The story behind DeepSeek’s breakthrough
17:49 Share

Recently, my colleague Eleanor Olcott took a trip to China's southern Guangdong province. She traveled to a tiny farming village to visit the childhood home of Liang Wenfeng, the founder of the artificial intelligence startup DeepSeek.

That success Eleanor's talking about has to do with something very big that happened last month.

You probably recall this startup DeepSeek shocked the world by proving that China could be competitive on AI. They showed they could potentially compete with the likes of American companies such as OpenAI or Meta. We've just walked past Liang Wenfeng's house.

You can hear firecrackers going off in the background because it's Chinese New Year and people singing karaoke as well. Liang and his parents left yesterday, apparently escorted by a bunch of security guards because they are being inundated by visitors from all over Guangdong coming to pay their respects. Liang Wenfeng became a celebrity practically overnight.

So much so that on Eleanor's visit to Liang's hometown, she found people who had traveled there just to see where this new hero of China had been raised. So I spoke to this young teenager who traveled from Guangzhou to Mililing, this village, and

And I asked him what he thought of Liang Wenfang. His response was really revealing and I think speaks to a broader attitude of the Chinese towards Liang. He said that he's a pragmatic technologist and actually because of him, that China has been able to do something which they haven't been able to do before, which is

not only compete but also surpass the likes of OpenAI. He's become this kind of rallying call for Chinese pride, for Chinese sense of confidence in technology. And he said that Liang has made a big contribution to the country, a big contribution to China. Now, all of this excitement around DeepSeek says a lot about the health of China's tech sector.

For years now, it's been in a tough place. Overall, the conditions are very difficult, especially for people in that kind of traditional internet space. We've seen structural changes to the economy, geopolitics, and also Xi Jinping's reforms. All of these are confined to dampen the entrepreneurial animal spirit in China. So with the odds stacked against him, how did Liang Wenfeng do it? I'm Michaela Tendera from the Financial Times.

And this is Behind the Money. If you've been watching the tech sector for a while, you might remember a different era in China. When internet giants like Alibaba and Tencent appeared unassailable, when their market cap was larger than tech giants like Facebook and Amazon, and foreign investors piled in to get a slice of these companies. But by the time that Eleanor joined the FT as our China tech correspondent in 2021, she

She told me that, thinking back on it, her start date on her beat felt a little bit late to the party. I think the starting point was really the end of 2020. We've seen a number of reforms that have basically made it less profitable to run some of these tech companies.

In reaction to this, right, we saw foreign investors retreating from China because it became very clear that regulatory power in China was very strong and that you couldn't be assured that your investments wouldn't be vulnerable to kind of sudden change in winds on the regulatory side. Beijing cracked down on tech companies that were considered to control too much of their market or to be not aligned with Communist Party values.

This led to a series of anti-monopoly investigations and restrictions on certain businesses like gaming and edtech. No surprise, this all hampered business. And all that intervention has been only part of the problem for China's tech sector. We obviously had the pandemic, which shut down large swathes of the country for several months on end.

And this one, in many ways, is the most important when it comes to the perspective of the entrepreneurial environment, is that as foreign investors have retreated, we've really seen state investors, so those are local governments and state-owned enterprises in China, stepping in and playing a much bigger role in bankrolling innovation. Here's the thing. Funding from the state has sort of always been there for startups in China.

But as Eleanor said, it's taken on a bigger role as foreign investors have pulled back. And that comes with some strings attached. State entities will, in essence, make you pay back the money they gave you if your company doesn't succeed. This obviously makes starting a business feel riskier and way less appealing.

And this is really detrimental to the overall entrepreneurial environment because you're suddenly asking a founder who's launching a risky company that if they are unable to exit that investment, they are personally liable to

for that. It's kind of a form of debt, as it were. So we've seen this kind of wave of founders going on the debtors blacklist and many, many really tragic stories of founders who, you know, have lost everything and have basically been forced to kind of put their house on the line, et cetera, because of this trend. Which brings us back to the question, how did Liang Wenfeng do it? How did he create DeepSeek?

Could you just tell me very briefly, in a sentence or two, what is DeepSeek? DeepSeek is a small AI startup based in Hangzhou. It was formed out of a quant hedge fund by this mysterious billionaire, Liang Wenfeng. And they've been quietly plugging away, releasing more and more impressive models since late 2023. But really, the world finally woke up to them at the start of this year.

