What I think is most interesting about it, though, is it upends the business model that Silicon Valley had been trying to win the AI race.
One thing that is interesting to watch is what happens with NVIDIA with its web 90 % margins.
This could be very dangerous for NVIDIA in the long run.
The more possible cases will be that China and the US are leading in their own perspectives at the tech industry, and each has their own unique advantages.
The Chat Lounge The Chat Lounge The Chat Lounge unpacks views and opinions on hot issues in a more casual way.
Welcome to The Chat Lounge.
I'm Thuyen. Joining our discussion on the open source AI bombshell Deep Seek are Professor Tobey Walsh, Chief Scientist from the AI Institute, University of New South Wales Australia, Andy Mock, a senior research fellow at the Center for China and Globalization, and Shen Jiahe, the chief investment officer of Beijing -based Nova Market Technologies.
It's a great pleasure to have you all on the chat.
Gentlemen, welcome.
Wall Street has been stunned.
A Chinese -quantitative hedge fund called high flyer fueled a massive sell -off of tech stocks, not with quant trading but with the release of the latest version of its large language model, DeepSeq R1.
It performs on par with open AI's O1 model which is its most advanced model across complex reasoning benchmarks on reduced chip levels at much lower costs.
Well let's first take a close look at the Deepwater Bomb dropped by High Flyer, the first question goes to you all.
So what's your evaluation of DeepSeek's performance especially compared to OpenAI's chat GPT?
I checked the comments on Chinese social media, some say it's fabulous and highly capable of even composing articles in classical Chinese, but some feel disappointed saying it didn't work that well as GPT does.
So what's your assessment then?
Professor Walsh, it's your first time on the show.
Shall we start with you, please.
Yeah. Well, I think hats off to the young and small team at DeepSeek for pulling this off.
The performance is very comparable to the best of the West.
Some ways it may be some places, it may be worse.
But the fact that they've been able to do it demonstrates the view that various people had.
That China was behind 6 months or more behind the US is actually false.
That China is right up there with the best tech companies in the US.
And then I think what's most impressive about it, it's not the performance, because performance, as I say, is very comparable to existing models like the best ones from open AI, but they did it with way less compute.
It's claimed that they used about $6 million to train it.
That's a 20th of what at least a 20th of what was used to train those comparable Western models.
So that really does open up huge great opportunities.
And also was the reason there was that that vast correctional tech stocks that said people realize that the West wasn't going to get to win.
And maybe there was a better, more efficient model that was available elsewhere.
Yeah. You used it yourself right?
What kind of questions did you ask?
Can you share with us?
It took a while to use because it's become the most popular app on the app store, Number One Download overtook Chad GPT.
So there was a bit of a backlog to get yourself registered.
It took about a day to actually get access.
But when I did get access, I tried to doubt on all the little trick reasoning questions that normally trip up smaller models.
I wasn't actually able to trip it up that easily.
So I, I was actually quite impressed by the performance.
It's also worth out, you know, two other things to be impressed by its performance.
One is that it's open.
You can download it and run it on your own computer completely free and so they've been able to to do it, you know, in such a compact way.
So I think it's, as I said, hats off to the team.
Yeah. The reasoning process is quite interesting.
And what about Andy?
Well, Thuyen, I have to say that on one hand this is a surprise but in a more important way, this is no surprise at all.
Right. You know, of course there's been a lot of rhetoric about China being behind on AI.
So, from this perspective, I think many people were stunned, especially in the Western world.
But when we look at the broader picture of China's interaction with the West, since 1949 on the establishment of new China.
This is just the latest incident.
So let's go back to Tian Justan, right?
With the China developing capabilities in the atomic bomb, hydrogen bomb launching a satellite.
Then again, if we go back to that time period the United States was aghast that poor backwards China is able to do this.
If we fast forward more recently into Hawaii, again, that Huawei not only survived, but thrived in against this unprecedented attack by a nation state against a private company.
And similarly, we see deep seek here, doing something that no one thought was possible at a more macro conceptual level.
And I think we ought not to be too surprised.
Now, let me talk a little bit about the AI part of this, which also, I think suggests that we perhaps should not be too surprised by this.
So senior executive scientists in this field, from Sam Altman to many others, have said that if you look at the three components of artificial intelligence breakthroughs, hardware, algorithms, and data, it's algorithms that lead to the biggest breakthroughs, not necessarily hardware.
And we only need to look at what enabled these breakthroughs in large language models generative of AI was an algorithmic breakthrough in a paper that was very cleverly titled, attention is all we need.
So when we look at the deep and broad pool of scientific talent in China, China graduates 2x the number of STEM PhDs that the United States does, the intensity, the ferocity of the competition in China, the rigor of its educational system.
Again I think we should not be too surprised if we see algorithmic breakthroughs coming from China.
But can you give us some example of yourself using it which you might think it's quite convenient or not so convenient?
