00:42:52.369
Makes sense. Yeah. And finally, and this is the

00:42:52.369 --> 00:42:54.369
one that we sort of discussed very briefly before

00:42:54.369 --> 00:42:58.289
we started recording, which is he then outlined

00:42:58.289 --> 00:43:02.269
a short -term and a long -term strategy. So the

00:43:02.269 --> 00:43:04.210
short -term strategy he thinks moving forward

00:43:04.210 --> 00:43:06.949
is called the multi -strategy integration, which

00:43:06.949 --> 00:43:10.150
is starting from now in the next one or two years,

00:43:10.289 --> 00:43:13.190
you can't just be using one strategy to be able

00:43:13.190 --> 00:43:16.010
to gain all of your profits. You need to be combining.

00:43:16.880 --> 00:43:19.820
And the example that I said that you said was

00:43:19.820 --> 00:43:21.960
not useful was combining high frequency with

00:43:21.960 --> 00:43:23.840
some of the multi -factor trading. And you're

00:43:23.840 --> 00:43:26.400
like, the Chinese market doesn't like high frequency

00:43:26.400 --> 00:43:28.860
at all. So there's no way of integrating that.

00:43:29.199 --> 00:43:33.139
Yeah, so I guess context for everyone who has

00:43:33.139 --> 00:43:37.079
not traded in the Chinese stock market. The Chinese

00:43:37.079 --> 00:43:41.099
government or regulators just allowed same -day

00:43:41.099 --> 00:43:44.340
trading in the stock market. This is only the

00:43:44.340 --> 00:43:46.780
stock market, so you can't trade anything. You

00:43:46.780 --> 00:43:50.239
can't buy anything and sell it on the same day.

00:43:50.760 --> 00:43:53.280
There's other ways to get around it, but it just

00:43:53.280 --> 00:43:56.079
makes high -frequency trading in the stock market

00:43:56.079 --> 00:44:00.019
very difficult. But commodities is fine. Commodity

00:44:00.019 --> 00:44:03.239
is fine, futures and options are fine, but those

00:44:03.239 --> 00:44:09.000
are very limited to non -retail traders and heavily

00:44:09.000 --> 00:44:12.159
regulated because there's been a couple of incidences

00:44:12.159 --> 00:44:15.820
in the past that created a lot of commotion to

00:44:15.820 --> 00:44:20.699
make the regulators more strict. And the government's

00:44:20.699 --> 00:44:25.260
crackdown on the within day or T0 trading in

00:44:25.260 --> 00:44:28.260
the stock market, I think, is also a way to protect

00:44:28.260 --> 00:44:30.480
the people who are in the stock market. Yeah,

00:44:30.480 --> 00:44:33.760
the emotional retail investors of China, i .e.

00:44:33.760 --> 00:44:39.340
my family. The volume is really large. And so

00:44:39.340 --> 00:44:43.380
very much unlike the U .S. scenario where I guess

00:44:43.380 --> 00:44:46.440
80 % of the stock market is consist of agencies

00:44:46.440 --> 00:44:49.039
and institutions and 20 % are just individuals,

00:44:49.239 --> 00:44:52.260
the Chinese is kind of the reverse. So you need

00:44:52.260 --> 00:44:55.880
to be protecting the people, the majority of

00:44:55.880 --> 00:44:57.559
the people who are in the market and that you

00:44:57.559 --> 00:45:01.260
cannot encourage such behavior to dominate, right?

00:45:01.340 --> 00:45:03.000
So that's kind of where that crackdown policy

00:45:03.000 --> 00:45:06.099
is coming from. And OK, coming back to 2019.

00:45:06.300 --> 00:45:09.539
So long term, he thinks that the industry is

00:45:09.539 --> 00:45:12.079
going to move towards quantify fundamental analysis

00:45:12.079 --> 00:45:15.920
and primarily tackling the inefficiencies that

00:45:15.920 --> 00:45:19.679
is still in the in the market. And he says that

00:45:19.679 --> 00:45:22.579
it's feasible and necessary as technical strategies

00:45:22.579 --> 00:45:25.460
will become less profitable, to be honest. Yeah,

00:45:25.539 --> 00:45:27.119
like just going back to fundamental analysis.

00:45:27.960 --> 00:45:31.880
Sorry, it's almost like saying. uh high frequency

00:45:31.880 --> 00:45:34.619
trading or quant based trading will need to make

00:45:34.619 --> 00:45:39.340
way for like a mix strategy yeah i'm blurring

00:45:39.340 --> 00:45:42.630
the lines between long -term quantum and long

00:45:42.630 --> 00:45:44.889
-term fundamental trading yeah sounds like very

00:45:44.889 --> 00:45:47.389
interesting very interesting yeah for it's i

00:45:47.389 --> 00:45:51.170
guess unique to china yeah yeah um yeah i guess

00:45:51.170 --> 00:45:53.650
very unique to china he's suggesting that long

00:45:53.650 --> 00:45:55.929
-term trading or like i guess intrinsic value

00:45:55.929 --> 00:45:57.929
value -based trading should definitely still

00:45:57.929 --> 00:46:00.150
be there but in the meantime like quant trading

00:46:00.150 --> 00:46:02.269
should still serve its purpose and should be

00:46:02.269 --> 00:46:04.329
improving in its strategies to improve efficiencies

00:46:04.329 --> 00:46:07.230
of the overall market anyways so the conclusion

00:46:07.230 --> 00:46:09.809
that he had by the end of the award ceremony

00:46:10.190 --> 00:46:14.429
is as a hedge fund, our mission, aka as HuanfengHF's

00:46:14.429 --> 00:46:17.510
mission, is to improve Chinese secondary markets

00:46:17.510 --> 00:46:20.659
efficiency. Which kind of, he did that. He has

00:46:20.659 --> 00:46:24.099
to say that in a public setting. What is he going

00:46:24.099 --> 00:46:25.780
to say? Like, we're not going to help the market

00:46:25.780 --> 00:46:27.800
liquidity. True, true, true, true, true. That's

00:46:27.800 --> 00:46:30.199
very true. So I think he did what he could, I

00:46:30.199 --> 00:46:33.000
guess, from a market participant's point of view

00:46:33.000 --> 00:46:36.079
or like as a leading quant firm's point of view

00:46:36.079 --> 00:46:38.699
and also as a software engineer. I think for

00:46:38.699 --> 00:46:43.519
anyone who's trading in 2019. Would have a positive

00:46:43.519 --> 00:46:45.679
overview of the market. Because it's positive.

00:46:45.900 --> 00:46:48.539
Yeah, exactly. Because it's positive. It's very

00:46:48.539 --> 00:46:52.079
much up for trending. All right. So, all right.

00:46:52.099 --> 00:46:55.940
So, moving on to the next chapter. sort of a

00:46:55.940 --> 00:46:58.139
midlife crisis kind of chapter, right? Like we

00:46:58.139 --> 00:47:00.300
all hate that chapter, but it's coming up. He

00:47:00.300 --> 00:47:04.920
turned 35. 35. Yeah. So that's not old guys.

00:47:04.980 --> 00:47:08.579
It's not old. 35 is completely not midlife. It's

00:47:08.579 --> 00:47:10.519
a perfect time to take a challenge. Let's just

00:47:10.519 --> 00:47:19.019
say. So 2021, remember 2021 was the best year

00:47:19.019 --> 00:47:22.119
that he had, right? He was managing 100 billion

00:47:22.119 --> 00:47:25.760
RMB in 2021. Firefly One was working just fine.

00:47:26.019 --> 00:47:31.099
And in 2021, he became the leading quant strategy

00:47:31.099 --> 00:47:35.019
hedge funds in China, top four to be specific.

00:47:35.760 --> 00:47:41.000
But in the meantime, later on in 2021 was a serious

00:47:41.000 --> 00:47:44.059
challenge and 2021 became the worst year for

00:47:44.059 --> 00:47:46.619
Liang Wenpeng as well. So officially, what happened?

00:47:46.880 --> 00:47:54.400
Officially, in November 2021, there was an incident

00:47:54.400 --> 00:48:00.619
where the fund experienced continuous max drawdown.