And why did the world wake up to them this year? So they released a series of models that were on par with leading ones from the U.S. rivals. So that was surprising, first of all, from a Chinese company. But it was doubly surprising, and this is really crucial, because they appeared to be trained on a far fewer number of chips.

So tell me a bit more about the founder of DeepSeek. We know that he's from this tiny village, but what about his professional life? Liang Wenfeng is a quant trader by background. He's a true engineer, someone who's really obsessed with the technology. He really believes in the power of machines and believes that humans are fallible and that we should be trying to replace humans.

you know, human fallibility with machines. So at the hedge fund, that means using algorithms instead of human decision-making. At DeepSeek, that means trying to achieve AGI, that is artificial general intelligence, i.e. the point at which machines will be capable of human-level critical thinking.

So Liang Wenfeng is a math nerd. He's been running a quant hedge fund called High Flyer for about a decade in China. But around 2021, he became obsessed with the potential of AI to build something larger than his trading shop. And he started buying Nvidia chips in bulk for that purpose. They're needed to train advanced AI models. And all that turned out to be really good timing.

That's because by 2023, when Liang launches DeepSeek, the U.S. has imposed really tough restrictions against selling NVIDIA chips to China.

People at the time just thought he was this weird, eccentric billionaire who was pouring money into a vanity project. But it turned out that that was actually a really prescient bet because once the race to kind of replicate open AI started in China, DeepSeek and Liang were in a really good position vis-a-vis some of the competitors.

Now, there's a few things that Liang does to get DeepSeek off the ground. The first is that he stays far, far away from that state-backed, strings-attached financing that I mentioned earlier. So this is an incredibly expensive endeavor. Not everyone has money to buy these expensive AI chips to build models. Liang Wenfeng is uniquely positioned to be able to do this. He's incredibly wealthy from his quant background.

trading days. So he's kind of been channeling these resources into what was initially somewhat of a side project, but has now become his full-term focus. So since he kind of self-funded this thing, Liang has the luxury of modeling DeepSeek in a way that's similar to the early days of some of the AI giants like OpenAI and DeepMind.

which is basically like a pure AI research lab. That means they're not really worried about generating revenue right off the bat. It's just experimenting.

So everything I said previously about the entrepreneurial environment being difficult because of the funding constraints doesn't really apply to this company. And actually part of the reason for their success, part of the reason why they've been pushing out such innovative models and have been able to open source those is because they've been under no pressure to commercialize. Liang also has something else that's unique. The manpower from his quant fund to use and put toward this AI research.

High Flyer had a group of extremely talented so-called systems engineers. These are guys who basically set up the computing infrastructure to do these trades. And it turns out that actually this kind of talent, the ability to construct the computing infrastructure to execute these trades is actually very, very helpful when you're also trying to squeeze as much computing power out of these chips in order to train large language models.

So Eleanor, what else do we know about Liang Wenfeng? He really gets into the weeds of how the technology is developing. And by all accounts is someone who is really kind of obsessed with the detail and is actively involved in the process of developing DeepSeek's technology. The result? A model that's being hailed as China's answer to Silicon Valley's AI prowess.

You may want to consider shopping through Instacart this winter if you like watching snow more than walking in it,

celebrate soup season, or own several pairs of fuzzy socks. Because staying home means staying warm, and some days just scream delivery. So this season, get everything you need right from the comfort of your robe. Download the Instacart app and get delivery in as fast as 30 minutes. Plus, enjoy $0 delivery fees on your first three orders. Service fees, exclusions, and terms apply.

Eleanor wrote her first story about Deep Seek back in June of last year, way before all of this attention from the West popped up. I asked her what it's been like to watch the story unfold in China and what she thought it all meant. Eleanor, so how did he end up on your radar so many months before, you know, this came out into the forefront of most people's awareness?

I first actually heard of DeepSeek a couple of years ago. One of my contacts said that a very mysterious quant fund had gathered one of China's largest cluster of NVIDIA GPUs

I didn't think much of it until this company started releasing their models. So they released their first one at the end of 2023, but they became more and more competitive. Yeah. And so I guess how surprising was it to be writing about, you know, a tech startup like this in the middle of 2024? Yeah.