Sure. So I am not a scientist in AI.
I would probably call myself a power user.
So I use pretty much every model available on the market.
and they all have their strengths, relative strengths, relative weaknesses.
I would say that deep seek is very much competitive again, only speaking from the perspective of the user.
So I have no complaints.
I think as an analyst, the thing that is most important is one that they were able to do this using only about, meaning what they claim is about 2000 H -800 in video chips, which are much less powerful.
of course a fraction of the number of ships that other major players have used.
And second, that they made it open source, meaning that anyone can use it, anyone can check their work.
The last thing I want to say that I think is actually the most important is more from a geopolitical perspective, not only that it caused the largest stock market decline in history, something like $600 billion, right, in India in a single day.
But what we're Witnessing is a sea change in attitudes towards China.
Mark Andreeza, who developed a first internet browser and now is one of the founding partners of Andreeza correlates, a top Silicon Valley VC fund, has been known to be a staunch China hawk.
Very few good things to say about China, and even when I've talked to him, I've been stunned by how reflexively negative he has been in the past about China.
China. And he recently posted on X that Deep Seek's model is an amazing contribution to mankind.
So let that sink in for a second.
Someone who is as well respected as vociferous a critic of China could say something so positive in an unqualified way about a Chinese company, a Chinese technology, I think is going to have long lasted geopolitical ramifications, just like we're seeing with TikTok and the reversal of the stance TikTok by President Trump and other senior European politicians.
Right, you've got a very comprehensive evaluation of that, but there was some twist as well because OpenAI is now saying that the DeepSeek used some technique called a distillation, which violated the contract.
And we'll come to that later on, but to...
what's your evaluation?
Have you used it and compared it with other large language models like a chat GPT?
I have been using this deep -sick recently.
I've been using it almost every day.
I found a very good way of using it is that if you're searching some very basic concept, it gets you a very perfect answer.
For example, I got one of my majors reading through the Chinese traditional literatures which were written about two or three thousand years ago, I read them about half an hour or one hour by every day.
It's a bit like westerners reading latings.
So in these kinds of traditional literatures there are many words that are not used in today's Chinese language, not at all, you just, I just don't know what they mean.
Previously, I have to check through the dictionary when I was a high school student.
Later on, we used the search engine for searching these Chinese traditional characters, but search engines really don't tell you the very they give the video to online dictionary, but you have to go through this dictionary by yourself again.
But nowadays, I mean, by using this deep seek, it's really amazing.
It just tells you exactly the answer of these characters.
So to me, it's more like serving a modern dictionary.
For example, I ask what's the English expression of the MAX seven, that's a large seven tech company.
So it tells you the very good answer.
So if you ask them the very basic answer, I mean, either it's open AI or type QPT or deep seek really give you very good answer.
So it's a very good assistant.
But on the other hand, the question is that if you ask them more complicated questions, it doesn't really have to be too complicated.
For example, I asked the deep seek this morning, what's the current valuation of the MAX7 companies?
And I thought maybe I don't have to bother myself to go through each one of their sticker in the financial terminal.
You know so I can just get the answer in one sheet, ask the dipstick what's evaluation of the MAX7 this morning?
And I specifically said I want the PE ratio and it gave me the answer that was the evaluation of these seven components last May.
So it's about eight months ago.
So it's exactly not the answer I wanted.
So this is a, you can look at this, is that if you ask the more basic questions about the basic concepts is they give you very exact answer.
But if you ask them a more complicated question, especially a question that's not deriving the answer simply from some literatures, but asking the computer to think, to really think, you can get very tricky answers.
This doesn't only happen with a user.
It actually gave me the answer that was the valuation of May 2024, and especially when you consider that this company is claiming that we are the leading AI, and this financial data is up to date every single day.
You can just read it really from any financial website.
It's just there. I spend about five seconds and find myself there.
I think this is my experience with using this kind of AI program, GIP -SIC, OpenAI, is that you can ask them very, very basic questions, and they give you very, exact answer.
It's very good answer.
You can rely on that.
But if you ask them a bit more complicated questions, then you don't get very good answer at all, and you can't trust them.
that so you say okay you give me also with 80 % accuracy a better spend some time myself to find 100 % accurate answer I can't rely on the 80 % answer if you see what I mean.
The Chat Lounge unpacks views and opinions on hot issues in a more casual way.
It maybe excels in a you know Chinese literature rather than you know finance but I got a classic question which is which number is bigger.
3 .11 or 3 .8. It gave very good reasoning and got the right answer.
And obviously, generally speaking, I think you all have a relatively high, I should say, evaluation of this new model.
But, Chagra, since you're in the investment field, you might be able to answer this.
Why is a hedge fund able to make such a big breakthrough in AI?
That's a big question for a lot of us right?
Yes that is really amazing.
