我国生成式AI和Web3.0的发展仍面临五个挑战

　　中国经济时报记者 徐蔚冰　　

China Economic Times reporter Xu Wei-Chang

　　2023年11月22日，中国社会科学院上市公司研究中心与社会科学文献出版社联合发布了《中国上市公司蓝皮书：中国上市公司发展报告（2023）》。　　

On 22 November 2023, the company research centre listed by the Chinese Academy of Social Sciences and the publishing house of social science published " The Blue Book of China's listed companies: China's listed company development report (2023) ".

　　蓝皮书指出，我国生成式AI和Web3.0的发展仍面临以下五个挑战。　　

According to the Blue Book, the development of our generation AI and Web3.0 continues to face the following five challenges.

　　一是确保模型开发的充足投资，并形成可竞争的市场结构，保持大语言模型在实现规模经济和促进基于生成式AI的产业应用次级创新之间的平衡。作为通用目的技术，对生成式AI的大语言模型投资具有沉淀成本特征和数字经济新一代基础设施投资性质。如果没有对大语言模型足够的投资，将很难在信息处理成本上实现规模经济，并促进最为广泛的产业应用。然而，大语言模型的规模经济特征意味着只有少数寡头开发者才会在竞争中脱颖而出，服务的垄断定价又有可能成为基于生成式AI的产业应用次级创新障碍。因此，只有确保形成模型开发可竞争的市场结构，使大语言模型成为按成本定价的数字经济新一代基础设施，才能将对大语言模型的充足投资最终转化为生成式AI的产业级低成本应用。尽管我国目前对大语言模型的预训练才刚刚起步，中心化互联网平台企业侧重于生成式AI的内部应用，各级政府引导基金对大语言模型预训练又缺乏投资动力，但借助来自国际的开源大语言模型，仍可在模型微调的中间层和引入基于人工反馈的强化学习的产业应用层大有可为，继续发挥在人力资本积累和国内市场规模方面的传统优势。　　

One is to ensure adequate investment in model development and to create a competitive market structure that maintains a balance between large language models achieving economies of scale and promoting sub-innovation in industrial applications based on generation AI. As a general purpose technology, investment in large language models of generation AI has the character of a sedimentary cost feature and a new generation of infrastructure investment in digital economy. Without sufficient investment in large language models, it will be difficult to achieve economies of scale in terms of the cost of information processing and to promote the broadest industrial application. However, the economic dimension of large language models means that only a few oligopolistic developers are emerging in competition, with monopolistic pricing of services likely to become sub-innovated in generating AI-based industries.

　　二是实现对以中心化平台企业为核心的互联网（Web2.0）实行Web3.0改造，建设基于区块链的新一代互联网，避免生成式AI节约的信息处理成本被不断上升的互联网引流和匹配成本所吞噬，充分激励第三方基于生成式AI的产业应用次级创新。我国当前数据分配主要实行的是“以数据交换服务”“谁占有谁受益”等事实上的分配机制。这样的分配格局正是以中心化平台企业为核心的互联网（Web2.0）主导的结果。Web2.0无法实现网络效应内生化，极易导致网络租金被中心化平台企业独占。在Web2.0上，互联网引流和匹配成本迟早会吞噬生成式AI节约的信息处理成本，并最终抑制基于生成式AI的产业应用次级创新。2022年12月，《中共中央国务院关于构建数据基础制度更好发挥数据要素作用的意见》明确提出数据要素的分配要由市场评价贡献、按贡献决定报酬，按照“谁投入、谁贡献、谁受益”的原则，强调可以通过分红、提成等多种形式来实现收益共享。因此，对以中心化平台企业为核心的互联网实行Web3.0改造势在必行。只有这样，才能实现网络效应内生化，真正满足数据要素按贡献参与分配的基本分配原则和要求，充分激励基于生成式AI的产业应用次级创新。否则，一旦基于生成式AI的产业应用次级创新受到抑制，生成式AI对生产率的提升作用极有可能像传统互联网一样昙花一现。　　

The second is the realization of the Web3.0 transformation of the Internet (Web2.0), which is centred on centralized platform enterprises, and the creation of a new generation of Internet-based block-based Internet-based Internet-based Internet-based systems that avoid the cost of generating AI-saving information processing that is absorbed by rising Internet-based flows and matching costs, which fully motivates third parties to apply industrial applications based on generation-based AIs. On Web 2.0, the Internet-based flow and matching costs are mainly distributed & & & & & & & & & & & & & & & ; who owns the benefits & benefits & benefits & benefits & benefits & & & & & & & & & & & & & & & & & & & & & & ;