00:48:00.860 --> 00:48:04.019
And I'll explain that in a bit. The return started

00:48:04.019 --> 00:48:07.679
to decrease. over and over again. And starting

00:48:07.679 --> 00:48:10.659
from November, in one and a half months, the

00:48:10.659 --> 00:48:13.360
asset that he was managing decreased from $100

00:48:13.360 --> 00:48:18.219
billion to $80 billion. And it was still continuing

00:48:18.219 --> 00:48:23.880
to decrease. It was so bad that Huanfang HF had

00:48:23.880 --> 00:48:28.079
to issue a public apology to the investors about

00:48:28.079 --> 00:48:32.429
the max drop -down. It was apologizing. and made

00:48:32.429 --> 00:48:34.849
a public claim that it will continue to reduce

00:48:34.849 --> 00:48:39.809
in size. It seems that, you know, better model

00:48:39.809 --> 00:48:43.949
is needed. They cannot be managing such a large

00:48:43.949 --> 00:48:47.869
amount of fund continuously. And, you know, a

00:48:47.869 --> 00:48:50.409
lot of things went wrong. And the apology was

00:48:50.409 --> 00:48:52.349
the only thing that was disclosed. Some other

00:48:52.349 --> 00:48:54.190
funds also did equally bad. And then there were

00:48:54.190 --> 00:48:55.710
guesses that they were using the same factor.

00:48:55.809 --> 00:48:57.929
And, you know, if you're using the same factor,

00:48:58.010 --> 00:49:01.070
then maybe it's not efficient. So everyone's

00:49:01.070 --> 00:49:03.170
using the same strategy, but they're using the

00:49:03.170 --> 00:49:07.789
same wrong strategy at 2021. Well, I cannot know

00:49:07.789 --> 00:49:11.420
like what. would be the exact reason, but the

00:49:11.420 --> 00:49:14.500
return for a lot of the funds at the end of 2021

00:49:14.500 --> 00:49:18.519
were bad. And the only assumption that was fair

00:49:18.519 --> 00:49:22.320
to make was maybe people were banking on the

00:49:22.320 --> 00:49:24.820
same factor and then the factor did not yield

00:49:24.820 --> 00:49:28.500
as good of a return. Yeah. So I guess what's

00:49:28.500 --> 00:49:30.559
the max drawdown in the quant trading and why

00:49:30.559 --> 00:49:33.679
is it very bad for quant funds? A drawdown is

00:49:33.679 --> 00:49:37.480
a decline in value of an investment or portfolio

00:49:37.480 --> 00:49:40.809
from its peak value. it sounds bad but it's also

00:49:40.809 --> 00:49:43.170
i think there's two factors one is the return

00:49:43.170 --> 00:49:46.730
actually declining yeah the other one is i can

00:49:46.730 --> 00:49:49.730
actually as an lp take out the money true um

00:49:49.730 --> 00:49:53.519
so a mix of both like it kind of it's in a circle

00:49:53.519 --> 00:49:55.719
so the return is not good then i'm gonna take

00:49:55.719 --> 00:49:57.699
my money out exactly i think there's specific

00:49:57.699 --> 00:49:59.980
times where you can take your money out like

00:49:59.980 --> 00:50:02.280
you can't just take it out every it's not like

00:50:02.280 --> 00:50:04.219
stock where you can sell it every day right right

00:50:04.219 --> 00:50:07.599
right um but during one month 20 that's a lot

00:50:07.599 --> 00:50:11.219
that's a lot yeah it's also a very rapid loss

00:50:11.219 --> 00:50:15.539
of i think credibility yeah you you just you

00:50:15.539 --> 00:50:18.519
know went up to 100 billion and But you were

00:50:18.519 --> 00:50:20.239
not able to sustain that. So I think that's a

00:50:20.239 --> 00:50:22.460
question for a lot of the con trading firms that's

00:50:22.460 --> 00:50:24.739
up and running these days. Like how long can

00:50:24.739 --> 00:50:28.139
you be sustaining your sort of level of asset?

00:50:28.460 --> 00:50:30.880
Exactly. Yeah. I mean, it's bad enough that he

00:50:30.880 --> 00:50:34.159
has to make a PR statement. Yeah. The challenge

00:50:34.159 --> 00:50:38.659
begins 2022. If you were Liang Wenfeng, you'd

00:50:38.659 --> 00:50:43.019
be thinking of, so, you know, how do I continue?

00:50:43.360 --> 00:50:46.380
How do I proceed from here? Are there ways to

00:50:46.380 --> 00:50:50.599
pivot? So here is the actual circumstances that's

00:50:50.599 --> 00:50:55.059
in the year 2022. Externally, there is continuous

00:50:55.059 --> 00:50:57.619
government crackdown, the one that we just talked

00:50:57.619 --> 00:51:00.179
about. The government does not support or like

00:51:00.179 --> 00:51:03.119
any type of financial arbitrage in the stock

00:51:03.119 --> 00:51:07.880
market. And Liang Wenfeng, he's an initial part

00:51:07.880 --> 00:51:09.480
of the strategy. We're all related to the stock

00:51:09.480 --> 00:51:11.219
market. But obviously, there is the commodity

00:51:11.219 --> 00:51:15.329
bit. But a part of his strategy... probably will

00:51:15.329 --> 00:51:17.650
have to retreat because of the government crackdown.

00:51:17.769 --> 00:51:22.659
Yeah, so the Chinese government has... I guess

00:51:22.659 --> 00:51:26.460
you can say a bigger hand. A big invisible hand.

00:51:26.659 --> 00:51:31.039
A big invisible hand on the market. Yeah. And

00:51:31.039 --> 00:51:35.039
if you look at Bloomberg News, you will find

00:51:35.039 --> 00:51:39.059
out that maybe as early as 2020, there are continuous

00:51:39.059 --> 00:51:41.639
news that's been published about how tight the

00:51:41.639 --> 00:51:43.579
control will be for Chinese quant trading firms.

00:51:43.739 --> 00:51:46.340
It goes on and on and on. The more profitable

00:51:46.340 --> 00:51:49.579
that the quant traders find, I guess the more

00:51:49.579 --> 00:51:52.730
regulations there must be. until today i think

00:51:52.730 --> 00:51:55.210
commodity is still very much welcomed in the

00:51:55.210 --> 00:51:56.949
market but it's just stock market there's a lot

00:51:56.949 --> 00:51:59.309
of rules and regulations yeah yeah it kind of

00:51:59.309 --> 00:52:01.909
makes sense you have to protect the retail investors

00:52:01.909 --> 00:52:06.150
a bit more the retail investors tonight yes um

00:52:06.150 --> 00:52:08.949
okay so also what happened externally is uh by

00:52:08.949 --> 00:52:12.349
the end of 2022 november 30th if you can remember

00:52:12.349 --> 00:52:15.550
what happened on that day is uh the very first

00:52:15.550 --> 00:52:19.820
version of chat gpt uh oh Published by OpenAI.

00:52:19.900 --> 00:52:23.300
Yeah, that is the end of 2022. It's been three

00:52:23.300 --> 00:52:27.340
years? Yeah, it has been three years. But the

00:52:27.340 --> 00:52:32.480
November 2022 Chachaputi wasn't the... chat to

00:52:32.480 --> 00:52:35.000
beauty later on like it wasn't it was 3 .5 it's

00:52:35.000 --> 00:52:37.260
what the one that made all the news right yeah

00:52:37.260 --> 00:52:41.480
yeah yeah 3 .5 exactly and so that was the external

00:52:41.480 --> 00:52:44.639
circumstances right so we were still locked down

00:52:44.639 --> 00:52:46.639
back then we were still locked down we were not

00:52:46.639 --> 00:52:49.099
even we were what were we doing we were we were

00:52:49.099 --> 00:52:53.239
none not november but i think november december

00:52:53.239 --> 00:52:56.000
is when uh everything kind of went haywire and

00:52:56.000 --> 00:52:59.139
yeah yeah yeah country opened up exactly so uh

00:52:59.139 --> 00:53:02.869
so while we were going into A lot of chaos. ChatDVD

00:53:02.869 --> 00:53:05.429
is being published. And also, the fund itself

00:53:05.429 --> 00:53:07.969
has not been performing very well throughout

00:53:07.969 --> 00:53:10.929
the second half of 2021. Performers were declining.

00:53:11.809 --> 00:53:15.030
And according to a lot of financial news, they

00:53:15.030 --> 00:53:18.030
word it as partly due to its AI mistiming trades

00:53:18.030 --> 00:53:19.929
on the market. Not sure how to understand that.

