近日，DeepSeek的横空出世在全球科技领域引发了广泛关注和深度思考。作为人工智能（AI）发展史上的标志性事件，DeepSeek不仅打破了“高算力和高投入是发展AI唯一途径”的迷信，还引领了AI行业进入以算法和模型架构优化为主的新时代。中国科学院院士李国杰在《科技导报》2025年第3期发表文章《DeepSeek引发的AI发展路径思考》，深入剖析了DeepSeek背后的技术突破及其对全球AI发展路径的影响。以下是李国杰院士对DeepSeek提出的七个关键问题的思考。
1.为什么DeepSeek会引起全球性的科技震撼？
DeepSeek的崛起无疑是AI领域的里程碑事件。其用户数量在7天内突破1亿，创造了用户增长速度的世界纪录。与此同时，芯片巨头英伟达（NVIDIA）的股价单日暴跌17%，市值缩水5890亿美元，创下美国上市公司单日最大损失纪录。这一现象表明，DeepSeek的成功不仅挑战了“高算力=AI霸权”的传统观念，还标志着中国科技公司从“追赶者”转变为“规则改写者”。
DeepSeek通过算法和模型架构的优化，显著降低了训练成本，并提升了模型运行效率。其V3和R1模型采用了混合专家模型（MoE）架构和多头潜在注意力机制（MLA），在降低显存占用的同时，大幅提升了推理效率。此外，DeepSeek的开源商业模式和低成本推理模型，吸引了全球众多科技巨头的青睐，包括微软、亚马逊云科技（AWS）和英伟达等公司纷纷在其AI服务平台上部署DeepSeek模型。
DeepSeek的成功揭示了AI发展的新路径：不再依赖单纯的高算力堆砌，而是通过算法优化和系统级创新实现技术突破。这一突破不仅震撼了全球科技界，也为中国在AI领域的崛起提供了有力证明。
2.“规模法则（Scaling Law）”是否已遇天花板？
规模法则（Scaling Law）是OpenAI等公司提出的AI发展理念，认为通过增加模型规模、数据量和计算资源，可以显著提升模型性能。然而，李国杰院士指出，规模法则并非科学定律，而是对技术发展趋势的猜想。近年来，大模型训练的实际效果表明，模型性能的线性增长需要模型规模、数据量和算力投入的高指数性增长，这种模式显然不可持续。
DeepSeek的出现迫使AI界重新思考技术发展路线：是继续追求高算力，还是转向算法优化和系统级创新？李国杰院士认为，DeepSeek的成功标志着AI发展从“大力出奇迹”的外延式阶段，转向集约化系统优化的内涵式阶段。虽然算力在AI发展中仍具有重要作用，但绿色发展和高能效计算将成为未来的主要方向。
3.发展“通用人工智能”（AGI）应选择什么道路？
通用人工智能（AGI）是AI领域的终极目标，但其定义和发展路径尚未形成广泛共识。OpenAI走的是“由通到专”的道路，即先开发通用基础模型，再“蒸馏”出行业垂直模型。而DeepSeek则选择了“由专到通”的路径，通过混合专家模型（MoE）等技术，集小智为大智，探索在受限资源下实现通用人工智能的可能性。
李国杰院士指出，通用与专用的竞争是技术发展的普遍现象。未来，AI发展可能会走向“通专融合”，即通用大模型的横向扩展与专用模型的纵向做精相结合，共同构建智能时代的产业新生态。
4.发展AI应该追求高算力还是高算效（高能效）？
高算力是否是AI的本质要求？李国杰院士认为，AI的初始动机是模拟人脑，而人脑的计算效率和能效极高，功耗仅约20W。相比之下，当前计算机的高能耗主要源于软硬件分离的数字计算模式。未来，AI发展应追求高算效和高能效，采用存算一体的模拟计算方式，降低能耗并提升计算效率。
DeepSeek的成功证明了低成本AI的巨大潜力。例如，斯坦福大学团队以不到50美元的云计算费用，成功训练出性能媲美高端模型的s1模型。这表明，AI的低成本化仍有巨大提升空间，而低成本是技术普及的基本前提。
5.“开源”为什么有这么大的威力？
DeepSeek的开源模式是其成功的关键之一。过去，开源大模型的性能始终与闭源模型存在差距，而DeepSeek的性能追上了闭源模型，增强了开源社区的信心。李国杰院士指出，开源模型不仅破解了企业数据隐私的难题，还实现了技术的民主化，让每个开发者都能轻松调用强大的AI工具。
开源模型的崛起改变了AI发展模式，使其不再受大公司的约束。DeepSeek的开源战略证明，在这场AI竞赛中，谁拥抱开源，谁就能赢得未来。
6.中国是否已具有在AI上引领全球的实力？
