The first batch of government and enterprises to use DeepSeek are even more anxious
The first batch of government and enterprises to use DeepSeek are even more anxious
Once rang the bell
·February 27, 2025 19:25
DeepSeek's' cold thinking 'after the fire: How difficult is the last mile for AI to land in government and enterprises
Today, in the AI craze sparked by DeepSeek, everyone is excited and enthusiastic about the powerful and intelligent AI experience. There has also been a wave of integration on the enterprise side, with many exploring the application of DeepSeek.
However, most people may overlook that despite DeepSeek's high cost-effectiveness and open source free nature, it has greatly lowered the threshold for AI usage and reduced the cost burden on enterprise organizations. But this path of application is actually not so easy to take.
When DeepSeek is deeply applied to the core production systems of enterprises and various aspects within the organization, a large number of questions will erupt: is one DeepSeek enough? Will there be more powerful models in the future? How will it be replaced or integrated at that time? How can existing data be fed into DeepSeek, or even the next stronger model? wait.
These issues are troubling CTOs, especially after corporate organizations have started integrating with DeepSeek, they are even more anxious.
The common pain of enterprise AI application from laboratory to production workshop
At present, several leading enterprises in various industries such as government affairs, finance, electricity, coal mining, and manufacturing have publicly announced the introduction of DeepSeek to carry out intelligent applications. According to statistics, currently over 40 out of 98 state-owned enterprises have launched DeepSeek applications.
Looking at specific application scenarios, most of them are still focused on general scenarios such as customer service, office, and research and development, and some top enterprises have begun to expand into core scenarios. How to fully empower the entire production process of enterprises with DeepSeek and enable big models to truly "go to work" is the most concerning issue for everyone next.
1、 Multi model collaboration is inevitable, and more innovative models will emerge in the future.
The human world is a complex and diverse world, with complex and ever-changing scenarios and problems. A single model cannot meet the needs of all scenarios. For example, a typical manufacturing enterprise has both intelligent Q&A and knowledge base scenarios in the sales and service process, as well as scenarios for factory intrusion, accurate warning of abnormal events, and optimization of production processes in the production process. All of these require natural language models, machine vision models, and predictive models to complete. DeepSeek performs well in natural language, scientific computing, and other fields, but it is not omnipotent and cannot cover all areas of AI applications.
While most government and enterprises are integrating with DeepSeek, the models previously deployed and used will still be used in parallel, especially machine vision and prediction, which are widely used in the industrial field, will still play an important role. In the long run, the iterative updates of model technology will become faster and faster, which means that there will definitely be new models or training paradigms emerging in the future.
Therefore, the coexistence and collaborative development of multiple models is an inevitable choice for the AI industry to meet diversified needs, and will not be replaced by the emergence of a specific model.
2、 The surge in computing power: The demand for computing power is increasing instead of decreasing, and the evolution of technology requires strong infrastructure.
The development of AI technology, especially the continuous evolution of deep learning models, has led to an exponential increase in the demand for computing power. From early simple neural networks to today's complex large-scale pre trained models, every technological breakthrough relies on powerful computing power support.
Although DeepSeek has reduced training costs by 90% through algorithm optimization, it has also lowered the threshold for computing power, stimulating more industries to embrace AI. Based on long-term development, both the popularization of AI applications and the iterative updating of models cannot be separated from the continuous improvement of computing power.
Therefore, the growth trend of computing power demand is unstoppable, which is an important material foundation for promoting the development of the AI industry and will not change with the emergence of individual technologies or products.
3、 Application is king: delve into industry scenarios with systematic thinking.
Every industry has its unique knowledge system and operational logic, and AI can only truly unleash its enormous potential by continuously delving into it, optimizing and customizing it. Therefore, in the implementation of AI applications in government and enterprises, it is necessary to deeply integrate general models and industry knowledge to maximize the value of AI technology. For example, in the financial field, while leading AI technology is important, it is also necessary to combine analysis of customer transaction data, credit records, social networks, and other multi-source data to build risk assessment models, in order to accurately determine customer credit risk, decide whether to approve loans, loan amounts, interest rates, etc., which are currently beyond the capabilities of general models.