00:53:20.150 --> 00:53:23.130
But, yeah, what do you want to say? So, I think

00:53:23.130 --> 00:53:25.590
the government published, there was also, like,

00:53:25.650 --> 00:53:27.690
stock -wise, the retail investors are also not

00:53:27.690 --> 00:53:31.980
so happy about their return. I remember the government

00:53:31.980 --> 00:53:35.320
kind of blamed it on the hedge funds. Yeah. So,

00:53:35.440 --> 00:53:37.500
yeah, that was during that time. Yeah, yeah,

00:53:37.500 --> 00:53:39.960
yeah. I can imagine. It would make sense. Yeah,

00:53:40.000 --> 00:53:41.820
it's a lot of money lost. Right, right, right.

00:53:41.900 --> 00:53:45.440
A lot of money that's being misused, mistimed.

00:53:45.559 --> 00:53:48.079
Yeah, but it's really, if you look at the historical

00:53:48.079 --> 00:53:51.579
fund return, it was pretty, it was briefly. It's

00:53:51.579 --> 00:53:54.480
decent. Yeah. It's decent. It was briefly down

00:53:54.480 --> 00:53:58.739
in 2021. I think people did panic and withdraw

00:53:58.739 --> 00:54:02.340
their money. But later on, the fund went on and

00:54:02.340 --> 00:54:05.139
it was doing decent returns. I did look at some

00:54:05.139 --> 00:54:08.400
of the numbers of the annualized return all the

00:54:08.400 --> 00:54:12.159
way from 2016 to 2021. In between 2016 and 2018

00:54:12.159 --> 00:54:16.519
for HF, the neutral strategies annualized return

00:54:16.519 --> 00:54:20.679
was in between 20 % and 30%. It was great. Pretty

00:54:20.679 --> 00:54:24.469
good, yeah. And 2019 to 2020, because of... flyer

00:54:24.469 --> 00:54:28.769
flyer one the annualized return and for the csi

00:54:28.769 --> 00:54:34.230
500 equivalent etf it was a it was between 30

00:54:34.230 --> 00:54:39.570
and 40 but in 2021 that return dropped down to

00:54:39.570 --> 00:54:43.510
15 oh so it's still positive i thought people

00:54:43.510 --> 00:54:47.250
are like it's still positive but that may be

00:54:47.250 --> 00:54:51.019
the best performing product uh if you were to

00:54:51.019 --> 00:54:53.739
be looking at all of the products that's from

00:54:53.739 --> 00:54:57.920
hf uh i'm i can maybe it's also there are funds

00:54:57.920 --> 00:55:00.599
that are not making money and this fund the the

00:55:00.599 --> 00:55:04.320
top money making fund is 15 exactly which is

00:55:04.320 --> 00:55:06.639
decent by any means yeah which is decent even

00:55:06.639 --> 00:55:12.320
with an etf i guess so we are moving back to

00:55:12.320 --> 00:55:18.219
2022 So I think by year 2022, because of all

00:55:18.219 --> 00:55:21.400
the internal and external reason, and also I

00:55:21.400 --> 00:55:23.099
think with Liang Wenfeng's personal interest

00:55:23.099 --> 00:55:29.079
development, he was finally able to decide that

00:55:29.079 --> 00:55:33.840
I think it's time to make my LLM company an official

00:55:33.840 --> 00:55:38.400
thing, a .k .a. the AI lab that was within HF

00:55:38.400 --> 00:55:44.559
fund. Moving on to... 2023. 2023. 2023 was the

00:55:44.559 --> 00:55:47.800
year that the artificial intelligence research

00:55:47.800 --> 00:55:52.960
branch and the hedge fund split. Exactly. Or

00:55:52.960 --> 00:55:56.739
he created a new company that is focused on.

00:55:57.400 --> 00:56:00.000
artificial intelligence. Exactly. Take us from

00:56:00.000 --> 00:56:04.739
here. All right. So the AI Liang Wenfeng was

00:56:04.739 --> 00:56:13.239
born. So 2023 May is when he registered the company,

00:56:13.340 --> 00:56:17.800
which is the parent company of DeepSeek. 2023

00:56:17.800 --> 00:56:21.659
may is when the company was officially announced

00:56:21.659 --> 00:56:25.119
deep seek with the announcement actually the

00:56:25.119 --> 00:56:30.099
statement was they're focused are on agi artificial

00:56:30.099 --> 00:56:32.659
general intelligence it was never about application

00:56:32.659 --> 00:56:35.960
it was always about infrastructure this is from

00:56:35.960 --> 00:56:38.440
day one when they established the company that

00:56:38.440 --> 00:56:41.219
was the goal very far -sighted they're very very

00:56:41.219 --> 00:56:44.659
similar to sam oldman exactly so that's actually

00:56:44.659 --> 00:56:47.099
very different from most companies in China.

00:56:47.159 --> 00:56:49.840
Most companies never claim that I'm going to

00:56:49.840 --> 00:56:53.059
have a vision of AGI. Right, exactly. Like I'm

00:56:53.059 --> 00:56:55.539
going to establish a company that is going to

00:56:55.539 --> 00:57:02.000
be benchmarked against open AI, against anthropic.

00:57:02.400 --> 00:57:05.260
I'm not about applications. I'm not about short

00:57:05.260 --> 00:57:09.579
-term ROI. This is very, I would say like non

00:57:09.579 --> 00:57:13.150
-Chinese company mindset in general. non -conventional

00:57:13.150 --> 00:57:17.989
non -conventional exactly um because over the

00:57:17.989 --> 00:57:20.469
past 10 years with the rise of tech companies

00:57:20.469 --> 00:57:24.170
in china it's always about short -term roi it's

00:57:24.170 --> 00:57:27.889
always about can you the the p firms are so spoiled

00:57:27.889 --> 00:57:31.090
that they want ipo in like within seven years

00:57:31.090 --> 00:57:34.929
right that's uh so that was a pretty bold statement

00:57:34.929 --> 00:57:39.940
to make and he actually stuck true to the statement

00:57:39.940 --> 00:57:43.380
to this day. So they published their first LLM

00:57:43.380 --> 00:57:47.480
model six months later in November and it's called

00:57:47.480 --> 00:57:53.519
DeepSeek V1. That made zero slash. Zero slash.

00:57:53.599 --> 00:57:56.960
Zero splash. I mean, nobody heard about them.

00:57:57.139 --> 00:58:00.400
Maybe people in the industry within China. Well,

00:58:00.440 --> 00:58:03.440
no, that is a boutique, cute little model that

00:58:03.440 --> 00:58:05.239
the Chinese has made, but nothing threatening.

00:58:05.440 --> 00:58:10.000
Right, exactly. And the first wave they really

00:58:10.000 --> 00:58:13.940
made was mid -2024, which is really one year

00:58:13.940 --> 00:58:16.340
after they established a company. Still pretty

00:58:16.340 --> 00:58:18.840
fast. And that's when they published DeepSeek

00:58:18.840 --> 00:58:24.469
V2. and uh there's kind of two innovations in

00:58:24.469 --> 00:58:27.489
this model that created some splashed in the

00:58:27.489 --> 00:58:30.750
research market so it's still not uh general

00:58:30.750 --> 00:58:33.449
yet like we haven't heard about it at that point

00:58:33.449 --> 00:58:37.929
right as a market as users of of uh language

00:58:37.929 --> 00:58:42.090
models so the first innovation is called mixture

00:58:42.090 --> 00:58:48.710
of experts moe and it's basically an idea that

00:58:48.710 --> 00:58:53.130
was That started in the 80s or 90s. Nothing new.

