Copyright Statement
Reprinted from Finance Magazine, copyright belongs to the original author, only for academic sharing. If there is any infringement, please leave a message for deletion.
Artificial intelligence is currently in a period of explosive innovation; only through continuous innovation can one remain at the center of the stage.
Written by Mark, Executive Editor of Finance
A Brief History of Artificial Intelligence
What Innovations Has DeepSeek Made?
Today, DeepSeek has become a challenger; has it really surpassed OpenAI? Not quite.DeepSeek has surpassed OpenAI in certain aspects, but overall, OpenAI still leads.
First, let’s look at the foundational models of both sides. OpenAI released GPT-4-o in May 2024, while DeepSeek released V3 on December 26, 2024. The Stanford University Center for Research on Foundational Models has a global comprehensive ranking of foundational models, with the latest ranking on January 10 of this year. There are six indicators, and after summing the scores, DeepSeek V3 scored 4.835, ranking first; GPT-4-o (May version) scored 4.567, ranking sixth. The second to fifth places are all U.S. models, with Claude 3.5 Sonnet in second place, scoring 4.819; the company developing this model, Anthropic, was only founded in February 2021.
★ Algorithm Section: The New Reinforcement Learning Algorithm GRPO.
The goal of reinforcement learning is to allow computers to learn autonomously and complete tasks without explicit human programming instructions, which is an important method toward general artificial intelligence. Reinforcement learning was initially led by Google, which used reinforcement learning algorithms when training AlphaGo, but OpenAI later took the lead, releasing two new algorithms, TRPO (Trust Region Policy Optimization) in 2015 and PPO (Proximal Policy Optimization) in 2017. DeepSeek has taken it a step further by introducing a new reinforcement learning algorithm, GRPO (Group Relative Policy Optimization), which significantly reduces computational costs while improving model training efficiency.
(GRPO Algorithm Formula. Source: DeepSeek-R1 Paper)
Can DeepSeek Continue to Innovate?
Of course, China is not the only one fighting; there are many excellent artificial intelligence companies in China. In fact, for many years, global artificial intelligence has been dominated by a dual peak of China and the U.S., with the U.S. peak being somewhat higher.
Nevertheless,I still have confidence in Liang Wenfeng and the DeepSeek team. From Liang Wenfeng’s few interviews, it can be seen that he is an idealistic person who is also pragmatic and has a keen business sense. He certainly understands technology, but he is not a technical genius; he is likely a technology entrepreneur like Jobs or Musk who can gather technical geniuses to create great products.
In an interview with “Undercurrent”, Liang Wenfeng said: “Our core technology positions are primarily filled by fresh graduates and those with one or two years of experience. Our selection criteria have always been passion and curiosity. When hiring, we ensure that values align and then use corporate culture to ensure a unified pace.”
“The most important thing is to participate in the wave of global innovation. For the past thirty years in the IT wave, we have hardly participated in true technological innovation. Most Chinese companies are accustomed to following rather than innovating. The real gap between Chinese AI and the U.S. is originality versus imitation. If this does not change, China will always be a follower.”
“Innovation is primarily a matter of belief. Why is Silicon Valley so innovative? It is primarily because they dare. We are tackling the most difficult problems. The biggest attraction for top talent is certainly to solve the world’s toughest problems.”
Jobs once said: Only those who are crazy enough to think they can change the world can change the world. I see the shadow of this saying in Liang Wenfeng.
However, we must not be blindly optimistic about China’s AI surpassing the U.S.; DeepSeek has not disrupted the path of large models driven by computing power, algorithms, and data. Many of DeepSeek’s innovations are due to constraints imposed by chips. For example, the communication bandwidth of NVIDIA’s H100 is 900GB per second, while the H800 is only 400GB, but DeepSeek can only use the H800 to train its models.
Recently, I have seen a lot of comments from both sides of the Pacific regarding DeepSeek. “Necessity is the mother of invention,” a saying from ancient Greece, has been reiterated by various experts. But conversely, DeepSeek’s ability to compete evenly with OpenAI’s similar products relies on leveraging algorithmic advantages to compensate for computational disadvantages. However, the competitors have been alerted; if they develop equally good algorithms with better chips, will the gap between China and the U.S. in large models widen again?
On the other hand, although DeepSeek can now adapt to domestic chips, considering the performance gap, the computational disadvantage is unresolved in the short term. Unless we can replicate the scenario where electric vehicles reverse the gasoline vehicles and achieve a leapfrog advantage, such as replacing silicon-based chips with quantum chips.
It is tragic to be caught in such thoughts—technological innovation should benefit all humanity, yet it is distorted by geopolitical factors. Therefore, we should applaud DeepSeek for resolutely taking the open-source route.
<img src=”https://mmbiz.qpic.cn/mmbiz_jpg/GibKatX4zYibQEicj9l2whRvqibLAKbNpiacz1vvXmoeiaRWUqialo0zBFU2SkHsaYoxnxZcFIibrLpDkWlic8Ae88srSQg/640?wx_fmt=jpeg”