It really tells you how strong the hedge funds really are.
I mean especially when we talk about the cons funds.
Because this high flyer fund has been very famous in China for their achievement in the quantitative investment field.
They've been losing quite a lot of computers.
They've been programming very good programs in these kinds of things.
So they have been working in this field for a long period of time.
Now they're turning their programming capability to the AI.
Well, I think that's probably the same thing for their investment, because you have to understand one thing.
It's that in the Quans' world, they're not just talking about understanding the price, the financial data, these kind of things.
What they are doing in some cases is that they're trying to understand what's happening on the internet.
With this understanding, which means they have to understand all the literatures and languages that people are using, traditionally, traditional investors, especially those investors that are following the daily movement of the market, what they are doing is that they reach through the newspaper in the morning and then they find out what's going on, and then they make their investment decisions in the morning meeting.
In the states' news world, there is one thing that is that you have news coming up 24 hours So there is a kind of a Quants program that tries to read and understand what's going on on the internet second to second.
And once they find there is something amazing, they automatically put out the investment decision, which is much faster than what human beings can do.
So this is what Quants are doing.
They try to understand the human language.
From this point of view, you can easily understand why Quants can't do a language model, because they have been working on this for a long period of time.
You used to understand the language to make money for themselves.
There are a lot of hedge funds in this world, especially, you know, in the States, a lot of them are very influential.
Professor Walsh, what's your interpretation of this phenomenon?
Well, in this case, you know, the hedge fund had to pivot because of the trading restrictions that were being put on in China because of the turbulence in the markets and they were looking for something else to turn all of their PhD brains and all of those GPUs that required to good at.
And they'd been able to do that successfully.
What I think is most interesting about it though is it upends the business model that Silicon Valley had been trying to win the AI race.
The business model was summed up by that slow scaling is all that you need and it'd been a race to see if it had the deepest pockets, who could throw the most money at the problem, throw more compute, more data, somewhat it was proving to be successful but But somewhat lazy strategy.
And what's interesting, what's exciting, is that DeepSeek demonstrates that you can actually get there in a much more efficient way.
And it's not going to stop there with DeepSeek.
The fact that we know you can do it means that lots of other people are going to be able to do this now.
It's a bit, you know, a bit like running the four minute mile.
Once the first human, Roger Bannister, ran the first four minute mile and proved it was physically possible, lots of people were then able to do it.
And the same will be true here.
There's already a project on Hugging Face.
Hugging Face is this collaborative platform where AI models are shared to rebuild DeepSeq, completely free, completely open.
Because although the weights of the model of DeepSeq are open, we don't actually know the data was trained on or some of the finer details.
It's not completely open to the world what DeepSeq have done.
So there is a project and I'm sure there'll succeed in a month or two, re -create it in an entirely open way.
And then so lots of people going to do it.
Not just in China, in Europe, possibly here, even in Australia.
Right. Something a little bit ironic is that it's developed by some hedge fund.
But, like Jacka just mentioned, ironically, it seems not so good at collecting financial information.
But hopefully they can solve that problem in the future.
But like Professor Walsh just mentioned, yeah, there are a lot of questions surrounding how DeepSeek collect data, right.
Open AI CEO Sam Woltman initially suggested DeepSeek copy the existing models, and President Trump's AI tar appointee David Sachs said there was what he called substantial evidence that DeepSeek distilled outputs from OpenAI models.
Professor Walsh, what's your evaluation of that claim then?
Oh, and I'm pretty sure that they did.
And there's a beautiful irony to this of course, because OpenAI and all the Western models that have been developed up to now have themselves been trained on data that was obtained, but not necessarily without consent.
And there's lots of lawsuits going down, class action suits going down in the US at the moment, and contesting that the New York Times is famously suing OpenAI for having trained on all their data.
There's some beautiful irony in the fact that OpenAI now has been the victim itself of data from its model being used in the training process for DeepSeek.
This idea of distillation, trying to distill the results of another model.
In fact, I mean, in a technical sense this is what's going to happen.
Every model in the future is going to be building on the outputs of other models.
Indeed, you know, a lot of social media these days is AI generated a lot of content.
So even if you're just scraping the web, not just accessing someone else's model, you're going to be training on the outputs of some other models.
So it's almost inevitable that you will be training on other people's output.
And indeed, that's, you know, partly how, how we stand on the shoulders of giants.
Models get better in part because they can train on the outputs of previous models.
Of And of course, there's question marks about the legality of that and the business sense of that, but from a technical sense, that's the way that it's always going to be going down from now, moving forwards.
Yes, you can use data collected by other models.
But according to OpenAI, such kind of techniques cannot be used for some rival products, right?
That's according to...
conditions of accessing open AI's models are that you can't use it to train a rival product.
You can understand from a business sense why they would want to do that.