00:58:53.449 --> 00:58:56.170
It's nothing new, but they used it, whereas a

00:58:56.170 --> 00:59:00.110
lot of the other models didn't. And to put it

00:59:00.110 --> 00:59:02.590
into application is something that a lot of people

00:59:02.590 --> 00:59:07.170
applauded them for. And the idea is, usually

00:59:07.170 --> 00:59:11.860
in an LLM model, you need a... billions of parameters

00:59:11.860 --> 00:59:16.139
right and that requires a lot of computation

00:59:16.139 --> 00:59:20.880
power so moe architecture actually only calls

00:59:20.880 --> 00:59:25.980
upon the experts so only if you have so right

00:59:25.980 --> 00:59:28.219
now their their model has six billion parameters

00:59:28.219 --> 00:59:32.619
i only call maybe a fraction of that that are

00:59:32.619 --> 00:59:36.659
my experts to this kind of questions and whatnot

00:59:36.659 --> 00:59:41.960
and so that reduced the cost by a lot yeah right

00:59:41.960 --> 00:59:46.059
um so that made some splash in the in the research

00:59:46.059 --> 00:59:49.239
within the industry right the other innovation

00:59:49.239 --> 00:59:51.500
was called not really innovation but also something

00:59:51.500 --> 00:59:54.500
that has used before is called mla but this one

00:59:54.500 --> 00:59:56.820
they kind of created them i think that they're

00:59:56.820 --> 00:59:59.739
the first one to really use this in llm models

00:59:59.739 --> 01:00:03.849
and this is basically a librarian i'm trying

01:00:03.849 --> 01:00:05.889
to put this into like simple term but like if

01:00:05.889 --> 01:00:09.429
you rather than keeping a library at your disposal

01:00:09.429 --> 01:00:12.289
with all the books you kind of condense everything

01:00:12.289 --> 01:00:15.630
into this book of glossaries but then you're

01:00:15.630 --> 01:00:18.329
allowed to zoom into any item and still extract

01:00:18.329 --> 01:00:21.230
the necessary information so another cost -saving

01:00:21.230 --> 01:00:25.530
effort to reduce training costs and a combination

01:00:25.530 --> 01:00:27.530
of those two and some of the other innovations

01:00:27.530 --> 01:00:33.260
kind of made some splash in May of 2024. Okay.

01:00:33.940 --> 01:00:36.739
And... Oh, I can't believe it's a year ago. It's

01:00:36.739 --> 01:00:39.320
only a year ago. Oh, wow. It's exactly a year

01:00:39.320 --> 01:00:44.840
ago. Yeah. Wow. Okay. The reason why all of their

01:00:44.840 --> 01:00:48.420
efforts so far has been around cost saving. It's

01:00:48.420 --> 01:00:50.179
not because they don't have money. We just talked

01:00:50.179 --> 01:00:53.360
about how much money they have. Right? Make them

01:00:53.360 --> 01:00:55.739
rain. Yeah. It's not because they don't have

01:00:55.739 --> 01:00:59.900
money. It's because they can't access... The

01:00:59.900 --> 01:01:05.880
top GPUs. The top performance GPUs. Because as

01:01:05.880 --> 01:01:13.880
of 2022, you can't get any of the A100 cards

01:01:13.880 --> 01:01:17.420
and video cards into China. Welcome to the band

01:01:17.420 --> 01:01:22.559
world. Yeah. So that's why he actually, the team,

01:01:22.800 --> 01:01:26.039
DeepSeek team, actually made an effort. to reduce

01:01:26.039 --> 01:01:28.539
costs and that's what later shocked the market

01:01:28.539 --> 01:01:31.159
start to process the problem from a different

01:01:31.159 --> 01:01:34.139
perspective if i can't buy more how about i just

01:01:34.139 --> 01:01:36.860
use the the downgraded version but use them more

01:01:36.860 --> 01:01:39.340
efficiently and use less computing power exactly

01:01:39.340 --> 01:01:47.099
yeah so it's ironic that the ban of high -performing

01:01:47.099 --> 01:01:52.039
gpus pushed them forced deep sea to be innovative

01:01:52.039 --> 01:01:57.559
which later shocked the entire us if not global

01:01:57.559 --> 01:02:02.099
generative ai industry exactly yeah prior to

01:02:02.099 --> 01:02:06.400
them i think all the other major um ai or aj

01:02:06.400 --> 01:02:09.099
or llm companies in the us they were competing

01:02:09.099 --> 01:02:13.820
on the size whoever has more gpus or has a larger

01:02:13.820 --> 01:02:17.980
data set or essentially have more money can compete

01:02:17.980 --> 01:02:23.409
further that is not wrong But the fact that Deepsea

01:02:23.409 --> 01:02:28.349
can do that made people realize that, oh, you

01:02:28.349 --> 01:02:32.250
don't need to have the most groundbreaking GPUs

01:02:32.250 --> 01:02:36.869
to be creating a very efficient and useful LLM.

01:02:36.949 --> 01:02:44.389
Yeah. Cool. Exactly. So then this is V2, then

01:02:44.389 --> 01:02:46.829
half a year later again. So every half a year

01:02:46.829 --> 01:02:50.940
they come up with another version. is the one

01:02:50.940 --> 01:02:54.280
that everyone kind of caught the world's attention,

01:02:54.599 --> 01:02:56.980
which is two things, actually, a combination.

01:02:57.260 --> 01:03:01.360
So they published V3, which is an extension of

01:03:01.360 --> 01:03:07.360
V2, which is their general AI model, in December

01:03:07.360 --> 01:03:11.840
of 2024. So about seven months later than V2.

01:03:12.320 --> 01:03:15.940
In January... they published, again, another

01:03:15.940 --> 01:03:19.260
model called R1, which is the reasoning model.

01:03:20.400 --> 01:03:23.980
Both together is what really made the news, but

01:03:23.980 --> 01:03:26.619
what I think that actually caught everyone's

01:03:26.619 --> 01:03:31.159
attention is because both Vivo and V2 are not

01:03:31.159 --> 01:03:34.079
mass releases. They were not on the app stores.

01:03:34.239 --> 01:03:37.820
There was no website that you can access to ask

01:03:37.820 --> 01:03:43.110
the LLM questions. True. Right? Users that are

01:03:43.110 --> 01:03:45.989
not in the industry that doesn't know how to

01:03:45.989 --> 01:03:51.150
use GitHub to clone the model itself don't have

01:03:51.150 --> 01:03:55.389
access to it until December. Right. And that's

01:03:55.389 --> 01:03:59.130
why it blew up. I think most people talk about

01:03:59.130 --> 01:04:04.769
the shock factor as, oh, it appeared as the top

01:04:04.769 --> 01:04:07.829
ranking app in terms of download in the App Store.

01:04:08.010 --> 01:04:10.719
And that's how we know about it. Right. Right.

01:04:10.860 --> 01:04:13.780
And that's how I knew about it. And that's, that's,

01:04:13.780 --> 01:04:15.880
that's what we, the, what we experienced during

01:04:15.880 --> 01:04:19.519
the Chinese New Year time. I know. The news published,

01:04:19.619 --> 01:04:23.099
it was January 20th or something or 25th. Yeah.

01:04:23.199 --> 01:04:25.360
And it came all of a sudden and the next day

01:04:25.360 --> 01:04:27.440
everyone is talking about it and it show up in

01:04:27.440 --> 01:04:29.920
the app store. It's literally when everyone is

01:04:29.920 --> 01:04:32.099
on vacation too. Yeah. Yeah. Like you don't do

01:04:32.099 --> 01:04:34.300
that to people who are on vacation. Like everyone

01:04:34.300 --> 01:04:37.599
then has to go, like if people who are, you know,

01:04:37.599 --> 01:04:42.250
trading in, in the u .s market they have to like

01:04:42.250 --> 01:04:44.289
the chinese trader who are trading in u .s market

01:04:44.289 --> 01:04:46.510
they woke up that day thinking that they're on

01:04:46.510 --> 01:04:49.269
chinese new year break oh man have to go back

01:04:49.269 --> 01:04:52.230
have to go back to work and sort out the nvidia

01:04:52.230 --> 01:04:56.909
chip exactly exactly so that that was one other

01:04:56.909 --> 01:05:00.719
shocker um obviously There's the price factor,

01:05:00.820 --> 01:05:03.920
right? So I think you might remember the $6 million

01:05:03.920 --> 01:05:08.159
training cost is what was published everywhere.

01:05:08.500 --> 01:05:11.380
How did the $6 million came about? Because it's

01:05:11.380 --> 01:05:13.380
an open source model, they actually publish a

01:05:13.380 --> 01:05:18.760
lot of stuff. So along with V3, they published

01:05:18.760 --> 01:05:21.900
an article in their blog saying a cluster of

01:05:21.900 --> 01:05:26.440
about 2000 Nvidia H800. So H800 is the China

01:05:26.440 --> 01:05:31.159
version of 8800 after the export bans. And it

01:05:31.159 --> 01:05:37.840
used a total of 2 .788 million GPU hours. And

01:05:37.840 --> 01:05:41.960
they estimated if you rent this GPU in the market,

01:05:41.960 --> 01:05:46.719
it's about $2. So 2 .788 million times $2 per

01:05:46.719 --> 01:05:54.079
GPU hour, that's 5 .6 million USD. That's the

01:05:54.079 --> 01:05:56.019
number that was quoted. That was the number that

01:05:56.019 --> 01:06:00.070
every single article. ended up publishing. Everyone

01:06:00.070 --> 01:06:02.829
thought they trained a model for $6 million.

