中国初创企业在AI语言模型领域的重大突破

A Chinese Startup’s Breakthrough in AI Language Models

七级（考研）偏易| 436词

刘立军供稿

Part I. Passage

A Chinese Startup’s Breakthrough in AI Language Models

The U.S. ban on sales of advanced artificial intelligence (AI) computer chips to Chinese companies has prompted innovation among Chinese firms. One notable example is DeepSeek, a startup founded in May 2023 by Liang Wenfeng, a former AI student and hedge fund manager. DeepSeek claims to have developed AI models that rival U.S. competitors using less advanced hardware at a significantly lower cost.  

DeepSeek’s large language models (LLMs) have gained attention for their efficiency and open-source nature, allowing users to view and modify the code. The company’s latest model, V3, reportedly matches the performance of leading closed-source models like OpenAI’s GPT-4o. According to a December 2024 technical report, V3 outperforms other open-source models and offers comparable results to top-tier alternatives.  

Experts are taking DeepSeek’s claims seriously. Jeffrey Ding, a political scientist at George Washington University, notes that V3 has closed the gap with leading LLMs and even outperforms GPT-4o in some benchmarks. Andrej Karpathy, a co-founder of OpenAI, praised DeepSeek’s achievements as a remarkable display of innovation despite resource constraints.  

China’s limited access to advanced AI chips has driven researchers to innovate with available hardware. DeepSeek’s success partly lies in its use of a Mixture of Experts architecture, which reduces computing power requirements by training only specific parts of the model for particular tasks. This approach not only enhances efficiency but also cuts costs. Training V3 reportedly cost $5.6 million, far less than the $78 million spent on training GPT-4o.  

DeepSeek’s V3 has practical applications in fields like climate prediction, medical research, and cosmology. Unlike its major competitors, DeepSeek operates independently of China’s tech giants, focusing solely on developing high-performing LLMs. Liang Wenfeng has emphasized that the company prioritizes research and innovation over business opportunities, with the ultimate goal of achieving artificial general intelligence (AGI), a form of AI that matches human cognitive abilities.  

However, DeepSeek faces challenges. Its open-source approach allows competitors to build on its methods, and limited access to advanced AI chips could hinder future progress. Analysts suggest that Chinese firms, including DeepSeek, must continue pushing the boundaries of software and systems innovation to remain competitive.  

Despite these challenges, DeepSeek benefits from operating in China, where Western AI models like ChatGPT are blocked due to censorship. However, DeepSeek’s V3 appears to navigate political sensitivities carefully. For instance, when asked about Tiananmen Square, it avoids the topic, but it provides a neutral and factual response when asked about the origins of the COVID-19 pandemic.  

DeepSeek’s achievements highlight the potential of ingenuity and innovation in overcoming technological constraints, marking a significant step forward for Chinese AI development.  

【词汇】

1. outperform v. 超过，胜过  

2. benchmark n. 基准，标准  

3. cosmology n. 宇宙学  

4. analyst n. 分析师  

5. factual adj. 事实的，真实的  

6. pandemic n. 大流行病，疫情  

7. ingenuity n. 独创性，巧妙  

Part II. Questions

Q1. What prompted innovation among Chinese AI firms like DeepSeek?

A. The U.S. ban on sales of advanced AI chips.  

B. The collaboration with China's tech giants.  

C. The high cost of training AI models.  

D. The global demand for climate prediction tools.  

Q2. How does DeepSeek’s V3 model compare to other AI models according to the December 2024 technical report?

A. It performs worse than GPT-4o in benchmarks.  

B. It outperforms other open-source models and rivals top-tier alternatives.  

C. It is exclusively used for specific tasks like cosmology research.  

D. It relies on advanced AI chips for superior performance.  

Q3. What is one advantage of DeepSeek’s Mixture of Experts architecture?

A. It avoids political sensitivities in AI responses.  

B. It ensures the model operates independently of China's tech giants.  

C. It reduces computing power requirements by focusing on specific tasks.  

D. It increases the cost of training AI models.  

Q4. Why might DeepSeek’s open-source approach create challenges for the company?

A. It allows competitors to replicate and build on its methods.  

B. It limits the company’s ability to navigate political sensitivities.  

C. It increases the cost of training AI models.  

D. It restricts access to advanced AI chips.  

Q5. What is the main idea of the passage?

A. Western AI models dominate due to technological superiority.  

B. Chinese AI firms are struggling to compete in the global market.  

C. Open-source models are inferior to closed-source alternatives.  

D. DeepSeek’s V3 model showcases innovation despite hardware limitations.  

Part III. KEY

Q1. A.【解析】细节题。根据“The U.S. ban on sales of advanced artificial intelligence (AI) computer chips to Chinese companies has prompted innovation among Chinese firms.”，可知美国禁止向中国公司出售先进人工智能计算芯片，促使中国企业司进行创新。因此，正确答案为A。

Q2. B.【解析】细节题。根据“According to a December 2024 technical report, V3 outperforms other open-source models and offers comparable results to top-tier alternatives.”，可知根据2024年12月的技术报告，V3在性能上优于其他开源模型，并提供与顶级同类模型相当的结果。因此，正确答案为B。

Q3. C.【解析】细节题。根据“DeepSeek’s success partly lies in its use of a Mixture of Experts architecture, which reduces computing power requirements by training only specific parts of the model for particular tasks.”，可知DeepSeek的成功部分归功于其使用的专家混合架构，这种架构通过针对特定任务仅训练模型的特定部分，来减少计算能力需求。因此，正确答案为C。

Q4. A.【解析】推理题。根据“DeepSeek faces challenges. Its open-source approach allows competitors to build on its methods, and limited access to advanced AI chips could hinder future progress.”，可知DeepSeek面临挑战。其开源方法允许竞争对手利用其方法进行开发，而对先进AI芯片的有限获取可能会阻碍未来的发展。因此，正确答案为A。

Q5. D.【解析】主旨题。根据全文内容，尤其是“a remarkable display of innovation despite resource constraints”和“DeepSeek’s achievements highlight the potential of ingenuity and innovation in overcoming technological constraints, marking a significant step forward for Chinese AI development.”，可知在硬件资源受限的背景下，DeepSeek的成就突出了在克服技术限制方面的创造力和创新潜力，这标志着中国人工智能发展的重要一步。因此，正确答案为D。

（本文图片来源于摄图网，版权归摄图网所有)