The technical reason why people want to do this is because whilst there's lots of interesting data, facts and economic prices and things like this that you can train on from the web, there isn't so much of the explanations of how you reason and solve problems.
and the the latest models their latest capabilities are because that they're better at doing reasoning.
They break problems down and work out how to reason through a problem.
Now humans you know humans haven't written down how they reason through their problems on very much on the web.
So there isn't much of that training data but these models can help with that.
So this is why that they've you know probably used I mean other models as training data because they need to get that reasoning training data in some way.
Indeed. I actually asked a DeepSeek how it works and whether there are still outputs from open AI models.
The answer, no. And it says, according to DeepSeek's official documentation and public statements, DeepSeek R1 is independently developed using proprietary data sets and training pipelines.
It does not rely on distillation or outputs from OpenAI models, like a GPT -4.
It leverages self -generated synthetic data, human annotated data sets publicly available legally compliant of corpora and it also notes the model architecture and training methodologies are designed in in -house emphasizing compliance with data privacy and intellectual property regulations.
And it also outlines key differentiators from open AI models in terms of a focus on search integration, language, and domain specialization, and regulatory compliance.
I'm not saying that its answer is direct evidence that it hasn't done so.
But Professor Walsh, can you help us understand what the answer told us, like in layman terms.
Well, from a legal perspective, that's the answer you would want them to say, the problem is and this is a criticism made not just of deep seek, I can make it of all the tech companies of open AI, Google is that they now no one tells you exactly what they've trained on that it would be possible for them to actually itemize what was exactly the training data they claim that it was all publicly available, You know, not copyrighted data that they've trained on.
But you have to take their word for it.
And indeed the suspicions that, you know, the U .S.
take yards themselves, have violated people's intellectual property.
Indeed, you know, I take this rather personally because I've written a couple of books about AI.
I know AI trained on my books and they didn't get my consent and I've received no compensation for having used my books.
So, you know, there are a number of lawsuits from authors like myself.
protesting the way that they've used our data and you know that right now because of all these laws it's going down all of the people building models now are rather quiet on exactly what was the data they claimed that they're doing the right thing but it's not clear that you know if they were transparent then we wouldn't know for sure but since none of them in the US China are being fully transparent it's hard to be sure.
This has been the chat lounge When we come back, we'll find out more about the ripple effect Deep Seek may generate on the market Stay with us The strong wind was howling and whistling He was the first Chinese citizen to graduate from Yale University in the mid -19th century I was born on the 17th of November She had prominent features Three of us were old enough to lend a helping hand He navigated between two vastly different cultures and moved further to realize his dream and promote understanding between the people of China and the United States.
Xieming Xing was a native of Hanyang.
I realized no danger.
China is really awakening.
Come and join us in discovering the incredible journey of Yong Wang and his autobiography, My Life in China and America.
Check out the audible stories on radio .cgtn .com and all major podcast platforms, to search for the podcast, books and beyond and find my life in China and America.
Welcome back to the chat lounge.
We continue our chat on the Chinese developed open source AI model, DeepSeek.
Some netizens argue open AI has long banned Chinese mainland addresses and you know how strict the blockade is.
If some individuals use open AI through some tricky means, it's still possible.
But if it used to train another AI with huge amount of data and access, will OpenAI not be able to detect it, is the question.
And Andy, some are saying that such kind of accusation is triggered, actually, by what you just mentioned, the big slide on Wall street?
So what's your take on that, Ben?
Well, I suppose it depends what you mean.
So my understanding of distillation is not necessarily that a company is taking data because data is usually thought of as the input into a model, but my understanding of distillation is using an existing AI model and in this instance, a large language model to train your own model.
And, you know, we could maybe think of this as a little bit cheating, in that, you know, you're already building on something that exists.
But, you know, I think conceptually, you know, all scientific advances are based on previous advances.
So, you know, I don't know if this is a legal question that they violate if it's open AI in this instance, on any their intellectual property rights around their model, I'm not sure of.
But, conceptionally, you know, again, I think any kind of scientific advance is always based on a previous world.
I think also, in a way, it seems like a very sensible thing to do, to use something that already exists, right?
I mean, why reinvent the wheel?
Of course, again, subject to legal constraints, et cetera.
I do wanna say a couple things here about these models.
I do know that some of them do have the ability to do real -time search.
For instance, Plexidy is one model that is known to be optimized for research, and OpenAI and even Deep Seek does have a button with a little globe on it where you press to search the web.
And I think that is one of the functions of that is designed to give you real -time information, whether that's the latest stock prices, the particular status of an airplane flight, etc. So, I think there's very much a moving target in things are developing so quickly, not just at the technology layer, but also at the product feature layer as these companies better understand what consumers, businesses, governments want in terms of product features.
So for me this is what makes it a very very exciting time.
And just very briefly I want to share a quick story that to me kind of summarizes this and that there's a an apocryphal story in American astronaut and a Russian astronaut talking about how they write when they're in space.