01:06:03.269 --> 01:06:07.809
And I think some article in the immediate aftermath,

01:06:07.829 --> 01:06:09.989
the first two days, people were comparing this

01:06:09.989 --> 01:06:12.650
number to the billions of dollars that OpenAI

01:06:12.650 --> 01:06:15.619
spent. And I was like... What are we doing, USA?

01:06:17.280 --> 01:06:20.039
The tens of years that Liang Wanfeng spent in

01:06:20.039 --> 01:06:22.440
hoarding the cards and doing the trainings himself

01:06:22.440 --> 01:06:25.360
and, you know, in the past 10 years and hiring

01:06:25.360 --> 01:06:27.679
the talents and hiring the people, none of that

01:06:27.679 --> 01:06:29.420
was counted. Yeah, none of that was counted.

01:06:29.739 --> 01:06:32.920
That's not fair. Yeah, and I think a day or two,

01:06:32.960 --> 01:06:35.559
I mean, immediately people start to kind of realize

01:06:35.559 --> 01:06:39.449
and publish articles, but... it was too late

01:06:39.449 --> 01:06:42.570
for the stock market. Just poor financial reporting.

01:06:42.809 --> 01:06:45.869
Yeah. That caused the stock market to react.

01:06:46.210 --> 01:06:49.369
So the stock market reacted. Nvidia saw its worst.

01:06:49.429 --> 01:06:52.469
I think the stock market overall saw its single

01:06:52.469 --> 01:06:56.369
stock worst return over one day, which is 17

01:06:56.369 --> 01:07:00.949
% down. And then obviously people started to

01:07:00.949 --> 01:07:03.449
realize, or they realized, but people started

01:07:03.449 --> 01:07:06.369
to publish articles saying that, you know, they...

01:07:06.760 --> 01:07:10.380
this is just the last run of that model, right?

01:07:10.500 --> 01:07:16.079
There's so many other costs, like human resources,

01:07:16.199 --> 01:07:18.699
for example, research and development, everything

01:07:18.699 --> 01:07:22.199
that went on before counts. Yeah. So if you add

01:07:22.199 --> 01:07:25.780
that up, I think there are some numbers that

01:07:25.780 --> 01:07:29.739
puts this whole project between 500 million to

01:07:29.739 --> 01:07:33.699
1 .6 billion, which... Overall, it's still cheaper

01:07:33.699 --> 01:07:37.039
than U .S. companies working on the same things.

01:07:38.500 --> 01:07:42.059
But the shock factor was late. No one expected

01:07:42.059 --> 01:07:44.980
a Chinese company that is not ByteDance, not

01:07:44.980 --> 01:07:49.300
Alibaba, not Tencent to come up with a model

01:07:49.300 --> 01:07:53.699
that on performance base matches that of ChatGPT's

01:07:53.699 --> 01:07:59.820
latest, probably better than Lama, and is open

01:07:59.820 --> 01:08:05.760
source. And it's free. And it's free. It's not

01:08:05.760 --> 01:08:08.460
free for API, but it's free for everyone to use.

01:08:08.679 --> 01:08:11.500
Even the reasoning model is free for everyone

01:08:11.500 --> 01:08:14.980
to use. Stanford published an article that says,

01:08:15.039 --> 01:08:20.479
we thought China was way behind on all the generative

01:08:20.479 --> 01:08:24.960
AI models that are coming up. Effectively, China

01:08:24.960 --> 01:08:28.399
is behind by three to four months with a lot

01:08:28.399 --> 01:08:33.100
less chips. Be alert. Yeah, all of these developments

01:08:33.100 --> 01:08:36.020
are because of... Liao Wenfeng couldn't find

01:08:36.020 --> 01:08:39.319
better resources. There's no chips, and... If

01:08:39.319 --> 01:08:41.819
you don't want American educated PhDs, we'll

01:08:41.819 --> 01:08:44.079
just find homegrown ones. Yeah, that's actually

01:08:44.079 --> 01:08:46.859
a really good point. Let's touch on that. He

01:08:46.859 --> 01:08:49.420
doesn't want... No, we're not saying that. He

01:08:49.420 --> 01:08:53.279
doesn't want American PhDs. Okay, sorry. I said

01:08:53.279 --> 01:08:55.640
that. He never said that. I don't think he ever

01:08:55.640 --> 01:08:57.699
disclosed that we don't want international talents.

01:08:57.880 --> 01:09:00.300
I think it's because he wanted, but he couldn't

01:09:00.300 --> 01:09:03.000
find. Or there's just not as many in the market.

01:09:03.000 --> 01:09:05.399
He's not by dance, right? So it's hard to find.

01:09:05.439 --> 01:09:07.680
Hard to attract talents. Yeah. It's still a small

01:09:07.680 --> 01:09:10.539
lab. Yeah. and he has a lot of options because

01:09:10.539 --> 01:09:13.760
it's a small lab like they're quite selective

01:09:13.760 --> 01:09:18.119
about talent there's someone uh complaining online

01:09:18.119 --> 01:09:21.539
about like i'm a huawei engineer top of the top

01:09:21.539 --> 01:09:24.800
of my school eight years of working experience

01:09:24.800 --> 01:09:28.020
deep seek never gave me an interview like hey

01:09:28.020 --> 01:09:32.729
um so there is a deep seek hiring philosophy,

01:09:33.189 --> 01:09:39.810
which is super interesting, which is, and not

01:09:39.810 --> 01:09:41.869
speaking for them that but this is from coming

01:09:41.869 --> 01:09:44.189
from stuff we gather from the headhunters that

01:09:44.189 --> 01:09:47.550
are talking about their hiring strategy, which

01:09:47.550 --> 01:09:52.590
is, they tend to hire folks that are new graduates,

01:09:52.670 --> 01:09:55.829
or folks that are still in the last two, three

01:09:55.829 --> 01:10:00.079
years of their PhD program, who studied domestically

01:10:00.079 --> 01:10:05.560
in china which is very well i guess counterintuitive

01:10:05.560 --> 01:10:08.539
because most of the other companies are actually

01:10:08.539 --> 01:10:14.520
trying to find talent from open ai from uh anthropic

01:10:14.520 --> 01:10:18.359
from from google and meta right and try to get

01:10:18.359 --> 01:10:22.699
them to come back to china the oversee students

01:10:22.699 --> 01:10:29.130
uh and to help the Chinese AI industry grow and

01:10:29.130 --> 01:10:31.489
offer them probably two times of their salary

01:10:31.489 --> 01:10:34.649
that they're making in the US. And they all went

01:10:34.649 --> 01:10:39.029
completely a different approach. And it worked.

01:10:39.770 --> 01:10:43.909
So it's kind of ageism, but they actually refuse

01:10:43.909 --> 01:10:46.670
to hire people that has more than five years

01:10:46.670 --> 01:10:49.729
or eight years of experience because the philosophy

01:10:49.729 --> 01:10:54.539
is that I'd rather hire someone who... are more

01:10:54.539 --> 01:10:57.479
creative or who doesn't have necessary experience

01:10:57.479 --> 01:10:59.680
but are smart and think about you know how to

01:10:59.680 --> 01:11:03.359
solve problems rather than someone who is who's

01:11:03.359 --> 01:11:05.399
good at what they're doing but have a specific

01:11:05.399 --> 01:11:07.859
way or is used to doing things a certain way

01:11:07.859 --> 01:11:11.199
because this person is experienced so it's very

01:11:11.199 --> 01:11:14.560
interesting yeah yeah willing to hire new and

01:11:14.560 --> 01:11:17.060
innovative people without any experience yeah

01:11:17.060 --> 01:11:21.939
yeah And so they currently have, I think, about

01:11:21.939 --> 01:11:26.279
150 or so employees. And I think a majority of

01:11:26.279 --> 01:11:29.539
their employees are 20 -something -year -olds,

01:11:29.619 --> 01:11:33.020
right? So it's very interesting. And I look at

01:11:33.020 --> 01:11:36.579
their top maybe 20 or so. Most of them are from

01:11:36.579 --> 01:11:40.520
Beida or his Zhejiang University. So Peking University,