In this process, a scientific methodology is required, which is a systematic engineering. DeepSeek provides enterprises with a new option for large-scale models, but the key steps and principles in the implementation process will not change. This includes multiple stages such as precise requirement sorting, efficient data preparation, model selection and customization, deployment, operation and maintenance, and integration with existing business processes and systems of the enterprise.
Firstly, enterprises need to clarify their business objectives and identify which business processes can be optimized or innovated through the use of large-scale models. Secondly, when deploying the model, it is necessary to consider the enterprise's IT infrastructure, data security, and privacy protection requirements, and choose an appropriate deployment method, such as private cloud, hybrid cloud, or edge deployment. Furthermore, high-quality industry data is needed to train and fine tune large models to match the actual scenarios of the enterprise. After the model is put into use, its performance needs to be continuously monitored, and timely adjustments and optimizations should be made based on business changes and user feedback. This entire methodology is a guarantee to ensure that big models can play a role in enterprises and achieve sustainable development, and will not be overturned by the emergence of new models.
Overall, the difficulties and fundamentals of AI application in government and enterprises have not changed, which determines the direction of AI application in government and enterprises. We must firmly grasp the basic situation, actively layout according to our own development needs and industry characteristics, and achieve stable and sustainable development.
Connecting the Last Mile: A Sustainable Evolution Platform is the Core
In fact, the hotspots of big models have been constantly changing. From ChatGPT to DeepSeek, in just 2 years, the market's C-position has already been replaced. In the development path of technology, change is inevitable, and uncertainty is also inevitable. For large government and enterprises, building an AI architecture with sustainable evolution capabilities is a necessary condition for enterprises to achieve stable and sustainable development in the field of AI as a major foundation for coping with future changes.
1、 A stable AI development platform. The AI platform is a bridge connecting the lower level software and hardware infrastructure with the upper level large model, and its technical architecture needs to be convergent, simple, and unified. The downward trend of AI platforms requires the integration of cloud platforms to encapsulate the complexity of underlying software and hardware, and to solve computational efficiency through elastic resource scheduling, ensuring the scheduling and operation of computing power, models, and various resources; Upward, it is necessary to support rapid adaptation of models from diverse sources, and provide a series of toolchains to support one-stop development and deployment of models, data, and applications. Through fine-tuning, evaluation, compression, and end-to-end standardization of deployment, models can be launched faster.
In addition, AI platforms also need to flexibly choose various deployment modes such as public cloud, private cloud, or a combination of both according to the needs of enterprises in different periods and scenarios. Meanwhile, for scenarios with edge business requirements, it is necessary to plan the architecture of cloud edge collaboration in advance.
Hybrid cloud is a more suitable solution. Taking Huawei Cloud Stack, a hybrid cloud solution provided by Huawei Cloud, as an example, various deployment modes such as Ascend Cloud Services, Full Stack Hybrid Cloud, and Edge Lightweight can be provided for government and enterprise customers to flexibly choose from based on their different AI needs at different stages. In the early stages of exploring applications, government and enterprise customers can quickly pilot and try out applications using Huawei Cloud Ascend Cloud Services with just one click. In the deep application stage, government and enterprise customers can also choose to push the models trained on Ascend Cloud to the local central cloud. Based on the full stack cloud service provided by Huawei Cloud Stack, combined with customers' local private data, the models can be fine tuned and trained to train specialized models that better match their own business needs.
In this way, government and enterprise clients do not need to migrate platforms or restructure architectures, and can smoothly complete AI application deployment and experimentation at different stages, efficiently achieving the digital transformation and upgrading of enterprises.
2、 Standardized implementation paradigm. The evolution of architecture also comes from its understanding and implementation of AI landing paradigms. Often, a platform with standardized landing paradigms can solve various problems related to the landing of government and enterprise AI applications in an orderly manner, which often requires a series of development tools to support implementation.
For example, for data development, a powerful set of data development tools with efficient data collection, cleaning, and preprocessing functions can effectively improve data quality. Based on these high-quality data, DeepSeek models can be trained and fine tuned to adapt to enterprise scenarios.