So the American astronaut was very proud says you know we spent 20 million dollars developing a pen that can work in zero gravity because we don't need gravity to pull the ink, you know, to the niblet so we can write.
We spent 20 million dollars on this and he's really really proud.
And he says to the Russian and what did you guys do?
The Russian looks at him and says we used a pencil.
Right. So I think again we see that this is what makes deep -seek, I think, so important in that it is not just a breakthrough on its own, but I think it also represents the spirit of the Chinese way of doing things in a very pragmatic and efficient way.
I'll throw in just a very quick technical detail here.
Most of the modeling done by OpenAI, Meta, et cetera, use what's called FP32, meaning that they carry, they process 32 bits of information.
Now, of course, if you're going to use billions and billions of you know, calculations, you know, that matters a lot, whereas what DeepSeek did is they used 8 bits, and of course, you lose some precision, but they were clever about picking the points where you can still reduce the length of this data by a quarter or two only quarters without really losing that much precision.
And there's a number of these clever hacks or insights that they've done.
Again, that I think is just emblematic of whether you're looking at electric vehicles, whether you're looking at mobile phones, etc., that it's a very pragmatic Chinese approach that really is starting to take the world by storm.
That's also why the stunt at Wall Street, right, and we've seen, like you all mentioned, the Silicon Valley companies like Open AI, Meta, reportedly panicking.
Well, I think this has been so important, right, to you in that the paradigm in the US is that you have to spend hundreds of millions of dollars, maybe now billions of dollars, training these models, and only a few companies can afford them.
And I think what DC has shown us is that that may not be true.
And this has implications not just for the US stock market, but think about the global South now that this could potentially empower maybe millions of businesses in the global South in Africa and Southeast Asia in Latin America that could drive again accelerate the adoption of AI, right?
Because again, it's a it's a well known, I think it's called Gibbon's Paradox, right?
But when you make something more efficient, actually, rather than use less of it, we use much more.
So this in the long run may actually be positive for tech companies, but certainly in the short term, I think it's caused a lot of panic as we've seen with a drop in NVIDIA share price.
Right. And actually, it triggered some horror scene on Wall Street Monday.
And so, Ja -Hoo, you're in the financial business, so do you expect the prices of AI -related shares may continue to crash at such a scale like what we've witnessed on Wall Street?
Well, it's very difficult to say.
I mean, quite a lot of people have been arguing that the Nasdaq and these MEX 7 are trading a bit too expensive.
Basically speaking, many news media has been saying that the falling of large US tech companies is because of this dipstick that is produced by small Chinese company.
It looks like China has been hurting the US market, but it's not the case.
The dipstick is just one reason.
One of the major reasons behind the large correction of these tech companies is basically because they've trading at a pretty high valuation.
I mean it's not comparable to the valuation that we have seen in the 2000 dot com bubble.
It's still a pretty expensive valuation.
I mean, if you look at the valuation of the Max 7, they'll be trading at about 30 to 40 times PE ratio.
And if you look at the U .S.
financial history, I mean, this is a pretty expensive valuation that in many cases, they have been too expensive.
And in many cases, they've resulted in a lot of corrections in their past.
So these things have been seen.
So this valuation basically is what's causing the correction of the market.
So when we talk about the future, it's really a tricky question because it's not that expensive comparable to what has been going on in 2000 where we saw a 100 times or even 200 times the ratio.
But this times, what we have seen is that it's not that expensive valuation.
But it's not cheap at all, I mean Warren Buffett has been selling Apple, which is one of the cheapest companies in these, you know, MAX 7.
I mean, Apple is trading as only about 30 plus times P.
It is one of the most profitable companies, but Warren Buffett has been selling that.
And if you look at the past, I mean, Warren Buffett has always been correct with his decisions.
So if you follow the path of Warren, then you might think about lowering your position in these companies.
Right. Good advice.
Professor Walsh, Metachief's AI scientist, Yan Lekun, has said the biggest takeaway from Deep Seek's success, was not the heightened threat posed by a Chinese competition.
The value of keeping AI models open source, so anyone can benefit.
It seems you also support such an idea.
Do you expect DeepSeek may change Sam Altman's mind and make a chat GPT open source again?
Maybe in the future?
I'm not sure that that's going to be enough to change Sam Altman's mind.
There's a financial aspect to this, which is what's the best business model, and open source models are very attractive for improving the uptake of these models.
But then there are other considerations, considerations around responsibility.
How do you ensure that the technology's used in a responsible way?
It's much harder when the models are open source to be able to put in restrictions to prevent people from doing harmful things, or saying harmful things, or using them to find out the recipes for bioweapons or whatever it is that you want to do because people can quickly remove those restrictions.
But certainly, I think it's promising that open source is doing it.