01:11:40.840 --> 01:11:43.500
Beida, or Zhejiang University. And a lot of them

01:11:43.500 --> 01:11:48.369
has PhD degrees. and they had their PhD degrees

01:11:48.369 --> 01:11:51.470
domestically. Yeah. Yeah. Gives a lot of hope

01:11:51.470 --> 01:11:54.810
for the LLM industry within China. But I think,

01:11:54.810 --> 01:11:57.869
you know, before recording, we both shared this

01:11:57.869 --> 01:12:00.869
very critical view about DeepSeek, right? Obviously,

01:12:01.149 --> 01:12:03.609
you know, great company, made a hit, has been

01:12:03.609 --> 01:12:05.710
working on this for a very long time, have their

01:12:05.710 --> 01:12:07.569
way of accumulating wealth, have their way of

01:12:07.569 --> 01:12:10.289
hiring talents. But the question that we both

01:12:10.289 --> 01:12:13.380
had was how long they can... or how long they

01:12:13.380 --> 01:12:15.720
can remain in this industry. What are your thoughts

01:12:15.720 --> 01:12:18.479
on that? This race will not end in the next two

01:12:18.479 --> 01:12:23.539
years. It will be in a while. And obviously,

01:12:23.640 --> 01:12:26.760
he can still have a group of talents and he can

01:12:26.760 --> 01:12:30.460
still be... Huanfang HF Fund, as of this point,

01:12:30.500 --> 01:12:32.199
is still up and running, but it's just much smaller

01:12:32.199 --> 01:12:35.539
in size. What do we think of the future of DeepSeek?

01:12:35.579 --> 01:12:38.359
How is that going to go? Personally, I'm very

01:12:38.359 --> 01:12:41.079
concerned. But there's a few factors. Let's run

01:12:41.079 --> 01:12:43.560
through some of them. So first of all, there's

01:12:43.560 --> 01:12:47.420
a whole cost issue, right? He managed to keep

01:12:47.420 --> 01:12:52.699
the cost down. And he cut down the API cost so

01:12:52.699 --> 01:12:56.239
much that ByteDance, Alibaba, OpenAI has to respond.

01:12:57.220 --> 01:13:00.619
And it's free for everyone. So I'm not sure if

01:13:00.619 --> 01:13:02.439
they're actually making money. I don't think

01:13:02.439 --> 01:13:06.130
they are starting to. making perhaps a little

01:13:06.130 --> 01:13:09.510
but right yeah they're not on the you know the

01:13:09.510 --> 01:13:11.689
extremely aggressive end of commercialization

01:13:11.689 --> 01:13:15.130
yet yeah and they're definitely on the cost reduction

01:13:15.130 --> 01:13:18.590
front like they could have priced their api a

01:13:18.590 --> 01:13:20.449
little bit higher but they chose not to because

01:13:20.449 --> 01:13:24.529
the i think the end goal is agi and so this is

01:13:24.529 --> 01:13:29.029
kind of all the path toward it but he doesn't

01:13:29.029 --> 01:13:32.270
have funding he uses his own funding or he uses

01:13:32.270 --> 01:13:37.050
funding from from uh hf fund an article i read

01:13:37.050 --> 01:13:40.090
that is that he actually at the beginning maybe

01:13:40.090 --> 01:13:43.289
2023 2022 when he wanted to establish deep seek

01:13:44.220 --> 01:13:47.359
He did go talk to investors to think about, you

01:13:47.359 --> 01:13:50.060
know, maybe I shouldn't fund 100 % myself. It

01:13:50.060 --> 01:13:53.020
should be, you know, us with the investors. We

01:13:53.020 --> 01:13:55.079
can maybe tap into their resource, et cetera,

01:13:55.119 --> 01:13:58.119
et cetera, which is most startups, AI startup

01:13:58.119 --> 01:14:02.859
path in China. DeepSeek didn't end up doing it

01:14:02.859 --> 01:14:06.800
because with money comes with obligations. And

01:14:06.800 --> 01:14:09.359
usually these obligations means commercialization

01:14:09.359 --> 01:14:12.800
early on. And return. And return. And that's

01:14:12.800 --> 01:14:15.399
not the intention. for a company that is looking

01:14:15.399 --> 01:14:19.560
at AGI. That's not what OpenAI has achieved.

01:14:19.800 --> 01:14:25.340
That's not the fundamental goal of Liang Wenfeng

01:14:25.340 --> 01:14:27.060
at this point. Liang Wenfeng's company feels

01:14:27.060 --> 01:14:30.439
very much like a university lab. Exactly. It

01:14:30.439 --> 01:14:32.420
doesn't feel like that he wants to be delivering

01:14:32.420 --> 01:14:34.619
anything commercial just yet. He's just hoarding

01:14:34.619 --> 01:14:36.579
a lot of really talented students and say, do

01:14:36.579 --> 01:14:38.539
whatever you want. Yeah, it's a research lab.

01:14:38.659 --> 01:14:42.000
It's a research lab funded by a really rich head

01:14:42.000 --> 01:14:44.960
fund investor. Interesting. Okay, I like that.

01:14:45.079 --> 01:14:46.800
Yeah. You just don't call it deep sea. You just

01:14:46.800 --> 01:14:48.420
call it a research lab. I think that's a more

01:14:48.420 --> 01:14:51.260
accurate depiction of what it is. Exactly. But

01:14:51.260 --> 01:14:57.319
the question is, this business model will only

01:14:57.319 --> 01:15:02.109
work when... You have enough belief in the AGI,

01:15:02.210 --> 01:15:09.189
in achieving AGI through LLMs. And you're going

01:15:09.189 --> 01:15:13.529
to continue to make money on your fund to supply

01:15:13.529 --> 01:15:16.329
this or you find another investor. But like you

01:15:16.329 --> 01:15:19.069
said, with investor come with obligations and

01:15:19.069 --> 01:15:24.369
can you take on these obligations without disrupting

01:15:24.369 --> 01:15:31.640
your research idealism? So, I mean, research

01:15:31.640 --> 01:15:36.579
idealism meaning that the employees of DeepSeek

01:15:36.579 --> 01:15:40.439
apparently can run any model without pre -approval.

01:15:40.479 --> 01:15:43.100
They're just like, I have this idea. Full freedom.

01:15:43.279 --> 01:15:46.460
Yeah, we have enough chips back then. Try it

01:15:46.460 --> 01:15:49.760
out, you know. Can this continue, right? Especially

01:15:49.760 --> 01:15:54.739
yesterday, NVIDIA, again, their report, their

01:15:54.739 --> 01:15:58.210
Q1 report came out and said, hey we can't deliver

01:15:58.210 --> 01:16:03.609
h20 anymore to china because of a recent restriction