In the model development process, a series of advanced tools are used to support the design, training, and optimization of the model. Including model training, fine-tuning, edge deployment, and conducting quantitative compression and model evaluation to help developers understand the strengths and weaknesses of the model and make targeted optimizations.
There is also an application development process where efficient and user-friendly application development frameworks can help users quickly integrate trained models into various application systems, including Prompt templates, pre built plugins, RAG, etc., achieving minute level AI application creation. At the same time, the visual interface design function can lower the development threshold, improve development efficiency, and enable non professional developers to quickly build fully functional AI applications.
3、 Accumulation of talent and experience. The evolution of architecture not only relies on technical capabilities, but also on the accumulation of talent and experience. This is a seemingly abstract ability, but in reality it still has traceable features. With the acceleration of AI applications in government and enterprises, these capabilities are becoming an irreplaceable asset of the platform.
By establishing a comprehensive experience accumulation mechanism, these experiences can be organized, summarized, and shared. For example, by establishing an internal knowledge base, regularly organizing experience sharing meetings and technical exchange activities, structured management of internal experience and knowledge can enable enterprises to quickly learn from past successful experiences and avoid detours when facing new large-scale model goals.
At the same time, develop a comprehensive AI talent training plan to attract and cultivate a group of high-quality AI professionals. Clarify the recruitment standards, training plans, growth channels, etc. for AI talents. At the same time, enterprises should create a good innovation atmosphere, encourage employees to try new technologies and methods, provide practical and innovative platforms for employees, stimulate their creativity and potential, and build a strong competitive AI talent team.
conclusion
Today, the AI craze sparked by DeepSeek is redefining the opportunities and challenges for government and enterprise organizations in the field of AI. Faced with the constantly growing demand for computing power and complex and ever-changing business scenarios, enterprises not only need to keep up with the pace of the times and actively embrace new models and technologies, but also need to take a long-term perspective, carefully select the correct AI architecture and evolution path based on their own actual situation and development needs.
This is not just a process of technology selection, but also a test of the company's strategic vision and execution ability. Only by selecting a stable, reliable, and flexible AI platform can we ensure the success of government and enterprise AI in key areas such as large-scale deployment, application development, data engineering, model training, and fine-tuning, and respond to the constantly changing future with a long-term stable attitude.
*The images in this article are all sourced from the internet
This article is from the WeChat official account "Ringing Talk". The author is Zeng Ringing, 36 Krypton has been authorized to release it.
The viewpoint of this article only represents the author himself, and the 36Kr platform only provides information storage space services.
第一批用上DeepSeek的政企，更“焦虑”了
今天，在DeepSeek掀起的AI热潮中，人人都在为强大且智能的AI体验而兴奋、狂热。企业侧也掀起了一场接入热潮，纷纷开展DeepSeek的应用探索。
然而，可能大多数人都忽略了——尽管DeepSeek的高性价比和开源免费，已经大大降低了AI的使用门槛，也减轻了企业组织的成本负担。但这条应用之路走起来其实并没有那么轻快。
当DeepSeek深度应用到企业的核心生产系统以及组织内部的方方面面，大量的问题将会爆发出来：是否一个DeepSeek就足够了？未来是否会出现更强大的模型？届时如何替换或整合？现有的数据如何投喂给DeepSeek，甚至是下一个更强的模型？等等。
这些问题困扰着CTO们，特别是企业组织已经开始接入DeepSeek之后，他们更焦虑了。
从实验室到生产车间，企业AI应用共同的“痛”
目前，政务、金融、电力、煤矿、制造等多个行业的多家头部企业已公开宣称引入DeepSeek开展智能化应用。据统计，目前98家央企中已有超过40家开展DeepSeek应用。
翻看具体的应用场景，大多数都还集中在客服、办公、研发等通用场景，有部分头部企业开始拓展到核心场景。如何让DeepSeek全面地赋能企业生产全流程，让大模型真正“上岗”是大家接下来最关心的问题。
一、百模千态：多模型协同是必然，未来也会出现更创新的模型。
人类世界是一个复杂多元的世界，场景和问题也是复杂多变的，单一的模型无法满足所有场景的需求。例如，典型的制造企业，既有销服环节的智能问答、知识库场景，又有生产环节的工厂的入侵、异常事件准确预警和生产工艺进行优化场景，这些都需要自然语言模型、机器视觉模型和预测模型才能完成。DeepSeek 在自然语言、科学计算等方面表现出色，但它不是万能的，更无法覆盖AI 应用的全部领域。
大部分政企在接入DeepSeek的同时，之前已部署使用的模型也会依然并行使用，尤其是目前工业领域应用比较多的机器视觉、预测等依然会发挥重要作用。从长远来看，模型技术的迭代更新会越来越快，这意味着未来一定会有新的模型或训练范式出现。
因此，多模型并存、协同发展是 AI 产业满足多样化需求的必然选择，不会因某一特定模型的出现而被替代。
二、算力激增：算力需求不减反增，技术的演进需要强大的基础设施。
AI 技术的发展，尤其是深度学习模型的不断演进，对算力的需求呈指数级增长。从早期简单的神经网络到如今复杂的大规模预训练模型，每一次技术突破都依赖于强大的算力支持。
尽管DeepSeek通过算法优化将训练成本降低90%，但是也降低了算力门槛，刺激了更多行业拥抱AI。基于长期发展来看，不论是AI应用普及还是模型迭代更新都离不开算力的持续提升。
因此，算力需求的增长趋势不可阻挡，这是推动 AI 产业发展的重要物质基础，不会因个别技术或产品的出现而改变。
三、应用为王：深入行业场景，以系统化思维。
每个行业都有其独特的知识体系和运作逻辑，AI 只有持续深入其中，不断优化和定制，才能真正发挥其巨大潜力。因此，在政企AI应用落地中，需要将通用模型和行业知识深度融合共同发挥作用，才能最大化发挥AI技术的价值。例如，在金融领域，AI技术的领先固然重要，但还需要结合分析客户的交易数据、信用记录、社交网络等多源数据，构建风险评估模型，才能精准判断客户的信用风险，决定是否批准贷款以及贷款额度、利率等，而这些都是目前的通用模型所做不到的。
在这个过程中，需要一套科学的方法论，是一个系统工程。DeepSeek 为企业提供了一种新的大模型选择，但落地过程中的关键步骤和原则不会改变。包括从需求的精准梳理、数据的高效准备、模型的选型与定制，到部署、运维以及与企业现有业务流程和系统的融合等多个环节。
首先，企业要明确自身业务目标，确定哪些业务环节可以借助大模型实现优化或创新。其次，在模型部署时，要考虑企业的 IT 基础设施、数据安全和隐私保护要求，选择合适的部署方式，如私有云、混合云或边缘部署。再者，需要高质量的行业数据，对大模型进行增训、微调，以匹配企业的实际场景。模型投入使用后，需要持续监测其性能，根据业务变化和用户反馈进行及时调整和优化。这一整套方法论是确保大模型在企业中发挥价值、实现可持续发展的保障，不会因新模型的出现而被颠覆。
综合来说，政企AI应用的难点和基本面并未改变，这也就决定了政企AI应用方向，牢牢把握基本盘，结合自身发展诉求和行业特性积极布局，实现稳健且可持续发展。
打通最后一公里：可持续演进的平台是核心
事实上，大模型的热点一直在变。从ChatGPT到DeepSeek，中间只是2年的时间，市场的C位就已经完成替换。在技术的发展路径中，变化是必然的，不确定性也是必然的。而对于大型政企而言，构建一个具备可持续演进能力的AI架构，作为应对未来变化的一大基本盘，是企业在 AI 领域实现稳健且可持续发展的必要条件。
一、稳定的AI开发平台。AI平台是下层的软硬件基础设施和上层的大模型之间联接的桥梁，其技术架构要收敛，要简单，要统一。AI平台向下需要结合云平台封装底层软硬件的复杂性，通过弹性资源调度解决算力效率，做好算力、模型和各类资源的调度和运营；向上需要支持来源多样的模型的快速适配，并且提供一系列工具链支持模型、数据、应用的一站式开发和部署，通过微调、评测、压缩、部署端到端标准化，让模型上线更快。
此外，AI平台还需要根据企业在不同时期和场景的需要，灵活选择公有云、私有云或是两者结合的多种部署模式。同时，对于有边缘业务需求的场景，需要提前规划云边协同的架构。
混合云是更为匹配的一种方案。以华为云提供的混合云方案华为云Stack为例，针对政企客户在不同阶段的不同AI需求，可以提供昇腾云服务、全栈混合云、边缘轻量化等多种部署模式供政企客户灵活选择。在探索应用初期，政企客户可以一键式使用华为云昇腾云服务快速开展应用试点尝鲜。等到了深入应用阶段，政企客户还可以有选择地将昇腾云上训练好的模型推送到本地中心云，基于华为云Stack提供的全栈云服务，结合客户本地私有数据进行模型的微调和增训，从而训练出更匹配自身业务诉求的专精模型。
如此一来，政企客户完全不需要迁移平台或是重构架构，就能顺畅地完成不同阶段的AI应用部署与尝试，高效实现企业的数智化转型升级。
二、标准化的落地范式。架构可演进还来自其对AI落地范式的理解与落实，往往一个平台越是具备标准化的落地范式越能有条不紊地解决好政企AI应用落地的各种问题，而这往往需要一系列的开发工具来支撑实现。
例如针对数据开发，一套强大的数据开发工具，具备高效数据采集、清洗、预处理功能，能够有效地提高数据质量。基于这些高质量数据，才能对DeepSeek模型进行增训和微调，适应企业场景。
在模型开发环节，通过一系列先进的工具来支持模型的设计、训练和优化。包括模型的增训、微调、边缘部署，并开展量化压缩和模型评估，帮助开发者了解模型的优缺点，进而进行针对性的优化。
还有应用开发环节，高效好用的应用开发框架可以帮助用户快速将训练好的模型集成到各种应用系统中，包括Prompt模板、预制插件、RAG等，实现分钟级AI应用创建。同时，可视化的界面设计功能，能够降低开发门槛，提高开发效率，使得非专业的开发人员也能够快速搭建出功能完善的 AI 应用。
三、人才和经验的积累。架构的可演进除了技术能力外，还离不开其在人才和经验方面的沉淀。这是一种看似抽象，但实际依然有迹可循的能力。随着政企AI应用进程加速，这些能力正转化成为平台的不可替代的财富。
通过建立完善的经验沉淀机制，将这些经验进行整理、归纳和分享。例如，通过建立内部知识库，定期组织经验分享会和技术交流活动，对企业内部的经验知识进行结构化管理，可以使得企业在面对新的大模型目时，能够快速借鉴以往的成功经验，少走弯路。
同时制定全面的 AI 人才培养计划，吸引和培养一批高素质的 AI 专业人才。明确 AI 人才的招聘标准、培养计划、成长通道等，同时，企业要营造良好的创新氛围，鼓励员工勇于尝试新技术、新方法，为员工提供实践和创新的平台，激发员工的创造力和潜力，打造一支具有强大竞争力的 AI 人才队伍。
结语
今天，DeepSeek引发的AI热潮，正在重新定义政企组织面向AI领域的机会与挑战。面对不断增长的算力需求和复杂多变的业务场景，企业不仅需要紧跟时代步伐，积极拥抱新模型、新技术，更要立足长远，从自身的实际情况和发展诉求出发，精心挑选正确的AI架构与演进路径。
这不仅仅是一次技术选型的过程，更是对企业战略眼光和执行能力的考验。只有选对一个稳定、可靠且具备灵活演进能力的AI平台，才能确保政企AI在大规模部署、应用开发、数据工程以及模型增训、微调等关键环节上取得成功，并以长期稳定的姿态应对持续变化的未来。
*本文图片均来源于网络
本文来自微信公众号“响铃说”，作者：曾响铃，36氪经授权发布。
该文观点仅代表作者本人，36氪平台仅提供信息存储空间服务。