　　三是大力开发以增强现实（AR）和虚拟现实（VR）为代表的新一代互联网终端，推动生成式AI必要的人类专家干预和相应的人机互动。无论是大语言模型功能泛化的微调和基于人工反馈的强化学习，还是基于生成式AI的产业应用次级创新都离不开必要的人类专家干预和相应的人机互动。由于能够实现沉浸式体验，AR和VR有望从狭隘的游戏娱乐场景解放出来，发展成为对人机互动最为友好的新一代互联网终端。由AR和VR代表的新互联网终端无疑有助于提高生成式AI人类专家干预和人机互动的效率。　　

Third, the new generation of Internet terminals, represented by reality (AR) and virtual reality (VR), is likely to be freed from narrow play scenes and developed into the new generation of Internet terminals that are most friendly to human interaction. The new Internet terminals, represented by AR and VR, will undoubtedly help to improve the efficiency of generating AI human intervention and human interaction.

　　四是积极打造终身学习体系，促进人力资本积累，充分挖掘由生成式AI驱动的智能产业革命潜力。生成式AI代表了数字化转型的最新发展方向，加剧了技术和需求结构变化，并进一步提高终身学习的重要性。除了包括生成式AI在内的ICT相关技能外，为了应对迅速变化的工作任务，还需具备与数字技术互补的认知能力。　　

The fourth is to actively build a lifelong learning system that promotes the accumulation of human capital and fully exploits the potential of the intelligent industrial revolution driven by the Genesis AI. Generating AI represents the latest direction of the digital transformation, exacerbates technological and demand structural changes and further enhances the importance of lifelong learning.

　　五是完善智能安全技术，稳健推进智能产业革命。智能安全技术是智能产业革命实现的前提，主要包括两方面内容:一要实现数据确权和隐私保护。尽管中心化平台企业拥有的隐私计算技术有助于促进对数据安全的保护，但受制于中心化的网络治理结构，其所承诺的数据调用需经严格授权程序却不完全可信。这就需要引入区块链，进一步保护数据安全。具体地讲，就是运用可验证计算、同态加密和安全多方计算等密码学的进步技术支持数据确权，使在不影响数据所有权的前提下交易数据使用权成为可能，并影响数据主体和数据控制者的经济利益关系。二要提高模型的可解释性和透明度。由于本质上属于对记忆学习过程进行模拟的算法黑箱，大语言模型不易及时发现错误，并纠正相应后果。这就需要引入基于人工反馈的强化学习，并同分析式AI相结合，双管齐下地修正大语言模型算法的黑箱性质，提高模型的可解释性和透明度。前者有望将人工智能和人类智力结合起来，实现面向模型的人工智能向面向智能体的人工智能的转变,确保生成式AI更好地为人的利益服务。后者则能够促进面向模型的生成式AI和面向工具的分析式AI的结合，有助于及时纠错和进行必要的问责。　　

The fifth is the improvement of smart security technologies and the robust advancement of the intelligent industrial revolution. Smart security technologies are a prerequisite for the intelligent industrial revolution: data validation and privacy protection are achieved as soon as it is achieved. While privacy computing techniques owned by centralized platform enterprises help to protect data security, they are subject to a centralized network governance structure in which the promised use of data is not fully credible. This requires the introduction of block chains to further protect data security. This requires, in particular, the introduction of enhanced learning based on manual feedback, combined with analytical AI, a two-pronged correction of the black box nature of the large language model, and influences the economic interests of data subjects and data controllers. The second is to improve the interpretability and transparency of models.

　　该蓝皮书认为，如果能够成功应对上述一系列挑战，就可以既实现大语言模型规模经济和产业化应用，又可以完成现有互联网的Web3.0改造，充分激励基于生成式AI的产业应用次级创新。只有这样，才能引发生成式AI和Web3.0双轮驱动的智能产业革命。

The Blue Book argues that if the above-mentioned series of challenges can be successfully addressed, both large-language model economies of scale and industrial applications can be achieved, and the Web 3.0 adaptation of the existing Internet can be completed, providing sufficient incentives for sub-innovation in industrial applications based on generation AI. Only then can the generation AI and Web3.0 two-wheel drive intelligent industrial revolutions be triggered.