Who Is The Mysterious Founder Of China's DeepSeek

Despite the global hype around China's DeepSeek, very little is known about the man behind it - Liang Wenfeng.

On today's Big Take Asia Podcast, host K. Oanh Ha talks to Bloomberg's Saritha Rai about the tech founder who led DeepSeek to the frontline of AI advances and what the company's rise tells us about the battle for AI dominance.

Here is a lightly edited transcript of the conversation:

K. Oanh Ha: Earlier this year, a new product from the Chinese AI startup DeepSeek shocked the world and rattled Wall Street.

Tom Mackenzie:  China's DeepSeek is freaking out the AI world right now. Tech stocks tumbled as its app surged to the top of the download charts.

Ha: But despite the global attention, very little is known about the man behind DeepSeek - Chinese entrepreneur Liang Wenfeng.

Saritha Rai: Liang Wenfeng is certainly a mystery figure.

Ha: Bloomberg's Saritha Rai covers artificial intelligence in Asia.

Rai: He's certainly one of the most inaccessible and low-key tech entrepreneurs that I've come across. Just to illustrate how private he is, we weren't able to find any pictures of him on the internet when we scoured through his website and all of that, but finally appeared in a really high-profile meeting with President Xi Jinping and that picture got out into the world and he was everywhere.

Ha: And what does this man of mystery look like?

Rai: He is slim, wears glasses but doesn't talk much

Ha: Baby-faced?

Rai:  Yes , I think we could describe him like that.

Ha: DeepSeek rarely answers questions about Liang, citing his privacy. But Saritha and her colleagues were curious about this man, whose AI systems turned the tech world on its head. So they spoke with dozens of people familiar with his work: from former employees and fellow researchers to investors and insiders in the industry.

Rai:  And what we found is that yes, he is extraordinarily low-key, very, shy, but extraordinarily driven and talented and passionate. And I think he has somewhat taken on DeepSeek as a sort of a mission to establish China in AI, trying to make sure that China is a force to reckon with in AI.

Ha: Welcome to the Big Take Asia from Bloomberg News. I'm Oanh Ha. Every week, we take you inside some of the world's biggest and most powerful economies, and the markets, tycoons and businesses that drive this ever-shifting region. Today on the show: Who is Liang Wenfeng? We learn about the mysterious tech founder who led DeepSeek to the frontline of AI advances. Plus, what does the company's rapid rise tell us about the US-China artificial intelligence race?

Ha: Saritha, thanks for joining us.  I'm fascinated by AI. You guys did such an interesting job with the story. I wonder if we can start with - who is Liang Wenfeng? What do we know about his roots?

Rai: So Liang is about 40 years old, born in a small village called Mililing in the Guangdong Province. His parents were school teachers, mainly taught primary school kids. He was extremely bright and went on to study at Zhejiang University and then also did his masters there.

Ha: At Zhejiang University, Liang and his friends immersed themselves in all things tech: machine learning, signal processing, electronic engineering. They even developed programs to trade stocks during the financial crisis. After graduating, Liang joined forces with two of his classmates and set up a quantitative hedge fund called High-Flyer Management.

Rai:  So quant funds basically work with mathematical models and statistical analysis to do stock trading. Humans are not involved in taking the decisions. At its peak, High-Flyer Management was managing something like $14 billion in assets, so it was quite a sizable fund. And - in its most successful runs, it was providing annualized returns, averaging 35% to its investors. So I would say that it was doing very well indeed.

Ha: According to former employees, High-Flyer had a geeky startup culture. Its early job postings boasted of attracting top talent from Google and Facebook, and said they were looking for math and coding "geeks" with "quirky brilliance."

Rai: The early job postings, also referred to Sheldon, who is this very awkward, main character in the prominent American sitcom called the Big Bang Theory

Sheldon Cooper: For example, I cry because others are stupid and that makes me sad.

Rai: Sheldon has a legion of fans and is extraordinarily funny without meaning to be. So you know, the whole culture of DeepSeek in the early days revolved around recreating some of that geeky, nerdy culture. There were free snacks, poker game nights. Everybody was dressed in T-shirts and slippers.