01:16:03.609 --> 01:16:08.229
the u .s government has put on us again so context

01:16:08.229 --> 01:16:13.810
china ordered about i think a million unit or

01:16:13.810 --> 01:16:17.550
so of h20 which is the chip that nvidia created

01:16:17.550 --> 01:16:20.489
especially for china to bypass the recent ban

01:16:20.489 --> 01:16:24.659
on gpu and china already ordered I think last

01:16:24.659 --> 01:16:27.000
year China already got one million or so, and

01:16:27.000 --> 01:16:30.000
then they ordered another million or so. NVIDIA

01:16:30.000 --> 01:16:32.899
already created some, right, in their inventory,

01:16:33.100 --> 01:16:34.880
about to ship to, some already shipped, some

01:16:34.880 --> 01:16:39.000
didn't ship. U .S. government, like April 9th

01:16:39.000 --> 01:16:42.539
said, hey, we're going to, you're going to need

01:16:42.539 --> 01:16:44.960
a license to ship that. We might not give you

01:16:44.960 --> 01:16:50.399
the license. I don't know, indefinitely. So one

01:16:50.399 --> 01:16:53.550
month later, NVIDIA's financial report. says,

01:16:53.789 --> 01:16:57.329
we got all of these inventory, we can't ship

01:16:57.329 --> 01:16:59.750
them, we're going to write them off. China is

01:16:59.750 --> 01:17:03.149
not getting those H20s. China is not getting

01:17:03.149 --> 01:17:07.449
those H20s, which means DeepSeek is not getting

01:17:07.449 --> 01:17:12.609
those H20s. So there's a lot of roadblock to

01:17:12.609 --> 01:17:18.390
the continual success of DeepSeek. But going

01:17:18.390 --> 01:17:23.020
back to January of this year, When they made

01:17:23.020 --> 01:17:26.840
the news, besides what we discussed, which is

01:17:26.840 --> 01:17:30.659
hitting the App Store top of the chart and just

01:17:30.659 --> 01:17:33.680
getting tractions, the $6 million costs that

01:17:33.680 --> 01:17:37.439
shocked everyone for training models, the other

01:17:37.439 --> 01:17:41.020
thing that really shocked the community that

01:17:41.020 --> 01:17:45.220
is actually working on generative AI models is

01:17:45.220 --> 01:17:51.489
R10 and also distal models. let me talk about

01:17:51.489 --> 01:17:56.409
r10 first so r10 is so the r1 is the reasoning

01:17:56.409 --> 01:17:59.189
model that if you go on deepsea .com you can

01:17:59.189 --> 01:18:01.869
use it if you click on the kind of i think it's

01:18:01.869 --> 01:18:04.609
called reasoning or whatever button right that's

01:18:04.609 --> 01:18:09.569
the equivalent of the o1 model published by chat

01:18:09.569 --> 01:18:16.869
gpt that one has a dash zero model which uses

01:18:17.800 --> 01:18:21.840
a technique that differs from every other model

01:18:21.840 --> 01:18:24.920
out there that's doing reasoning. And the best

01:18:24.920 --> 01:18:28.560
way to think about this is the R1 model or the

01:18:28.560 --> 01:18:32.479
most reasoning models out there are actually

01:18:32.479 --> 01:18:36.340
brilliant thinkers and mathematicians who have

01:18:36.340 --> 01:18:38.520
gone through formal education, who have been

01:18:38.520 --> 01:18:43.359
taught how to think through a heavily human -evolved

01:18:43.359 --> 01:18:46.880
post -training process where people like us,

01:18:47.449 --> 01:18:50.590
Tell the model, I think this is good, I think

01:18:50.590 --> 01:18:53.510
this is not. Human reinforcement. Exactly. Human

01:18:53.510 --> 01:18:57.470
feedback. Human feedback to make the model perform

01:18:57.470 --> 01:19:06.350
well. The R10 is unlike someone who went through

01:19:06.350 --> 01:19:09.569
a formal education, someone who taught themselves

01:19:09.569 --> 01:19:13.489
math and physics and everything else. There was

01:19:13.489 --> 01:19:18.569
no human involved. the model itself actually

01:19:18.569 --> 01:19:22.210
solves complex questions by asking themselves,

01:19:22.430 --> 01:19:25.170
is this the right way? How do I go about this?

01:19:25.369 --> 01:19:30.430
So you kind of reward yourself as the model every

01:19:30.430 --> 01:19:34.590
step of the way. So that was kind of a different

01:19:34.590 --> 01:19:37.189
way that was always discussed in the research

01:19:37.189 --> 01:19:41.470
community and DeepSeek proved that it works.

01:19:41.649 --> 01:19:45.819
In terms of... Well -performing Chinese models.

01:19:45.920 --> 01:19:48.119
Actually, Alibaba published one that's called

01:19:48.119 --> 01:19:51.119
Quinn. That was pretty good. But it didn't make

01:19:51.119 --> 01:19:55.140
the same splash as DeepSeek because it really

01:19:55.140 --> 01:19:59.279
was going along the same lines as most US models.

01:19:59.880 --> 01:20:02.819
DeepSeek actually really made a different innovation

01:20:02.819 --> 01:20:06.420
and folks in the US was looking at how this works

01:20:06.420 --> 01:20:09.520
and they're like, wow, hey, we should try that.

01:20:09.880 --> 01:20:14.300
So that was one thing that they did. kind of

01:20:14.300 --> 01:20:16.720
really shocked everyone. The other thing was,

01:20:16.859 --> 01:20:20.359
it's the first time that you can distill a reasoning

01:20:20.359 --> 01:20:25.880
model. And what distilling means is you have

01:20:25.880 --> 01:20:28.640
this big model that sits in the server somewhere

01:20:28.640 --> 01:20:31.119
that as a user, you can kind of, if you go on

01:20:31.119 --> 01:20:33.220
deepsea .com, if you use any of the apps, you

01:20:33.220 --> 01:20:37.720
can call on the API, right? That big model requires

01:20:38.829 --> 01:20:41.170
It's the 6 billion one. So it requires a lot

01:20:41.170 --> 01:20:47.050
of resources to train and to use. But as an individual

01:20:47.050 --> 01:20:49.770
hobbyist, you can kind of distill that model

01:20:49.770 --> 01:20:53.170
and make your own mini model using instead of

01:20:53.170 --> 01:20:56.710
600 billion parameters, you can train yours on

01:20:56.710 --> 01:20:59.930
7 billion parameters, for example. And you can

01:20:59.930 --> 01:21:04.010
recreate that model by kind of distilling it

01:21:04.010 --> 01:21:07.970
and putting it on your MacBook. Not our MacBook

01:21:07.970 --> 01:21:11.470
Air, but like a better MacBook. But essentially,

01:21:11.569 --> 01:21:14.310
it shows that how open source and how accessible

01:21:14.310 --> 01:21:16.909
it is. Exactly. So that more people can enter

01:21:16.909 --> 01:21:19.430
into this ecosystem. And then the more developer

01:21:19.430 --> 01:21:23.470
it is, the bigger the community is, the less

01:21:23.470 --> 01:21:25.689
that people will switch to other closed sourced

01:21:25.689 --> 01:21:29.449
LLMs. Right. Which is very smart. Yeah. And the

01:21:29.449 --> 01:21:31.689
fact that they went with open source actually.

01:21:33.420 --> 01:21:37.239
which kind of also shocks everyone um so so people

01:21:37.239 --> 01:21:41.840
can then use this and distill the reasoning model

01:21:41.840 --> 01:21:44.220
and the only other comparable reasoning model

01:21:44.220 --> 01:21:48.279
if you remember is no no not the open source

01:21:48.279 --> 01:21:51.779
one the only comparable one is 04. right which

01:21:51.779 --> 01:21:55.079
is closed source closed source is harder to distill

01:21:55.079 --> 01:21:56.899
the model because you don't have the parameters

01:21:56.899 --> 01:22:00.359
right so you actually have to create a billion

01:22:00.359 --> 01:22:03.609
times to really get a good distillation, which

01:22:03.609 --> 01:22:06.710
is very costly by any means. So there's no really

01:22:06.710 --> 01:22:12.109
good way. So DeepSeek R1 was the first distillable

01:22:12.109 --> 01:22:15.869
reasoning model for the community. And imagine

01:22:15.869 --> 01:22:19.210
all the copyists just going like, yeah, cheap

01:22:19.210 --> 01:22:24.149
model. Let's get it on my laptop. So it ended

01:22:24.149 --> 01:22:26.590
up, and I think the CEO of Hugging Face actually

01:22:26.590 --> 01:22:30.529
quoted this. After the DeepSeek models are once

01:22:30.529 --> 01:22:35.350
published, over 500 derivative models were created

01:22:35.350 --> 01:22:40.289
on Hugging Face. And these 500 or so models,

01:22:40.510 --> 01:22:42.789
and this was two months ago, so the number probably

01:22:42.789 --> 01:22:46.970
went up, racked up a total of 2 .5 million downloads.

01:22:47.489 --> 01:22:50.569
Wow. Yeah. This is the developer community. This

01:22:50.569 --> 01:22:52.609
is the developer. 2 .5 million for the developer

01:22:52.609 --> 01:22:57.329
community. Right. Wow. Yeah. And guess who? open

01:22:57.329 --> 01:23:01.750
source deep six model who is the one and only

01:23:01.750 --> 01:23:07.090
so uh perplexity oh really yeah so perplexity

01:23:07.090 --> 01:23:12.010
actually uh used open so if you go to perplexity

01:23:12.010 --> 01:23:14.810
you can actually find deep seek oh yeah okay

01:23:15.310 --> 01:23:18.250
And they actually did it not for completely free

01:23:18.250 --> 01:23:20.149
because they have to train it, but it's pretty

01:23:20.149 --> 01:23:23.649
free and DeepSeat didn't get a cent. And that's

01:23:23.649 --> 01:23:26.890
the beauty of open source. So what Perplexi did,

01:23:27.130 --> 01:23:31.640
which is super smart, is that they... I guess

01:23:31.640 --> 01:23:35.380
this is called retrained. So they open sourced

01:23:35.380 --> 01:23:38.640
a version of the DeepSeek R1 model that's been

01:23:38.640 --> 01:23:41.140
post -trained to provide unbiased, accurate,

01:23:41.279 --> 01:23:45.220
and factual information. And by unbiased, accurate,

01:23:45.359 --> 01:23:49.920
and factual, it means they've created prompts

01:23:49.920 --> 01:23:54.800
that would trigger censored responses. Censored

01:23:54.800 --> 01:24:01.210
meaning party censored. And they generate a response

01:24:01.210 --> 01:24:07.189
to these prompts to make the model answer these

01:24:07.189 --> 01:24:12.210
questions as if they were not censored. Interesting.