I think the other thing is the fact you can do it with smaller models, not the lazy scaling, throwing money at the problem.
Just a historical note, people are really shocked and surprised by DeepSeek that they shouldn't have been because this isn't the first time that people have demonstrated that you can take a state -of -the -art model and do it with much less resources.
When 2020 to the end of 2022 it cost somewhere around $30 -odd million dollars to train.
Six months later, researchers at the University of California doing some of the tricks that we were just talking about.
Using lower precision floating -point operations, 4 -bit and 8 -bit arithmetic knots, 32 -bit, we're able to build a chat GPT clone for $300 of Amazon Web Service credits.
So the ultimate target is to run it on your smartphone, the ultimate target is to do it with 20 watts of power, just like the 20 watts of power your brain does.
So we're trying to do it with less, and that's always going to make it much easier for the technology to be deployed, is to make it much simpler.
So we shouldn't have been a surprise and the stock market shouldn't have been so troubled that people were starting to do models with way less compute, because we've been doing that for the last couple of years, any time my models come out, people have shown you can do it more efficiently, sometimes by orders of magnitude.
Like you mentioned, maybe open source has its negative impact to some extent.
But it seems DeepSeek, the founder of DeepSeek, Rianwenfong, or the company, decides to keep going on on death's a path, and who's also said that the problem they are facing has never been funding but the export controls on advanced ships.
So let's move on to the outlook or prospects of deep -seek.
Andy, what are the prospects for deep -seek against a backdrop of the US restrictions or even tighter restrictions on chip exports?
Well, I think it's natural that he would say this.
You know, of course, any company anywhere around the world would be better off if they had unfettered access to top of the line Nvidia chips.
And we have to recognize too, you know, for Chinese companies, it's politically motivated.
But these are just hard to buy.
So, you know, one thing that is, is interesting to watch is what happens with Nvidia with its what, 90 % margins because it's high end, super profitable chips are so in demand.
And what deep seek is showing is that maybe you know the world doesn't need that many of them.
So this is could be very dangerous for video in the long run.
But to talk about the future deep seek, I think it's interesting that they have some parallels with meta.
And that one reason I think it's going to be very difficult for a company like open AI to go open source is that the way they make their money is by selling access to their AI models.
Whereas Meta makes their money selling advertising.
And they can give away the AI for free, which I'm sure is part of the strategic calculus for why they chose to open source it.
And I would see something similar with deep seek as well in that the skill set to do AI is very similar to hedge fund.
You know, there's algorithmic timing, there's predictive analytics.
So you, of course, have to hire very advanced, highly skilled mathematicians, even theoretical physicists to do this work.
Because at the end of the day, it's really just a lot of math.
So I think this, again, is part of the strategic calculus for DeepSeek.
The only thing I would add here too, is that this is an enormous win for China, by making it open source.
Because China has been positioning itself.
As a provider of global public goods, such as the Belt and Road initiative, and now an open source AI model that anyone in the world, especially those in the Global South, can use.
So unless something dramatically changes, I see a pretty bright future for Deepsea.
And I think, of course, it's natural to complain about the lack of access to high -end video chips.
But at the same time there's a saying that extreme constraints leads to the greatest breakthroughs in design.
And I think this is exactly what we're seeing here.
The Chat Lounge. The Chat Lounge unpacks views and opinions on hot issues in a more casual way.
Right, like you mentioned, it's something positive, also agreed with by President Donald Trump.
is called DeepSeek a wake -up call for American industries.
You've envisioned a rosy picture for DeepSeek, but it seems it already suffered some setback in the States because a lot of government departments or not just in the States but in Australia as well.
Government officials are telling departments or some government -related companies to stop using DeepSeek.
So, Well, to you, if I may, just very briefly, I mean, this is inherently a political question.
And I think we can look at TikTok for some guidance here.
And I think whether that's DC or some other future tech company coming from China, as long as they can find a way to provide decisive political value to powerful leaders.
They will find a way to thrive.
So I think it's still very early and again it's very natural for the American government, the other Five Eyes government to all see this as a national security threat, but I I think the next move will be on the part of DeepSeek, or you know, these other companies, DeepSeek's the product, you know, but the company to find a way if that's part of their strategy.
Because again, given the cost advantage that they have, you know, I think there's going to be tremendous demand everywhere around the world for what they have to offer.
Can I get a bit of a prediction?
Sure. Which is, I think it's going to unfold like the internet has unfolded.
Because essentially today we have two internets.
We have the internet that runs in the West and it provides like Google and meta.
And then in China you have all these, you have all these alternative rival services and it's a complete in some senses, it's almost like it's a separate internet and the same is going to happen with AI because of these geopolitical tensions, they're going to be Western offerings and Chinese offerings.
And people are going to choose whichever, you know, is aligned with their geopolitics, not necessarily, because the technology is going to be pretty comparable.