Ha: Sounds like a great place to work

Rai: Yeah, it was really an unorthodox, startup culture. Unlike what you'll probably see in the big tech companies in China, such as Alibaba and Tencent.

Ha: And how did Liang transition from doing quant financials into AI and building DeepSeek?

Rai: Liang was always extraordinarily interested in machine learning and artificial intelligence and then a few months after OpenAI, you know, launched ChatGPT, the chat bot that became an overnight global success - it was then the spring of 2023, a few months had passed after the launch of ChatGPT - and Liang then announced that DeepSeek would be set up. In its early manifesto, DeepSeek talked about shunning mediocrity and tackling the big challenges in AI, and of course, ultimately cracking artificial general intelligence.

Ha: The manifesto also laid out DeekSeek's ambition - to position China as a leader in cutting edge technologies.

Rai: You know, Liang has given two interviews, rare as they may be. In both those interviews, he's talked about bringing China's AI ecosystem to the forefront of where the world is. You know, China has been accused constantly of being a copycat. He wanted an AI China to chart a different path.

Ha: DeepSeek worked fast. Since 2023, it's released over half a dozen AI models and helped pioneer a technique called sparsity - which enabled those models to train and run more efficiently. Developers started to take note. Then, earlier this year -

David Gura: Get back to that top story now, DeepSeek shaking up global tech...

Rai: When they released their reasoning model R1, that caused such an upheaval in the industry and caused a trillion-dollar stock market meltdown. That's when the world really started paying attention to this secretive AI entrepreneur in China.

Ha: And Saritha, what is so groundbreaking about DeepSeek's R1 model?

Rai: The AI industry until recently was always about, billions of dollars spent in building the infrastructure, the data centers, the graphics processing units in the data centers that would train these models. But what DeepSeek did was show that its models could match or even outdo on some benchmark measures what the latest OpenAI or Anthropic models were doing, and with far less computational power, with far less resources and as DeepSeek claimed with far less capital as well.

Ha: So how did Liang and his team manage to achieve true innovation - at what it says is a fraction of the cost? And what does DeepSeek's success say about the AI race between China and the US? That's after the break.

Ha: For much of the last decade, the US has tried to restrict China's access to semiconductors. Tensions reached a fever pitch in 2022 and the following year, when Washington targeted Beijing with two rounds of chip export controls.

Jon Erlichman: Nvidia and shares of semiconductor companies have slumped today after the Biden administration said it would tighten restrictions on exports of AI chips to China, now Nvidia told Bloomberg...

Ha: That limited sales from American firms like Nvidia, whose cutting-edge chips are used by tech companies to help train their AI models. The move presented a significant challenge for developers in China, but as Bloomberg's Saritha Rai says, it also forced them to develop workarounds.

Rai: Necessity is always the mother of innovation. This has been proven, by AI developers in China, nevermind the export curbs, they've still gone on to build good models that have benchmarked with the best around the world.

Ha: And one of the most innovative approaches from DeepSeek is the sparsity technique we mentioned earlier.

Rai: Now sparsity is something to do with building a model without having the high-end computational power. It's when a large language model doesn't have to be entirely harnessed to give a, an answer to a query. Instead, Liang and his fellow developers tried to apportion the expertise of the model into smaller expert groups and then only harness those groups that required to be used. So, in doing that, they made it much more computation efficient, and also much more cost efficient.

Ha: Is it basically, instead of using your whole brain, are you using just certain parts of your brain to do that computation?

Rai: That's exactly right, Oanh. You know, instead of entirely using every little gray cell in your brain, it only fires up those neurons or little portions in your brain that contain that particular field of expertise. And then bring that to, you know, respond to a query or give an answer to a particular question, whether it's a command or a coding question.

Ha: The sparsity breakthrough impressed DeepSeek's competitors, but its price point is what ultimately made headlines. DeepSeek said it cost them just $5.6 million to train its V3 model - that's far less than estimated $100 million OpenAI spent on its most advanced version of ChatGPT.

Rai: Now there is definitely a whole lot of skepticism around that number because just the infrastructure, the training of the model, the talent, and the time it takes, all of it adds to quite a sizable sum of money. So, the skepticism is warranted. People have estimated that there was no way DeepSeek could have pulled that off without at least a billion dollars or more.