DeepSeek的成功标志着中国AI产业从“技术跟跑”向“技术并跑和领跑”迈进。尽管中国在AI基础研究和核心技术上与美国仍有差距，但近年来中国在AI领域的进步速度令人瞩目。根据《日本经济新闻》的统计，中国作者在机器学习顶级会议上的论文数量在过去4年增长了8倍。
李国杰院士认为，AI产业本质上是拼智力的新兴产业，具有明显的不对称性。DeepSeek的成功证明，中国已有一批创新型小企业进入世界前列，展现出引领全球的实力。
7.中国实现AI自立自强要如何发力？
实现AI自立自强，不仅需要国家的顶层规划和资金支持，还需要构建自主可控的产业生态。李国杰院士指出，英伟达的“护城河”并非GPU芯片本身，而是其CUDA软件生态。中国需要开发一套比CUDA更优秀的自主可控的AI软件工具系统，重构AI软件生态。
此外，中国AI投资市场的萎缩值得警醒。2023年，美国AI投资达到672亿美元，是中国AI投资的8.7倍。政府和资本界应合力构建健康的科创金融生态，为创新提供动力。
结语
DeepSeek的崛起不仅是技术突破，更是AI发展路径的深刻变革。李国杰院士的“七问”深入剖析了DeepSeek背后的技术逻辑和产业影响，为中国乃至全球AI的未来发展提供了重要启示。随着AI技术的不断演进，中国有望在AI领域实现从“跟跑”到“领跑”的跨越，为全球科技发展贡献更多中国智慧。
作者简介：梁世杰 中医高年资主治医师，本科学历，从事中医临床工作24年，积累了较丰富的临床经验。师从首都医科大学附属北京中医院肝病科主任医师、著名老中医陈勇，侍诊多载，深得器重，尽得真传！擅用“商汤经方分类疗法”、专病专方结合“焦树德学术思想”“关幼波十纲辨证”学术思想治疗疑难杂症为特色。现任北京树德堂中医研究院研究员，北京中医药薪火传承新3+3工程—焦树德门人（陈勇）传承工作站研究员，国际易联易学与养生专委会常务理事，中国中医药研究促进会焦树德学术传承专业委员会委员，中国药文化研究会中医药慢病防治分会首批癌症领域入库专家。荣获2020年中国中医药研究促进会仲景医学分会举办的第八届医圣仲景南阳论坛“经方名医”荣誉称号。2023年首届京津冀“扁鹊杯”燕赵医学研究主题征文优秀奖获得者。事迹入选《当代科学家》杂志、《中华英才》杂志。
Academician Li Guojie's "Seven Questions" to DeepSeek: Deep Thoughts on the Development Path of AI
Recently, the emergence of DeepSeek has triggered widespread attention and deep thinking in the global technology field. As a landmark event in the history of artificial intelligence (AI) development, DeepSeek not only broke the superstition that "high computing power and high investment are the only way to develop AI", but also led the AI industry into a new era dominated by algorithm and model architecture optimization. Academician Li Guojie of the Chinese Academy of Sciences published an article titled "DeepSeekinduced Thinking on AI Development Path" in the 3rd issue of "Science and Technology Guide" in 2025, which deeply analyzed the technological breakthrough behind DeepSeek and its impact on the global AI development path. Here are Richard Li's thoughts on the seven key questions raised by DeepSeek.