01:24:12.289 --> 01:24:16.609
So they created an uncensored version of DeepSeek

01:24:16.609 --> 01:24:21.050
hosted in the U .S. on Perplexity servers and

01:24:21.050 --> 01:24:26.170
marketed that successfully. Oh, that's very smart.

01:24:26.409 --> 01:24:29.720
It's a great way. And all they have to put up

01:24:29.720 --> 01:24:33.340
in terms of cost is the post -training and server.

01:24:33.819 --> 01:24:39.600
All in all, I do think this is a very successful

01:24:39.600 --> 01:24:43.359
attempt for a Chinese LLN company. Definitely.

01:24:43.380 --> 01:24:47.220
I like that Liang Wanfeng is running the company

01:24:47.220 --> 01:24:50.840
like a college lab. It's giving the freedom to

01:24:50.840 --> 01:24:55.539
the researchers. And I like that he open sourced.

01:24:56.079 --> 01:24:58.739
He even embraced with an open mind, I think.

01:24:58.760 --> 01:25:01.560
He's not making money off of all the copies that's

01:25:01.560 --> 01:25:04.300
being made, but he is letting people to use it

01:25:04.300 --> 01:25:07.079
and doing their own creativities. So I think

01:25:07.079 --> 01:25:09.720
that takes a lot of courage. Yeah. But I also

01:25:09.720 --> 01:25:12.420
think this is a bit calculated. For sure. Because

01:25:12.420 --> 01:25:16.140
you are not the incumbent. Definitely. You're

01:25:16.140 --> 01:25:19.109
trying to... become the incumbent and the best

01:25:19.109 --> 01:25:22.069
way to do it is to do it with a community and

01:25:22.069 --> 01:25:25.310
to build a community you open source right yeah

01:25:25.310 --> 01:25:29.850
so it's actually quite smart yeah but i guess

01:25:29.850 --> 01:25:33.310
the the question is um deep seek i think is very

01:25:33.310 --> 01:25:37.949
far far away from um you know being a public

01:25:37.949 --> 01:25:39.729
company i think it still takes a lot of years

01:25:39.729 --> 01:25:41.329
and i don't even know whether that's the way

01:25:41.329 --> 01:25:46.489
for them to go but with the information that

01:25:46.489 --> 01:25:49.869
we were able to find online we can only paint

01:25:49.869 --> 01:25:54.090
you a very faint picture of what it is and what

01:25:54.090 --> 01:25:57.069
it is not. And I would highly recommend that

01:25:57.069 --> 01:26:00.069
we just wait it out and see what could be the

01:26:00.069 --> 01:26:02.289
future hold for them. I really look forward to

01:26:02.289 --> 01:26:04.689
what they will be doing. But I do think this

01:26:04.689 --> 01:26:06.930
is a highly competitive arena. And what Liang

01:26:06.930 --> 01:26:08.970
Feng is doing, he's definitely trying his best

01:26:08.970 --> 01:26:12.010
to, you know, finding talents and, you know,

01:26:12.010 --> 01:26:15.569
bring up the most interesting, I guess, algorithms

01:26:15.569 --> 01:26:18.109
or the models that he can create. I think it

01:26:18.109 --> 01:26:20.550
did a really good job to stir up the market.

01:26:21.539 --> 01:26:27.260
And it's sending positive momentum to the research

01:26:27.260 --> 01:26:31.659
industry, to the research community. And I hope

01:26:31.659 --> 01:26:35.140
that despite all the GPU bans, they can continue

01:26:35.140 --> 01:26:38.319
or other companies in China can pop up and challenge

01:26:38.319 --> 01:26:42.520
the incumbents to create more competition, to

01:26:42.520 --> 01:26:45.899
create more innovation at this space. Thanks

01:26:45.899 --> 01:26:49.819
to DeepSeek, actually OpenAI's API price dropped

01:26:49.819 --> 01:26:53.310
a lot. right and thanks to deep seek now we have

01:26:53.310 --> 01:26:56.609
a lot more options when it comes to chinese llms

01:26:56.609 --> 01:26:59.630
so i i don't know about you but i like i know

01:26:59.630 --> 01:27:03.029
dc has his performance limitations but i I tend

01:27:03.029 --> 01:27:05.970
to ask a lot of Chinese questions to DeepSeek,

01:27:05.989 --> 01:27:09.170
and I find that you get better answers than asking,

01:27:09.210 --> 01:27:13.609
say, chat GPT. Right. Yeah. Well, I take all

01:27:13.609 --> 01:27:16.470
of my AI tools with a grain of salt. DeepSeek

01:27:16.470 --> 01:27:18.829
does well in some things. It can do some very

01:27:18.829 --> 01:27:21.430
basic secretary work for me. Well, maybe I'm

01:27:21.430 --> 01:27:23.130
just asking too simple of a question, but I think

01:27:23.130 --> 01:27:25.390
it still hallucinates a lot at this point. Yes.

01:27:26.460 --> 01:27:29.079
But I would hope that it will continue to improve

01:27:29.079 --> 01:27:30.760
in the quality of answers that it can provide.

01:27:31.000 --> 01:27:34.119
And I really do enjoy the thinking process. I

01:27:34.119 --> 01:27:37.819
always open that thinking bar. It's like it responses

01:27:37.819 --> 01:27:42.000
quite well to different questions and different

01:27:42.000 --> 01:27:47.399
contexts. So it kind of shocks me sometimes that

01:27:47.399 --> 01:27:49.979
the type of responses the thinking process would

01:27:49.979 --> 01:27:52.899
label. So today I asked a question and somehow

01:27:52.899 --> 01:27:56.130
from my tone. I think the thinking process that,

01:27:56.210 --> 01:27:59.489
you know, this question seems quite urgent. What

01:27:59.489 --> 01:28:01.949
were you asking? I mean, it was just a really

01:28:01.949 --> 01:28:04.770
like brief. I didn't type as much. It misunderstood

01:28:04.770 --> 01:28:07.810
me. But the fact that, you know, because my question

01:28:07.810 --> 01:28:10.729
was very brief, it somehow concluded that maybe

01:28:10.729 --> 01:28:13.869
I have a sense of urgency. It's quite interesting,

01:28:13.909 --> 01:28:18.390
right? It's so akin to a human that it's... a

01:28:18.390 --> 01:28:21.369
little bit scary at this point it's trying to

01:28:21.369 --> 01:28:23.489
understand like what what else you're trying

01:28:23.489 --> 01:28:27.310
to convey to the machine yeah it's been very

01:28:27.310 --> 01:28:29.829
fun researching for deep seek super fun yeah

01:28:29.829 --> 01:28:33.529
super frustrating but because it changes all

01:28:33.529 --> 01:28:36.029
the time it's like two people that have no quant

01:28:36.029 --> 01:28:40.350
trading background no limited llm background

01:28:40.350 --> 01:28:43.239
we're llm users doesn't mean we're Developers,

01:28:43.260 --> 01:28:47.420
right? I think we have to thank the folks that

01:28:47.420 --> 01:28:49.960
are behind our research that helped us with fact

01:28:49.960 --> 01:28:52.859
-checking. Well, thank you for all the QuantTrader

01:28:52.859 --> 01:28:55.760
friends that I've reached out for your continuous

01:28:55.760 --> 01:28:57.699
support and not hating me as a friend who is

01:28:57.699 --> 01:29:00.640
constantly asking financial questions. And thank

01:29:00.640 --> 01:29:03.560
you for all the friends that I reached out that

01:29:03.560 --> 01:29:09.039
was helping me. answering my questions with a

01:29:09.039 --> 01:29:12.180
12 -hour time zone difference. And if you have

01:29:12.180 --> 01:29:15.739
any questions or comments, please do leave us

01:29:15.739 --> 01:29:19.380
a message. And if there's any other rising companies

01:29:19.380 --> 01:29:21.560
that you're interested in hearing us talk about,

01:29:21.699 --> 01:29:23.720
please do leave a comment as well. We'll see

01:29:23.720 --> 01:29:26.060
you in the next episode. Bye. Bye.