One of the various Chinese AI hopefuls looking for to tackle corporations like OpenAI comes from an unusual source: a quant fund that dominates the country's financial sector.
High-Flyer Capital Management, a Chinese quantitative hedge fund, has grown into an asset manager with around 60 billion RMB ($8 billion) in assets since its founding in 2015, using some AI and algorithms to discover patterns or variables that might affect stock prices.
The company has now translated this information and infrastructure into a strong AI model that has been published and which, based on experts, is by no means inferior to leading Western approaches. DeepSeek-V2 can answer questions, write code and think logically.
DeepSeek costs significantly lower than its competitors, about two RMB per million tokens issued – or words returned per query – which sparked a price cutting war amongst Chinese artificial intelligence providers.
Per week after the launch in May, tech giant ByteDance cut prices to as little as 0.60 RMB per million tokens issued. Rival Alibaba then cut usage prices for a few of its models by as much as 97 percent, and Baidu made two of its Ernie models available without spending a dime.
The launch of the brand new model, which has quickly attracted hundreds of Chinese developers, shows that despite their early pioneering roles in generative AI, tech giants like Baidu and Alibaba face stiff competition from more agile upstarts. It has also put China's highly competitive generative AI race within the highlight.
“The gap between the US and China shouldn’t be as big as everyone thinks,” said Liu Qingfeng, founding father of the Chinese AI group iFlytek, recently at a tech meeting in Macau. “In many areas, our (models) are higher than theirs.”
DeepSeek's development is being fueled by funds from its sister hedge fund High-Flyer. Its funds have returned 151 percent since 2017, or 13 percent annualized, and have been generated in China's battered domestic stock market. The country's benchmark index, the CSI 300 index, which tracks China's 300 largest stocks, has risen 8 percent over the identical period, based on research firm Simu Paipai.
In February, Beijing cracked down on quant funds, blaming their high-speed algorithmic trading for a stock market sell-off earlier this 12 months. Since then, high-flyer funds have lagged the CSI 300 by 4 percentage points.
High-Flyer and DeepSeek didn’t reply to requests for comment.
The quant fund began in a Chengdu apartment where founder Liang Wenfeng, a pc science graduate from Zhejiang University, experimented with automated stock trading, local media reports. According to his profile on the China Asset Management Association registry, he worked as a freelancer until 2013, when he founded his first investment firm.
According to manager Cai Liyu, by 2021 all of High-Flyer's strategies were using AI, employing strategies just like those of the highly profitable hedge fund Renaissance Technologies. “AI helps extract invaluable data from huge data sets that could be useful for predicting stock prices and making investment decisions,” he said during a roadshow streamed online this 12 months.
Cai said the corporate's first computer cluster cost nearly 200 million RMB and High Flyer is investing a few billion RMB to construct a second supercomputer cluster that may span an area in regards to the size of a football field. The majority of their profits are being ploughed back into their AI infrastructure, he added.
The second cluster, now accomplished, connects greater than 10,000 of Nvidia's cutting-edge processors with servers and storage, giving DeepSeek the computing power to coach a big model, based on archived versions of the corporate's website. The group acquired the Nvidia A100 chips before Washington restricted their supply to China in mid-2022.
“We all the time desired to do experiments on a bigger scale, so we all the time tried to make use of as much computing power as possible,” founder Liang told Chinese tech site 36Kr last 12 months. “We wanted to search out a paradigm that might fully describe all the financial market.”
The company is one among six Chinese groups with greater than 10,000 A100 processors, that are widely considered the processing power threshold for self-training large models, based on Guosheng Securities. The other five are all Chinese tech giants, but their combined processing power pales compared to U.S. corporations. Meta has said it should have processing power similar to nearly 600,000 of Nvidia's more advanced H100 chips by the tip of the 12 months.
Tests by research groups rank DeepSeek-V2 as among the finest LLMs on this planet. Researchers on the University of Waterloo in Canada ranked it among the many top 10 models behind OpenAI's GPT-4, Anthropic's Claude and Chinese competitor 01.AI.
DeepSeek's model can also be open source, so AI researchers can study and duplicate its structure.
“The architecture of the model is exclusive,” says Andrew Carr, chief scientist at Cartwheel, a US-based AI animation startup. “DeepSeek has taken this concept of expert mixing, where you break a model down into smaller parts, with a whole lot of little experts, to the intense.”
Carr said the model could be very near Meta's latest Llama 3 model but cheaper, costing about one-hundredth the value of OpenAI's GPT-4 and one-fifth the value of Anthropic's Claude 3 Haiku.
Tiezhen Wang, an engineer at New York-based AI research center Hugging Face, said the DeepSeek team reduced the quantity of knowledge the model has to recollect while allowing it to “handle more tasks concurrently without slowing down.”
In China, the pricing strategy has helped attract developers. Wang Zixu, a programmer from northern China, said he switched to DeepSeek due to OpenAI's GPT-4's lower prices as a programming aid.
Despite the associated fee advantage, some industry experts said DeepSeek could potentially make a loss as a result of its low price. Its computing power could also fall further behind that of competitors as Nvidia launches recent chips which are banned from exporting to China.
Nevertheless, High-Flyer's AI offshoot has set itself the goal of being the primary to attain artificial general intelligence, the purpose at which machines have greater cognitive abilities than humans.
“We imagine AGI is the overwhelming fantastic thing about model x data x compute power,” said a DeepSeek job posting. “Join us on a 'deep search' on the journey to AGI!”