Ha: Also in DeepSeek's favor is that AI startups like it have a staunch ally in China's government and President Xi Jinping. Saritha says Xi sees generative AI, robotics and other high-tech ambitions as beneficial to the state's agenda - part of a larger push for self-reliance in key technologies. And DeepSeek's success has spurred much bigger rivals such as Alibaba, Tencent, and ByteDance to release their own AI models.

Ha: Saritha, DeepSeek's model is entirely open-sourced at this point. That means any individual or company could incorporate DeepSeek's algorithms into their own programs. Why did the company choose this approach and why is that important.

Rai: Open-source, on one level, you could say that it is democratizing AI and taking it out into the world. But let's not forget that China's AI models would've otherwise found fewer takers around the world if they were proprietary models and were on par in terms of what they cost with Western companies such as OpenAI. By making it cheap and by making it open-source, China allowed people around the world to quickly take a look at the models and begin using them, allowing them to be adopted much faster in the business and AI ecosystem thereby outdoing the likes of OpenAI. Now, that's huge. It's not only about democratizing models, it's strategically about making sure that you cut out your competitor by making things so cheap that the world adopts it quickly and then it becomes mainstream.

Ha: As a result, Microsoft and Amazon both offer DeepSeek on their cloud services. And DeepSeek's models have been incorporated into Perplexity, an AI-powered search engine that also offers models from OpenAI and Anthropic.

Rai: There is definitely a question about, you know, how fast AI is advancing and there's a fear around the world about having all of the controls rest only with one or two companies in the world. I think that was what DeepSeek and others were trying to put out a message in the world, saying that all of the controls cannot be left to one or two companies and the proprietary models that they are building, it should be much more democratic. Therefore I think the open-source philosophy is about de-risking, concentration and allowing more people to build with technologies that are much more available.

Ha: Is there also potentially a kind of a clash of cultures or a clash of values as well when it comes to building AI between the West's approach and the Chinese approach?

Rai: Very clearly, because if you look at early models of DeepSeek or even, the, you know, not tweaked or fine-tuned models of DeepSeek, they are very much working within the boundaries of China's censorship rules. For instance, you cannot ask it questions about Taiwan or Xi Jinping without it giving a very bland, official answer. Whereas if you take that same model and you can train it with other data and make it culturally suitable to different geographies, that's one of the things that DeepSeek learned early on, that by open-sourcing the model and by giving developers and users a chance to customize to their own cultural context, DeepSeek could find much more quicker and faster adoption around the world than by controlling a lot of it and controlling it in a way that it could only give China-friendly answers around the world.

Ha: And while some applaud China's innovations in AI, many in the US suspect darker reasons for the success. An April report from a US House of Representatives committee alleged "significant" ties between DeepSeek and the Chinese government. It concluded that the company unlawfully stole data from OpenAI. The Chinese Embassy rejects those claims as groundless. Meanwhile, DeepSeek and Liang haven't commented on the House report.

Ha: Saritha, it seems like there is very much of an arms race of sorts when it comes to AI, certainly between the US and China at this moment.

Rai: It's very much a race and I think it would be too early to call a winner. All I can say is that. a year ago, I would not have called it as a close race. It's a marathon, but you have to go at a sprinting pace. We're really at the very beginning of it and for whichever country cracks the race, there's a lot of economic gains to be had. So every country, particularly the US and China, do not want to let up in AI.

Ha: And what are the challenges ahead right now for DeepSeek that you see?

Rai: I think one of the main challenges is what next? What can they do that out does what they've already done. But there's also, I think, for DeepSeek, competition within its own home ground. A bunch of China companies such as Alibaba and ByteDance and Tencent are building models that are outdoing DeepSeek's last flagship model. So there's this pressure on DeepSeek to do better. But also I think there is also the question around commercializing these models. How are companies like DeepSeek going to make money? There is no clear answer yet whether DeepSeek wants to make money, and if it does, how will it make money?

(This story has not been edited by NDTV staff and is auto-generated from a syndicated feed.)