Just like the internet providers you have, various internet services, social, media and so on.
But what if its access is totally banned?
You cannot download it as right?
Well, you already have that, you can't access Google in China.
So similarly, you won't be able to, perhaps you won't be able to access TikTok in the West, you're going to get restrictions and firewalls are in place.
Again, that reinforces, is in a technical sense, the geopolitical separation that we have, we've ended up essentially with two internets.
We'll end up with essentially two A .I.
worlds as well. And another question is, has DeepSeek R1 performance made Trump's $500 billion Stargate project obsolete?
Not at all. There's two fronts to the A .I.
race. One is who can build the best, smartest, cheapest AI models and Deepsea is definitely throwing the cat amongst the pigeons there, who's going to win that part of the race?
But there's a second part of the race that we don't talk about much, it's not very visible, which is how you deliver those AI models, the infrastructure, the digital infrastructure that delivers the internet.
We take it for granted that we can just open Netflix and stream a movie and it's instantly there in high quality video and we forget what goes into that, that we forget that companies like Google own satellite networks, they own underseas cables, they own data centers in most countries.
It takes a huge amount of infrastructure that they've been building.
They've been building it for the last couple of decades and they're now gonna spend even more billions to actually provide, serve up those models at scale on a planetary -wide scale to billions of people.
That's still going to need significant investments.
And good news for those companies actually, DeepSeek is actually fantastic, good news for those companies.
Because it tells them maybe they need slightly less infrastructure than they were planning, because they can run slightly smaller models.
But you're still going to need huge amount of infrastructure.
You're still going to need those undersea cables.
You're going to still need satellite.
You're still going to need data centers to do the grunt work.
So it's been made, though, those, you know, those infrastructure investments a bit more manageable, a bit more reasonable.
But you're still gonna have to make those big bets.
But if Deepsea puts those Wall Street wolfs in awkward a situation, would they mass all measures they can to contain it?
I mean, let's not overreact here.
I mean, in the tech sector, tech stocks sell -off on Monday.
Actually some tech stocks went up, Amazon's tech stocks went up modestly, because the it correctly foresaw that actually this is good news for Amazon because most of Amazon's business is not AI models, but is that infrastructure, is all those data centers Amazon web services, that's the most profitable part of Amazon as a business indeed.
So that part of it, the delivering the AI infrastructure still stands, and indeed, we know who wins that game and certainly in the West we know who wins that game, it's going to be those big tech giants, it's going to be the Googles and the Amazons.
No one is going to be able to make the scale of investments necessary, to be able to deliver it up.
And similarly, the comparable Chinese alternatives, again, there's a huge amount of legacy hardware infrastructure they built, been very hard to catch up with that.
So it was always going to be a case.
For example, the OpenAI partnered with someone like Microsoft to deliver their exercises, and Tropic have got in paid with Google, or the DeepSeek, I haven't checked, you know, but DeepSeek won't be necessarily be serving up all of their model, they'll be paying some web service to do that for them, who provided that infrastructure, that's you know, that infrastructure still got to be built and still be paid for.
The good news is maybe we need slightly less of it than we thought we would.
Right, actually it's already under attack and the company said all addresses were chased back to the states.
That's not some good news for for DeepSeek, right?
Not at all, but I'm sure that's temporary.
That's part of the infrastructure.
You have to provide infrastructure that's resistant to these denial of service attacks.
Again, that's all part of it.
That's not really AI.
That's part of the digital backbone that the internet and digital services run on.
Again, people have to invest money to build that infrastructure so that it's resistant to the hackers and the people, mischievous people who want to call.
cause you harm, so that's all part of delivering AI.
You're gonna have to pay for that stuff as well.
DeepSeek's model doesn't change any of that equation.
You're still gonna have to make those big investments to deliver digitally to people, you know, the billions of people on the planet.
Right, it also has some other impact.
DeepSeek was developed by a quantitative hedge fund company.
So, Jacko, what are the implications there going short on U .S.
AI related stocks? I really don't think DeepSig has really that much to do with the US -related tech -related companies.
I mean, if we look at the bright side, what we're looking at is that the AI technology is not being used all around the world, you know?
You should be striving about the tech companies.
And basically, if you look at the open AI and DeepSig, all these companies' products, they are amazing, they are wonderful.
But the question is that, how much people are going to pay for them?
So, this is another question that is, you know, you can find the answer by saying, okay, Chinese are coming for the job, so that we're not getting money because China's developing a cheaper model.
It's not the case, because OpenAI itself is not making that much money at the moment.
So, it's always the question for tech companies about where they are going to get the income, who is going to pay for them, especially for these kind of new AI programs.
I mean, if you look at Apple, if you look at Amazon, they're already making quite a lot money because they have made successful applications for their tech products in their programs, but for the open AI, deep seek, the question really is, who is going to pay?
Are they going to get enough money?