It wasn't so surprising because actually generative AI has been one area where there's been quite a bit of financing, nowhere near the scale of the US players, but has received a lot of money and explicit kind of government support. So it wasn't really surprising in the bigger context, right? Like this is a complicated picture. It's not the same in every single one of these sectors, right?

But actually, you could also argue, aside from that, that DeepSeek is kind of the exception that proves the rule, right? It hasn't raised any external financing. They haven't been under pressure from their state investors or VCs to show that they can make money. So did you have the chance to interview Liang while you were working on that reporting last year? I mean, have you ever met him?

No, I did not interview him then. I've seen a whole bunch of Western media since this has come out asking for interview requests. And I found that quite funny because that's just not how things work in China, right? Like if

you're a tech executive under such intense scrutiny, given the kind of heightened geopolitical environment, like they're not going around giving interviews to the Western media. He has given a couple of interviews to the Chinese press a year or so ago. And basically what's been happening since then is people have just been, you know, trying to extract as much information from the very, very limited amount of public information there is out there on this guy. Yeah.

And why is that? Is the pressure you're talking about something that's internal or external? They're worried about both. They're worried about the Washington reaction because D.C. has, you know, put a lot of limits on Chinese companies that they view as making too many strides in certain areas of high-end technology. And they're also worried about the Chinese reaction because Washington

We're just kind of emerging out of this era where big tech companies that were making lots of money have been under intense scrutiny in China. So it's kind of a lose-lose situation if you go out and you're very loud about your successes. Eleanor, they made this big splash. DeepSeek made this big splash. Obviously, Liang is getting all of this attention in China. But what is next for this company? What is on the near horizon for DeepSeek and for Liang?

So deep seeks really in growth mode right now, they're going to be continuing to plug away and releasing more and more competitive models. It's unclear, like precisely what their commercial plans are, if they have any, I've been hearing that they're on a big

spree at the moment. And actually, this is a really good moment for DeepSeek because, you know, previously they might have found it difficult to court the best AI talent because they were somewhat of an unknown. But right now, like they are the hottest company. They are the open AI of late 2022. Like everyone wants to be a part of this organization.

It's not entirely clear where they go from here, whether or not they're going to commercialize their technology. But what is clear already is that they've sparked off a kind of chain reaction in China where all of the other generative AI players are trying to absorb and replicate some of their learnings and really focus on some of the techniques that DeepSeek has been pioneering to drive down the cost of training and running these models.

So do you think there's a scenario in which Liang gets kind of too big for himself? I mean, is that something that you'd worry about if you were him? That is a very good question, right? One thing that will be interesting to watch, and I think potentially is a bit of an issue for companies like DeepSeek, is Beijing has thus far been supportive of generative AI companies and

But, you know, he's looking to build AGI. He's looking to build machines that are capable of human level critical thinking. What will that mean for Beijing once they start seeing, you know, this technology becoming more and more competent? I do foresee there being some tensions between, you know, the way the technology is evolving and actually the political response to that. Something to keep an eye on for the future.

Behind the Money is hosted by me, Michaela Tendera. It's produced by me, Safiya Ahmed, and Katya Kamkova. Sound design and mixing by Sam Giovinko and Joseph Salcedo. Original music is by Hannes Brown. Topher Forges is our executive producer. Cheryl Brumley is the global head of audio. Thanks for listening. See you next week.

Hi, this is Freddie Wong from Dungeons and Daddies, and this episode is sponsored by Rocket Money. Houston, we have a savings. That's right. Rocket Money is a personal finance app that empowers you to save more, spend less, and take control of your financial life. Basically, you can see all of your checking, savings, credit cards, and investments in one convenient place. You can understand your spending trends, and most importantly, you can categorize those expenses individually.

and set up a custom budget by identifying top spending categories. Listen, you want track bank statements? It's very difficult. Use Rocket Money to help keep track of all your finances. Rocket Money has over 5 million happy members and it's saved its users over $1 billion across all the app's premium features. Let Rocket Money help you reach your financial goals faster. Get Rocket Money today at rocketmoney.com slash all in one. That's rocketmoney.com slash all in one.

From Paris, France to Paris, Texas, use your American Express Green Card and transform everyday purchases into memorable experiences. Earn three times membership rewards points on travel and on transit with Green for Memex. Learn more at americanexpress.com slash explore dash green. Terms apply.