1. Why is DeepSeek causing a global tech shock?
The rise of DeepSeek is undoubtedly a milestone event in the field of AI. Its users hit 100 million in seven days, setting a world record for user growth. At the same time, chip giant NVIDIA (NVIDIA) shares fell 17% in one day, the market value of 589 billion U. S.dollars, the largest singleday loss on record for a listed company.This phenomenon shows that DeepSeek's success not only challenges the traditional concept of "high computing power = AI hegemony", but also marks the transformation of Chinese technology companies from "catchupers" to "rulechangers".
DeepSeek has significantly reduced training costs and improved model efficiency through optimization of algorithms and model architectures. Its V3 and R1 models adopt a hybrid expert model (MoE) architecture and multihead potential inference mechanism (MLA), which greatly improves inference efficiency while reducing memory footprint. In addition, DeepSeek's open source business model and lowcost inference model have attracted the attention of many global technology giants, including Microsoft, Amazon Web Services (AWS) and Nvidia, which have deployed DeepSeek models on their AI service platforms.
The success of DeepSeek reveals a new path for the development of AI: no longer relying on the simple accumulation of high computing power, but achieving technological breakthroughs through algorithm optimization and systemlevel innovation. This breakthrough not only shocked the global technology community, but also provided strong evidence for China's rise in the field of AI.
2. Has the Scaling Law Met the Ceiling?
The Scaling Law is an AI development concept proposed by companies such as OpenAI, which believes that by increasing the size of the model, the amount of data and computing resources, the performance of the model can be significantly improved. However, academician Li Guojie pointed out that the law of scale is not a law of science, but rather an assumption of technological development trends. In recent years, the practical effects of large model training have shown that linear growth in model performance requires a high exponential increase in model size, data volume, and computational investment, a pattern that is clearly unsustainable.
The emergence of DeepSeek has forced the AI community to rethink the development path of technology: should it continue to pursue high computing power or shift to algorithm optimization and systemlevel innovation? Academician li Guojie believes that the success of DeepSeek marks a shift in the development of AI from the outwardoriented phase of "great efforts leading to miracles" to the inwardoriented phase of intensive system optimization. Although computing power still plays an important role in AI development, green development and highefficiency computing will become the main direction in the future.
3. What path should be taken to develop "General Human Intelligence" (AGI)?
Artificial general intelligence (AGI) is the ultimate goal in the field of AI, but there is still no broad consensus on its definition and development path, OpenAI takes the road of "from general to specialized," that is, first developing a general basic model, and then "distilling" the industry vertical model. DeepSeek, on the other hand, has chosen the path of "from specialized to general," through technologies such as Hybrid Expert Model (MoE) to explore the possibility of realizing general AI under limited resources.
Academician Li Guojie pointed out that competition between generics and specialties is a universal phenomenon of technological development. In the future, the development of AI may move towards "the integration of general and special", that is, the horizontal expansion of general large models and the vertical refinement of special models are combined to jointly build a new industrial ecosystem of the intelligent era.
4. Should AI development pursue high computing power or high computing efficiency (high energy efficiency)?
Is high computing power the essential requirement for AI? Academician li Guojie believes that the initial motivation of AI is to simulate the human brain, and the human brain has extremely high calculation efficiency and energy efficiency, with power consumption of only about 20W. In contrast, the high energy consumption of current computers is mainly due to the digital computing model that separates hardware and software. In the future, the development of AI should pursue high computational efficiency and high energy efficiency, adopt the simulation computing mode of storage and computation integration, reduce energy consumption, and improve computing efficiency.
The success of DeepSeek proves the great potential of lowcost AI. The Stanford team, for example, has managed to train an S1 model that performs as well as highend models for less than $50 in cloud computing. This shows that there is still great room for improvement in the lowcost of AI, and low cost is the basic premise for the popularization of technology.