And that is really a hard question to answer, because it's not a traditional business pattern that we can find.
We don't have this in the past.
So it's reasonable for the analysts on the wall streets and financial centers in Beijing to make predictions about the possibility in the future, but it's really hard to tell.
You can only make predictions facing, you know, 70, 60 percent accuracy.
So this is a question we can face in the AI companies at the moment.
Coming to our last question.
Meta, AI Chief Scientist, Jan Lekun said, people who see the performance of DeepSeek and think China is surpassing the U .S.
in AI, you are reading this wrong.
So my question is, do you envision China being able to surpass the US in AI in the future or foreseeable future?
Let's begin with maybe Andy.
Sure, thank you Tuyenin.
I believe that Jan Lekun believes that this is a triumph of open source, not necessarily the triumph of one country over another.
But I will say, again, given the depth and breadth of China's technological talent pool, the size of its market, and the ferocity of the competition in the domestic market, that it would be quite remarkable if China did not establish leadership in key areas of artificial intelligence.
Thanks for that. And next, Jacquel, please.
Wow, I think it's really difficult to tell whether China's going to surpass US, or US is going to surpass China.
I mean, currently, it's very obvious that the US is having quite a lot of leadership with tech companies, but China's following up.
You know, currently, if you look at China's per capita GDP, we're about one -fifth of the United States, or one -sixth of the United States.
If you look at this from the per capita GDP, then we're lagging quite behind.
If you look at the tech company, we are following up, but how do you evaluate that China surpasses the United States?
Do you evaluate this from the number of companies that China is having compared with the US?
Then China is definitely surpassing the United States, because we've got four times larger in the population.
Are you looking at the value of the whole tech company and the whole tech industry?
Then currently, China is following up, but it's not surpassing the US.
So in the future, I think the more possible cases will be that China and the U .S.
are leading in their own perspectives, the tech industry, and each has their own unique advantages.
For example, U .S. is very advanced with basic mathematics and physics, which China is trying to catch up with still acting behind.
But if you look at e -commerce, China got a much larger e -commerce market because we got a larger consumption market.
So each market or each country and economy will have their leading aspect.
There is no direct competition, say, who is a winner is going to lose?
We are both going to have very thriving tech markets in the future and we can definitely learn a lot from each other.
I think that is the meaning of modern economics is not finding out who is winning, who is losing, but both are thriving out from the development of technology.
Fair enough. And the last one, Elise, Professor Walsh, please.
Yeah, I agree. I don't think there is one winner.
It's like saying you know AIs often compared to electricity, we didn't have one winner of electricity, China is a major user of electricity and they've got some wonderful companies that build electrical devices, just as the US is, just as Germany is, just as lots of countries get to use electricity, exploit it, profit from it, the same is going to be true for AI, so it's going to be leader technology, intelligence is widely distributed around the planet, there's 8 billion intelligence around the planet, in particular, it's going to be 8 billion ultra -rich intelligence around the planet.
There won't be one winner.
The deep -stick news, though, is good news, because it actually demonstrates that it is going to be available, not just to the OpenAI and Google, with the DeepPockets to build it, it's going to be available to lots of people with much shallower pockets.
Hopefully, the U .S.
side, especially, U .S.
government officials, can realize that, too.
And with that, we wrap up this session of the chat lounge.
Many thanks to Professor Toby Walsh, Chief Scientist from the AI Institute, University of New South Wales, Australia, Chen Shagla, Chief Investment Officer of Beijing -based Novomarcade Technologies, and Andy Mach, Senior Research Fellow at the Center for China and Globalization for sharing your insights.
The show is available on all major podcast platforms.
Please email us your comments at radio at cgtn .com.
To you, thank you for listening.
We'll have more Check at the Chat Lounge next week.
I cannot marry you while your mother recalls you.
Stepping to the world of timeless elegance with Mrs. Spring Fragrance, a captivating audiobook by Edith Eaton, read by Man Leng.
Experience the lively, yet complex lives in Chinese -American communities at the turn of the 20th century as Edith Eaton's delicate prose weaves together themes of tradition, change, love, and identity.
forget all that belongs to the joy of life.
Follow the journey of Mrs. Spring fragrance and witness the struggles and fights of those living on a rapidly changing exotic land while holding on to their cultural traditions or more importantly to their identities.
He could hardly believe that his daughter was seriously opposed to becoming the wife of a such a good looking prosperous young merchant.
Join Books and Beyond for Mrs. Spring Fragrance.
Let the stories come to life in your imagination.
Subscribe to Books and Beyond for free on your favourite podcast platforms or log onto radio .cgtn .com for more.
Sightline Story brings you all things sports -related.
The hottest topics, latest events, juiciest stories all with a very personal take.
Subscribe to Sideline Story Podcast for heated sports discussions, covering events that are happening in China and around the world.
