Kai-Fu Lee is arguably the most prominent ‘face’ for artificial intelligence outside of Silicon Valley. Having once run Google’s business in China, he knows how to talk to a Western audience, all the while leading venture investing into China’s AI ecosystem through his firm, Sinovation Ventures.
More recently he founded AI.01 in Beijing, a company meant to develop its own large-language models, á la ChatGPT. This year he abruptly shifted the startup’s business focus, he said at the HSBC Global Investment Summit in Hong Kong.
That was in response to the launch of DeepSeek’s most sophisticated LLMs and the way they were built – using open-source protocols. Now Lee is leading the company to develop applications that can operate on top of DeepSeek, including in finance.
There are several obvious areas where China’s financial institutions are using AI, both established versions of machine learning as well as LLMs: to spot risks, detect fraud, personalize customer service, and automate more operational processes.
But the most exciting area – the one with a dramatic track record – is to generate investment strategies and to automate trading.
High-Flyer and AI in quant
DeepSeek was founded in 2023 as a non-profit by Liang Wenfeng. This was Liang’s second company. His first, which he continues to run, is Hyper-Flyer Capital Management in Hangzhou – a hedge fund.
Liang co-founded High-Flyer with some buddies from Zhejiang University in 2016 to try to run investment strategies using machine learning to identify mispriced stocks and time their buying and selling. The critical component to making High-Flyer work were his student engineers and access to a lot of Nvidia chips.
The firm racked up enormous gains over its first five years, spawning copy-cat quants. By 2021 it managed Rmb90 billion ($12.4 billion) of assets and had notched annual returns of 20 percent to 50 percent, clobbering the CSI 300 index. Excited boosters began to refer to High-Flyer as China’s answer to Renaissance Technologies, the legendary algo-pioneering quant firm that beat markets consistently for almost three decades.
High-Flyer hit a rough patch, losing money and returning client monies, amid higher volatility in Chinese equities. Liang’s timing had been lucky, as the Chinese equities market had bottomed in 2015. Now conditions had changed, and his data-fed models couldn’t keep up. Worse, in 2024 the government blamed hedge funds for the underperforming market, imposing restrictions on how they traded. Last year, High-Flyer was forced to close its market-neutral products. Its AUM fell to about $7 billion.
DeepSeek’s breakthrough
Liang is a techie at heart, not a financier, and he founded DeepSeek as a non-profit developing LLMs using open-source protocols. There was a pragmatic aspect to this direction: because DeepSeek is a research lab, not a commercial outfit, it doesn’t have ‘customers’, so it doesn’t need to adhere to China’s strict data privacy laws. This gave it a much freer hand; indeed, it’s unlikely a commercial firm could have sourced data from the US. This helps explain why DeepSeek’s initial versions of an LLM resembled that of Llama, the open-source LLM developed by Facebook.
But its engineers – mostly youngsters drawn from Zhejiang University – had to make hard choices in the wake of US efforts to block China from accessing Nvidia chips. The breakout came with the release of DeepSeek R1, with R standing for reinforcement, that is, pure reinforcement learning.
Kai-Fu Lee said: “DeepSeek cracked the code of AI teaching AI.” Even if this achievement is more of an engineering feat than model innovation, it has huge implications for hedge funds like High-Flyer – and for asset management and banking at large.
In the US, the pioneers of large-language models for generating text – OpenAI, Anthropic, Google, Microsoft, etc – have tried to solve genAI problems by throwing at it ever greater amounts of data and computing power. Hence the multi-billions being earmarked for vast data centers and nuclear power plants.
This fund raising is arguably far out of proportion to whatever revenues these companies will make selling AI services.
Scaling inference
Worse, there is only so much data to go around, and the scaling laws that should drive better models (and adoption) are slowing down, meaning recent upgrades of ChatGPT have been underwhelming and require ever longer training periods to show results.
But, Lee notes, inference is also scaling. Inference is the process by which a trained model provides an output – such as a prediction, or a conclusion – based on new, unseen data.
Lee’s analogy is that training is akin to reading more books to boost your knowledge, whereas inference is how to think more critically, making more out of your bookish education.
“The shift in research has gone from training models to teaching AI to think, to challenge itself, to think about problems longer, and to generate better answers.”
