News center > News > Headlines > Context
Dark Side of the Moon releases large mathematical model to benchmark OpenAI o1
Editor
2024-11-25 15:35:01 2,880

Dark Side of the Moon releases large mathematical model to benchmark OpenAI o1

Image source: generated by Unbounded AI

China’s artificial intelligence is quickly catching up with OpenAI. On the afternoon of November 16, China’s general artificial intelligence company Dark Side of the Moon announced: the release of a large mathematical model k0-math. This is the first time the company has launched a model product with enhanced reasoning capabilities. According to the company, the mathematical capabilities of k0-math can benchmark against the two publicly available models of the world's leading OpenAI o1 series, o1-mini and o1-preview. Specifically, in MATH, the most commonly used mathematical ability benchmark in the industry, the k0-math model scored 93.8 points, exceeding o1-mini's 90 points and o1-preview's 85.5 points. The score of k0-math is second only to the 94.8 score of the complete version of o1. It should be noted that although the k0-math model is good at solving most difficult mathematical problems, the current version cannot answer difficult-to-describe geometric graphics problems. In addition, this product still has some limitations that need to be overcome, including for overly simple mathematical problems, such as 1+1=? For this kind of question, the k0-math model may "overthink" and answer some answers that deviate from common sense. Yang Zhilin, the founder of Dark Side of the Moon, said in response to a question from Fortune: For reinforcement learning, "data" is a core issue. If the reward mechanism for large models can be done well in the future, unnecessary "data" will be avoided. Overthinking”. "We currently do not have any restrictions on the length of the answer, allowing artificial intelligence to think freely. Perhaps we can inhibit overthinking by changing the reward structure. This is the problem we want to solve next," he said. Dark Side of the Moon is one of the most highly valued large artificial intelligence model companies in China. Kimi smart assistant is the core product of Dark Side of the Moon. It currently has 36 million users. According to Alibaba’s financial report, Alibaba invested US$800 million in fiscal year 2024 to purchase a 36% stake in Dark Side of the Moon. It can be seen that Dark Side of the Moon was valued at approximately US$2.2 billion at that time. In May this year, new investors such as Tencent and Gaorong Capital joined, which also pushed the valuation of Dark Side of the Moon to over US$3 billion. Yang Zhilin, 31, graduated from Tsinghua University and received a PhD in computer science from Carnegie Mellon University in the United States. He has worked in the Meta AI and Google AI R&D teams. Kimi has experienced rapid growth over the past year. Because Kimi supports 2 million words of lossless contextual input, it performs well in text parsing and long text processing. This advantage gives it a unique advantage in tasks such as reading comprehension, document analysis, and long-form writing. In April 2024, the number of visits to the web version of Kimi Smart Assistant reached 20.04 million, an increase of 60.2% from the previous month, and the number of visits exceeded Baidu’s Wen Xinyiyan. At present, China's basic large model companies are facing fierce competition, and various companies including Tencent, Baidu, Alibaba and ByteDance have launched large model products. In the competitive landscape, ByteDance’s productsThe "bean bag" brand is becoming Kimi's most formidable competitor. In early November, Doubao ranked second in the global AI product list (App) list (aicpb.com), second only to ChatGPT. Since the beginning of the year, the cumulative downloads of Doubao have exceeded 100 million. As far as China is concerned, the top three are Doubao, Wen Xiaoyan and Kimi owned by Baidu, with monthly active users exceeding 10 million each. In response to the competition with Doubao, Yang Zhilin said that he is not too concerned about the competition itself. "Because competition itself does not produce value." He said, "Only by launching better technologies and products can we create greater value for users. This is our core issue now."

Yang Zhilin is most concerned about The data is user retention rate. He believes that Kimi has only reached the primary stage of general artificial intelligence. The user retention rate is positively related to the maturity of the technology. As the technology continues to improve, Kimi's user retention rate will naturally increase. He did not answer directly what the current user retention rate of kimi is, but only said that the improvement of this data needs to be "never-ending." However, many investors and Kimi competitors that Fortune spoke to said that they have not yet experienced the k0-math product and cannot comment on its functions and actual effects. Kimi's continuous launch of new products reflects the competition in the field of artificial intelligence between China and the United States. The industry generally believes that artificial intelligence is mainly composed of three major elements: algorithm, computing power and data. At the computing power level, the United States has an absolute advantage; at the algorithm level, Chinese companies are gradually catching up; and at the data and application scenario level, China has the will Applying artificial intelligence to various scenarios and gaining the ability to use data effectively is a major advantage. (Fortune Chinese)

Keywords: Bitcoin
Share to: