大模型下半场将更聚焦在垂直化应用以及生态化发展。
离大模型横空出世已经过去一年有余，在AI大模型的浪潮下，各大科技企业争先恐后的推出了自家的大模型产品。与此同时，各行业企业也对大模型保持着高度关注，一些其他行业的企业也都纷纷跨界布局大模型相关产品。
如果说，各大厂商纷纷推出大模型产品形成“百模大战”的局势，是大模型这场“战役”的上半场的话，那么这场“战役”的下半场将更聚焦在大模型的垂直化应用以及生态化发展。
一、从通用大模型到行业大模型
《北京市人工智能行业大模型创新应用白皮书（2023年）》中显示，截至2023年10月，我国10亿参数规模以上的大模型厂商及高校院所共计254家，分布于20余个省市/地区。商业咨询机构爱分析的报告称，2023年中国大模型市场规模约为50亿元，预计到2024年这一数字将达到120亿元。
显然，2024年，大模型将继续其火热的现象，进一步渗透到各行各业的数字化进程中。
大模型真正的价值在于行业侧的应用落地，就目前业内对大模型的认知来看，绝大多数人对大模型相关产品的发展观点类似于互联网——消费级只是开始，产业级价值更大。但就如同互联网一样，消费互联网发展迅速，甚至已经接近“天花板”；产业互联网也仅是近年来在政策引导，数字技术驱动下，逐步发展提速。 数字化转型网（www.szhzxw.cn）
大模型的发展与互联网的发展类似。回看刚刚过去的2023年，以ChatGPT为代表的语言大模型发展迅速，除了ChatGPT、必应以外，国内众多厂商也纷纷布局，包括华为、阿里、百度、京东、科大讯飞、商汤等在内的众多科技公司也都积极布局，纷纷推出了各自的大模型产品。
不过回看这些大模型产品，大多是聚焦C端，也就是终端用户的产品，为用户提供了办公、生活上的辅助。
不过，企业对于大模型产品还是保持开放的态度，据IBM商业价值研究院最近发布的调研报告显示：有四分之三的受访CEO认为，部署先进的生成式人工智能将为企业带来竞争优势。
大模型真正的价值是：产业侧应用的落地，也就是行业级大模型产品成熟度的提高。据市场研究机构预测，到2025年，全球生成式AI市场规模将达到100亿美元以上。其中，企业级生成式AI市场将占据相当大的份额，成为最大的应用领域之一。 数字化转型网（www.szhzxw.cn）
行业大模型是指针对特定行业或领域的大模型，这种模型针对特定领域的任务进行了优化和定制。相比通用大模型，行业大模型专业性更强、性能更优。不过行业大模型在发展的过程中，与通用大模型也存在着不少差别。
目前已知的主要差别是——行业大模型不仅需要通用的语料库，还需要针对不同行业，不同场景的专业语料库。神州数码副总裁CTO李刚曾对钛媒体表示，行业大模型具有极强的专业性，需要大量行业专业知识库，“目前，这个行业知识库的语料需控制在20%，不多不少。”李刚强调，“超过20%，训练出来的大模型可能就‘不会说人话’，造成沟通障碍；少于20%，又不具备行业的专业性。”
二、医疗、法律、金融等行业率先落地
目前行业级大模型还处于发展的初级阶段，虽然有很多家企业推出了行业大模型，但是应用并不成熟。就现阶段大模型在各个行业的应用现状，以及发展趋势来看，医疗、金融、法律等行业正率先落地较成熟的行业级大模型应用。 数字化转型网（www.szhzxw.cn）
智源研究院大模型行业应用负责人周华曾对钛媒体表示，现阶段，容错性比较高的通用领域大模型成熟度较高，在类似智能客服、文档处理等方面能够发挥更多辅助作用，还有通用领域的文生图应用，以及通过检索增强技术缓解部分幻觉问题的专业领域应用，都是目前企业比较好落地的一些应用场景。
无独有偶，IEEE标准协会新标准立项委员会副主席兼IEEE数字金融与经济标准委员会主席林道庄也有着相似的看法，林道庄表示，目前，大模型的应用主要集中在“三产”（服务业）居多，重点是辅助人更快、更好地服务其他人，而行业级的大模型也有望在服务业相关领域率先落地成熟度较高的应用。
就目前的发展现状来看，行业大模型发展较快的行业主要有金融、医疗、法律等。
金融行业方面，2023年3月，彭博首度针对金融业推出大型语言模型BloombergGPT，引发市场对金融垂直领域大模型的关注；6月，哥伦比亚大学联合上海纽约大学推出FinGPT 。
在国内，同年7月，华为全新发布盘古大模型，金融行业大模型正是其中数个行业通用大模型之一；同年9月，蚂蚁集团正式发布自研“蚂蚁基础大模型”，以及在此基础上进行定制的“蚂蚁金融大模型”。
虽然众多具备大模型能力的公司都在积极布局金融行业大模型产品，不过基于金融行业的特殊性——对安全合规要求极高，大模型在金融行业具备完全成熟落地的能力仍需时日。
张劲曾对钛媒体表示，金融行业不同于其他行业，其监管要求极高，从技术上看，像贷款审核等业务其实已具备初步落地的技术能力，但因为安全合规的要求，大模型只能在其中起到解放生产力的辅助作用。
现阶段，大模型在金融行业的应用主要还是集中在风险评估和管理，以及知识图谱平台搭建方面。在风险评估方面，大模型可以通过分析大量的历史数据和实时信息，预测市场风险、评估信用风险等，为金融机构提供更加准确和及时的风险管理决策支持。
另一方面，将大模型与知识图谱平台结合，用大模型代替NLP技术，金融机构可以在提升效率的同时，提升风控水平。 数字化转型网（www.szhzxw.cn）
除了在金融行业有望在今年有较成熟的应用场景出现以外，林道庄对钛媒体表示，像医疗、法律咨询、教培、娱乐等风险要求较低、偏服务的行业，在今年有望有较多成熟的应用场景落地。“在2024年，大模型将会在一些有人辅助校验、风险及精准程度要求较低的行业落地，通过大模型辅助人类去进行工作，可以使更多人享受到更好的服务。但还远不到代替人的能力。”林道庄强调。
以医疗行业为例，通过行业大模型对大量医疗数据的学习和分析，可以自动识别病变特征，辅助医生进行疾病诊断，提高诊断准确率与诊断效率。对此，林道庄表示，现阶段，我国医疗资源比较紧缺，许多人排了很久的队才能看上病，而医生也很忙，通过大模型的辅助，能帮助医生快速的识别诸如X光片、CT等病历，从而大幅提升医生工作效率，也能降低患者等待的时间。“对于医疗、法律及翻译服务等这类信息或能力严重不对称的行业，大模型的介入将大大提高服务提供的效率，推动服务流程的标准化。”林道庄强调。
除“增效”以外，大模型在医疗行业的落地还可以帮助患者和医院实现“降本”，通过自动化和智能化的医疗辅助系统，可以降低医疗成本，提高医疗服务的效率和质量。
另一方面，行业大模型在医疗行业还可以帮助医生优化治疗方案。通过大模型对对患者病情、病史、药物过敏等情况进行综合分析，为医生提供更加全面和个性化的治疗方案，提高治疗效果。
在林道庄看来，行业大模型能率先落地的垂直行业具有一个普遍的共性——知识密集型行业，“行业大模型能率先落地的行业一定是可以通过知识密集提供价值的行业，”林道庄指出，“某种程度上讲，通过知识收集、知识管理，实现辅助人生成内容、提出决策建议，而不是代替人的目标。”
在医疗、金融等行业之外，大模型在诸如工业、制造业等行业的落地还仍需时日。对此，周华对钛媒体表示，现阶段，大模型主要的能力还是体现在文字、文档处理，一般性的聊天和较浅显的专业问答方面，以及通用视觉领域的问答和生成方面，对于逻辑推理要求和准确性要求高的专业语言领域，涉及学科和工程相关图片、视频识别的专业视觉领域，以及“文生视频”等能力仍需技术迭代，“这种具备多模态能力的大模型产品目前仍难以在行业侧的落地应用的能力，”周华如是说，“2024年，多模态模型将成为大模型领域各大厂商角逐的焦点。而随着多模态模型成熟度不断变高，会有更多的行业应用场景出现。”
三、安全依旧“不容忽视”
行业大模型在落地应用的过程中，企业也会遇见很多问题，除了大模型的“幻觉”问题以外，最大的一个问题就是安全的问题。 数字化转型网（www.szhzxw.cn）
数据安全自从互联网出现以来，一直是所有参与者共同关注的焦点问题，在AIGC时代，亦是如此。大模型在为我们带来更多便利，提高效率的同时，也带来了日益严峻的安全挑战。
当下，数据已成为企业，乃至国家重要资产，数据安全、隐私保护等问题也成为各行业企业关注的焦点，据IBM Security 发布的《2023年数据泄露成本报告》中显示，仅数据泄露一项，2023年全球数据泄露的平均成本达到 445 万美元，创该报告有史以来以来最高记录，较过去 3 年均值增长了 15%。
而随着大模型相关产品的落地，数据安全面临的威胁也势必将越来越大。IEEE调研显示，2024年将会出现其他更具威胁的网络安全问题，包括勒索软件攻击（2024年为37%，高于2023年的30%）、网络钓鱼攻击（2024年为35%，高于2023年的25%）以及内部威胁（2024年为26%，高于2023年的19%）。
以目前技术发展来看，AIGC主要从三个方面给网络安全带来更大的挑战，分别是：恶意软件/网络攻击、分布式拒绝服务攻击，以及网络监控和隐私侵犯等；另一方面，企业在AIGC还面临着数据泄露、数据篡改/伪造等数据安全方面的挑战。
在林道庄看来，虽然AI时代，企业面临着更大的网络安全和数据安全的挑战，但是在这个过程中，企业也可以充分应用AI的能力进行预警、防御，“通过AI自动学习识别，预测一些潜在的风险，并将这些风险自动隔离起来，”林道庄如是说，“通过AI的赋能，让企业级防火墙具备更强的应变能力。”
立足安全领域，其实已经有不少头部的厂商尝试将AIGC能力整合进安全产品及解决方案中，例如，云起无垠2023年底发布了名为「SecGPT」的网络安全大模型开源项目；360于2023年9月开放360智脑大模型，据了解，该模型是全国首个原生安全大模型；腾讯安全在混元大模型基础上投喂安全知识语料库二次训练出安全大模型，并基于安全大模型打造了腾讯云AI安全助手……..
确实，AIGC就好像一把“双刃剑”，一方面，带来了更多便捷，提升了效率；另一方面，基于AIGC技术的网络攻击，以及给数据安全的威胁也会越来越多。 数字化转型网（www.szhzxw.cn）
而对于几乎所有企业来说，选择应用数字技术的时候，所有一切的前提就是——安全。企业在享受AIGC带来的技术红利的过程中，对于安全的考虑也将成为AIGC发展的一个重要赛道。
翻译：
2024, how to land the highly anticipated industry model?
The second half of the large model will focus more on vertical application and ecological development.
It has been more than a year since the big model was born, and under the wave of AI big models, major technology companies have rushed to launch their own big model products. At the same time, enterprises in various industries also maintain a high level of attention to large models, and some enterprises in other industries have also cross-border layout of large model-related products.
If the major manufacturers have launched large model products to form a “hundred model war” situation, is the first half of the “battle” of the large model, then the second half of the “battle” will be more focused on the vertical application of the large model and ecological development.
First, from the general model to the industry model
According to the “Beijing Artificial Intelligence Industry Large model Innovation Application White Paper (2023)”, as of October 2023, there are 254 large model manufacturers and colleges and universities with a scale of more than 1 billion parameters in China, distributed in more than 20 provinces and regions. The report by business consultancy Analytica said that the size of China’s large model market will be about 5 billion yuan in 2023, and the figure is expected to reach 12 billion yuan by 2024.
It is clear that in 2024, large models will continue their hot phenomenon and further penetrate into the digitization process of various industries.
The real value of the big model lies in the application of the industry side, as far as the industry’s cognition of the big model is concerned, the vast majority of people’s development views on the big model-related products are similar to the Internet – consumer grade is only the beginning, and the industrial grade value is greater. But just like the Internet, the consumer Internet is developing rapidly and even approaching the “ceiling”; The industrial Internet has only gradually accelerated in recent years under the guidance of policies and driven by digital technology.
The development of big models is similar to the development of the Internet. Looking back at the past 2023, the language large model represented by ChatGPT has developed rapidly, in addition to ChatGPT and Bing, many domestic manufacturers have also laid out, including Huawei, Ali, Baidu, Jingdong, iFlytek, SenseTime and many other technology companies have also actively laid out, have launched their own large model products. 数字化转型网（www.szhzxw.cn）
However, looking back at these large model products, most of them are focused on the C end, that is, the end user’s products, providing users with office and life assistance.
However, enterprises are still open to large model products, according to a recent survey by the IBM Institute for Business Value: three-quarters of the ceos surveyed believe that the deployment of advanced generative artificial intelligence will bring competitive advantage to the enterprise.
The real value of the large model is: the landing of the application on the industry side, that is, the improvement of the maturity of the industry-level large model products. According to market research institutions, the global generative AI market will reach more than $10 billion by 2025. Among them, the enterprise-level generative AI market will occupy a considerable share and become one of the largest application areas.
A large industry model is a large model of a specific industry or domain that is optimized and customized for a specific domain task. Compared with the general large model, the industry large model is more professional and better performance. However, in the process of development of the industry large model, there are also many differences with the general large model.
At present, the main difference is that the industry large model not only needs a general corpus, but also needs a professional corpus for different industries and different scenarios. Li Gang, vice president of Digital China, told titanium media that the industry model is highly professional and requires a large amount of industry knowledge base, “At present, the corpus of this industry knowledge base needs to be controlled at 20%, no more or less.” Li Gang stressed, “More than 20%, the trained large model may ‘not speak human language’, resulting in communication barriers; It is less than 20 percent and does not have the expertise of the industry.”
Second, medical, legal, financial and other industries took the lead in landing
At present, the industry-level large model is still in the initial stage of development, although many companies have launched the industry large model, but the application is not mature. In terms of the application status of large models in various industries at this stage, as well as the development trend, medical, financial, legal and other industries are taking the lead in landing more mature industry-level large model applications. 数字化转型网（www.szhzxw.cn）
Zhou Hua, head of large model industry application at Zhiyuan Research Institute, told titanium media that at this stage, the general field of large models with high fault tolerance is relatively mature, and can play a more auxiliary role in similar intelligent customer service, document processing, etc., as well as the application of Vincenne diagram in the general field, and the professional field application of alleviating some illusion problems through retrieval enhancement technology. Are some of the application scenarios that enterprises are better to land at present.
Coincidentally, Lin Daozhuang, vice chairman of the IEEE Standards Association New Standards Project Committee and Chairman of the IEEE Digital Finance and Economic Standards Committee, also has a similar view, Lin Daozhuang said that at present, the application of large models is mainly concentrated in the “three industries” (service industry), the focus is to assist people to serve others faster and better. The industry-level large model is also expected to take the lead in the application of higher maturity in the service industry related fields.
As far as the current development status is concerned, the industries with rapid development of large industry models mainly include finance, medical care, and law.
In the financial industry, in March 2023, Bloomberg launched BloombergGPT, a large-scale language model for the financial industry for the first time, causing the market to pay attention to the large model in the financial vertical field; In June, Columbia University and New York University Shanghai launched FinGPT. 数字化转型网（www.szhzxw.cn）
In China, in July of the same year, Huawei released a new Pangu grand model, and the financial industry grand model is one of several general industry grand models; In September of the same year, Ant Group officially released the self-developed “Ant Basic large model”, and the customized “Ant Financial large model” on this basis.
Although many companies with large model capabilities are actively laying out large model products in the financial industry, it still takes time for large models to fully mature and land in the financial industry due to the particularity of the financial industry – high requirements for security compliance.
Zhang Jin once told titanium media that the financial industry is different from other industries, its regulatory requirements are extremely high, from a technical point of view, such as loan audit and other businesses actually have the initial landing of technical capabilities, but because of the requirements of security compliance, the large model can only play a supporting role in liberating productivity.
At present, the application of large models in the financial industry is mainly concentrated in risk assessment and management, and the construction of knowledge graph platform. In terms of risk assessment, large models can predict market risks and assess credit risks by analyzing a large amount of historical data and real-time information, so as to provide more accurate and timely risk management decision support for financial institutions.
On the other hand, by combining large models with knowledge graph platforms and replacing NLP technology with large models, financial institutions can improve their efficiency and risk control level at the same time. 数字化转型网（www.szhzxw.cn）
In addition to the financial industry is expected to have more mature application scenarios this year, Lin Daozhuang told titanium media that such as medical, legal advice, education and training, entertainment and other risk requirements are lower, partial service industries, in this year is expected to have more mature application scenarios landing. “In 2024, the large model will land in some industries with human-assisted verification, low risk and accuracy requirements, and more people can enjoy better services through the large model to assist humans to work.” But it’s not even close to replacing people.” Lin Tao Zhuang stressed.
Taking the medical industry as an example, the study and analysis of a large number of medical data by industry large models can automatically identify pathological characteristics, assist doctors in disease diagnosis, and improve diagnostic accuracy and efficiency. In this regard, Lin Daozhuang said that at this stage, China’s medical resources are scarce, many people queue for a long time to see the disease, and doctors are also very busy, through the assistance of large models, can help doctors quickly identify such as X-ray, CT and other medical records, so as to greatly improve the efficiency of doctors, but also reduce the waiting time of patients. “For industries where information or capabilities are severely asymmetrical, such as healthcare, legal and translation services, the intervention of large models will greatly improve the efficiency of service delivery and promote the standardization of service processes.” Lin Tao Zhuang stressed. 数字化转型网（www.szhzxw.cn）
In addition to “efficiency”, the landing of large models in the medical industry can also help patients and hospitals achieve “cost reduction”, through automated and intelligent medical assistance systems, medical costs can be reduced and the efficiency and quality of medical services can be improved.
On the other hand, industry large models in the medical industry can also help doctors optimize treatment plans. Through a comprehensive analysis of the patient’s condition, medical history, drug allergy and other conditions through a large model, doctors can be provided with a more comprehensive and personalized treatment plan to improve the treatment effect.
In Lin Daozhuang’s view, the vertical industry that the industry large model can take the lead in landing has a general commonality – knowledge-intensive industry, “The industry that the industry large model can take the lead in landing must be the industry that can provide value through knowledge-intensive,” Lin Daozhuang pointed out, “To some extent, through knowledge collection and knowledge management, we can achieve the goal of assisting human content and making decision-making suggestions, rather than replacing people.” 数字化转型网（www.szhzxw.cn）
In addition to healthcare, finance and other industries, it will still take time for large models to be implemented in industries such as industry and manufacturing. In this regard, Zhou Hua told Titanium Media that at this stage, the main capabilities of the large model are still reflected in text, document processing, general chat and shallow professional Q&A, as well as Q&A and generation in the field of general vision, for the professional language field with high requirements for logical reasoning and accuracy, involving the professional visual field of picture and video recognition related to disciplines and engineering. As well as “Vinson video” and other capabilities still need technical iteration, “this kind of large model products with multi-modal capabilities are still difficult to apply in the industry side,” Zhou Hua said, “in 2024, multi-modal models will become the focus of competition among major manufacturers in the field of large models.” As the maturity of multimodal models continues to increase, there will be more industry application scenarios.”
Third, security is still “can not be ignored”
In the process of the application of the industry’s large model, enterprises will also encounter many problems, in addition to the “illusion” problem of the large model, the biggest problem is the problem of security. 数字化转型网（www.szhzxw.cn）
Data security has been a common concern of all participants since the advent of the Internet, and this is also true in the era of AIGC. While large models bring us more convenience and improve efficiency, they also bring increasingly severe security challenges.
At present, data has become an important asset of enterprises and even the country, and issues such as data Security and privacy protection have also become the focus of attention of enterprises in various industries. According to the “2023 Data breach Cost Report” released by IBM Security, the average cost of global data breaches in 2023 reached $4.45 million. It was the highest in the history of the report and a 15% increase over the average of the past three years.
With the landing of large-scale model-related products, the threat to data security is bound to become greater and greater. According to the IEEE survey, other more threatening cybersecurity issues will emerge in 2024, including ransomware attacks (37% in 2024, up from 30% in 2023), phishing attacks (35% in 2024, up from 25% in 2023), and insider threats (26% in 2024). Up from 19% in 2023).
From the perspective of current technological development, AIGC brings greater challenges to network security mainly from three aspects: malware/network attack, distributed denial of service attack, network monitoring and privacy invasion; On the other hand, enterprises in AIGC also face data security challenges such as data breaches, data tampering/forgery.
In Lin Daozhuang’s view, although the era of AI, enterprises are facing greater challenges of network security and data security, but in this process, enterprises can also fully apply the ability of AI for early warning and defense, “through AI automatic learning to identify, predict some potential risks, and automatically isolate these risks.” Lin Daozhuang said, “Through the empowerment of AI, enterprise-grade firewalls have stronger resilience.” 数字化转型网（www.szhzxw.cn）
Based on the security field, in fact, there are already many head manufacturers trying to integrate AIGC capabilities into security products and solutions, for example, Yunqi Infinity released a network security large-scale model open source project named “SecGPT” at the end of 2023; 360 opened the 360 Brain large model in September 2023, and it is understood that the model is the first native safety large model in the country; Tencent Security feeds security knowledge corpus for secondary training on the basis of hybrid large model to produce a large security model, and builds Tencent Cloud AI security Assistant…….. based on the large security model
Indeed, AIGC is like a “double-edged sword”. On the one hand, it brings more convenience and improves efficiency; On the other hand, network attacks based on AIGC technology and threats to data security will also increase. 数字化转型网（www.szhzxw.cn）
And for almost all businesses, when choosing to use digital technology, the premise of everything is security. In the process of enjoying the technical dividends brought by AIGC, the consideration of safety will also become an important track for the development of AIGC.
本文由数字化转型网（www.szhzxw.cn）转载而成，来源于ITValue；编辑/翻译：数字化转型网宁檬树。
免责声明: 本网站(https://www.szhzxw.cn/)内容主要来自原创、合作媒体供稿和第三方投稿，凡在本网站出现的信息，均仅供参考。本网站将尽力确保所提供信息的准确性及可靠性，但不保证有关资料的准确性及可靠性，读者在使用前请进一步核实，并对任何自主决定的行为负责。本网站对有关资料所引致的错误、不确或遗漏，概不负任何法律责任。
本网站刊载的所有内容(包括但不仅限文字、图片、LOGO、音频、视频、软件、程序等) 版权归原作者所有。任何单位或个人认为本网站中的内容可能涉嫌侵犯其知识产权或存在不实内容时，请及时通知本站，予以删除。