Lee’s guess is that with inference now scaling, improvements to LLMs will accelerate. If previous upgrades to ChatGPT or Claude took two years, DeepSeek shows it can be done in three months.
Models upon models
Why? Because AI is now teaching itself. Training models has been based on humans writing algorithms, humans modeling architecture, humans inputting data. Now the biggest large-language models can be used to train smaller, more niche ones; reasoning models teach immature models how to self-assess. DeepSeek is training itself on Alibaba’s LLM.
“DeepSeek figured out how the reasoning engine works,” Lee said. OpenAI doesn’t show the chain of thought behind how ChatGPT comes up with responses. DeepSeek now does. And what it is doing is reinforcement. Earlier LLMs rely on human reinforcement: people tell ChatGPT when its responses are poor, so the model improves. DeepSeek’s R1 version is automating that process.
In practical terms this means it needs access to less new data, and it learns at a much lower cost – while delivering results comparable to ChatGPT.
This opens the door to making small-language models just as powerful as their larger brethren. Many companies have been developing SLMs to train them with their own data, rather than rely on the datasets of hyperscalers like AWS or Microsoft. This ensures more accurate responses, and also protects the data. It also makes responses more ‘explainable’ and less a black box. DeepSeek makes SLMs cheaper, more accessible, but also as powerful as a big LLM trained on the internet.
China quants boom
This is already transforming China’s funds management industry. High-Flyer now has plenty of copycats in China’s $700 billion hedge-fund industry. Because OpenSeek is open source, it is easier for firms to use its LLM and build their own investment and trading apps on top. Baiont Quant, Wizard Quant and Mingshi Investment are some of the funds enjoying big returns this year on the back of AI-led strategies.
But this has already leaked well beyond quant firms. Individuals are able to code their own strategies using DeepSeek. Dozens of firms in China’s $10 trillion funds industry are doing the same, but it’s really the opening of these tools to retail investors that represents a huge change.
It is no coincidence that Chinese equities have boomed in the past three months after a dismal 2024 – the rise coincides with the sudden craze for investing in AI-led funds like High-Flyer, or in traditional fund houses deploying their own AI strategies, or people just doing it themselves.
China may not be an easy template for what might happen elsewhere. Retail investors have always dominated its stock markets; local institutional investors are relatively small and not influential. (Another difference is the ability of the government to occasionally intervene in a massive way, drawing on a ‘national team’ of state-funded institutions to prop up markets.)
But with LLMs open source and increasingly affordable, they are rapidly becoming commodities. This is why DeepSeek’s R1 model sent a shiver of fear down the spines of US tech companies (Nvidia lost $600 billion of market value over a few day’s trading). Their huge investments, and proprietary business models of the badly-named OpenAI, look vulnerable to disruption.
Capital markets disruption
For financial institutions, though, developments in China are pointing to cheaper and faster and more reliable ways to transform their businesses. Most banks, insurers and asset managers are using AI in some version or other to analyze market sentiments, personalize conversations with customers, identify risks, and tailor investment advice. And if they’re not, the ability to do so is becoming much easier.
This will continue regardless of what happens to the businesses of DeepSeek or OpenAI. DeepSeek still has plenty of its own challenges: although it has found ways to cut its demand for chips and energy, it is still fighting tech giants for resources. So far it has been successful inside of China, but it will want to grow its business in other markets, which will set it in competition with American companies.
Meanwhile the demand side in financial services faces its own constraints. Banks’ data remains siloed, many firms still operate on legacy (that is, dangerously outdated) mainframe systems, and regulations vary from one place to the next. GenAI when used for credit scoring or investment tips will inevitably face ethical questions. And we can expect resistance among many bank executives who will see their staffing cut to ribbons.
For Kai-Fu Lee, these are short-term considerations. He has pivoted 01.AI to abandon building its own LLM in favor of developing applications for enterprises: AI agents, putting intelligence into apps, helping enterprises build their own software.
“The original belief is that one company would build an artificial general intelligence and the rest of the world would be forced to adopt that,” Lee said. “Now it seems the best model will be made available to everyone, and the rest of the world can just build applications on top.”
If so, this could accelerate changes across financial services. Retail investors using genAI to write their own investment programs? It could all end in tears (genAI hasn’t abolished market downturns). But these changes seem far greater than the current market enthusiasm they are engendering.