5. Why is Open Source so powerful?
DeepSeek's open source model is one of the keys to its success. In the past, the performance of the open source model has always been a gap with the closedsource model, and the performance of DeepSeek has caught up with the closedsource model, enhancing the confidence of the open source community. Academician li Guojie pointed out that the open source model not only cracked the problem of enterprise data privacy, but also realized the democratization of technology, so that every developer can easily call powerful AI tools.
The rise of opensource models has changed the development model of AI, making it no longer bound by large companies. DeepSeek's open source strategy proves that in this AI competition, who embraces open source, who can win the future.
6. Does China have the strength to lead the world in AI?
The success of DeepSeek marks the progress of China's AI industry from "technology following" to "technology following and leading". Although China still has a gap with the United States in AI basic research and core technology, China's progress in the field of AI in recent years has been remarkable. According to the Japan Economic Journal, the number of papers presented by Chinese authors at top machine learning conferences has increased eightfold in the past four years.
Academician li Guojie believes that the AI industry is essentially an emerging industry that relies on intelligence and has obvious asymmetry. The success of DeepSeek proves that China has a number of innovative small enterprises to enter the forefront of the world, showing the strength of leading the world.
7. How should China work hard to achieve AI selfreliance?
Realizing AI selfreliance requires not only the toplevel planning and financial support of the country, but also the construction of a controllable and independent industrial ecology. Li pointed out that the "moat" of Yingweida is not the GPU chip itself, but its CUDA software ecosystem. China needs to develop an AI software tool system that is more independent and controllable than CUDA, and reconstruct the AI software ecology.
In addition, the shrinkage of China's AI investment market is worth warning. In 2023, AI investment in the United States reached 67.2 billion US dollars, which is 8.7 times that of China's AI investment. The government and the capital community should work together to build a healthy science and innovation financial ecosystem to provide impetus for innovation.
Conclusion
The rise of DeepSeek is not only a technological breakthrough, but also a profound change in the development path of AI. Academician Li Guojie's "Seven Questions" deeply analyzed the technical logic and industrial impact behind DeepSeek, providing important enlightenment for the future development of AI in China and even the world. With the continuous evolution of AI technology, China is expected to make the leap from "running behind" to "leading the way" in the field of AI and contribute more Chinese wisdom to global scientific and technological development.
Author Bio: Liang Shijie is a senior medical practitioner in traditional Chinese medicine with an undergraduate degree. He has been engaged in traditional medicine clinical work for 24 years and has accumulated a wealth of clinical experience. Following Chen Yong, chief physician of liver disease at Beijing Traditional Medicine Hospital, affiliated with Capital Medical University, and renowned old Chinese medicine, he has been treated for many years and received great attention. He specializes in the treatment of difficult diseases using "conversational traditional therapy" and special treatments combined with the academic ideas of Jiao Shude and Guan Yubo's ten-level diagnosis.He is currently a researcher at the Shude Tang TCM Research Institute in Beijing, a fellow at the new 3 + 3 project of traditional Chinese medicine flame inheritance in Beijing - a scholar at the inheritance workstation of Jiao Shude's protégés (Chen Yong),He is a standing committee member of the International Expert Committee on E-learning and Health Care, a member of the Jiao Shude Academic Heritage Special Committee of the Chinese Association for the Advancement of Chinese Medicine Research, and the first cancer specialist to be included in the chapter of the Chinese Pharmaceutical Culture Research Association. Won the 2020 China Association for the Promotion of Traditional Chinese Medicine Zhongjing Medical Branch held the eighth session of the Medical Saint Zhongjing Nanyang Forum "Classic Prescription Famous Doctor" honorary title. The winner of the first Beijing-Tianjin-Hebei "Pingui Cup" Yanzhao Medical Research Essay Award in 2023. His work was featured in the journal Current Scientist and the journal Chinese Talent.
