近日，由大湾区科学论坛（GSF）主办的双语季刊《大湾区科学论坛》（GSF Journal，ISSN 3734-8912）2026年第一期刊发图灵奖得主、法国科学院院士、法国国家工程院院士、欧洲科学院院士、美国艺术与科学学院院士、美国国家工程院院士、美国科学院外籍院士、中国科学院外籍院士、斯发基斯可信自主系统研究院院长、南方科技大学计算机科学与工程系杰出访问教授约瑟夫・斯发基斯（Joseph Sifakis）文章《人工智能：洞察挑战，把握机遇》。全文转载如下：

人工智能：洞察挑战，把握机遇

自本世纪初以来，人工智能开启了科技革命的新纪元，也成为了人类历史的转折点。如今，机器已经能在知识生产与应用方面比肩人类，这为科技进步开辟了新的视野。但需要特别指出的是，尽管人工智能取得了令人瞩目的成就（生成式人工智能的出现更是将其推向顶峰），但总体而言，人工智能仍处于起步阶段。一项关于人工智能在实体经济中有效应用的分析表明，目前这些应用主要是提供问答服务的助手类产品，例如 ChatGPT 和 DeepSeek。人工智能只能提供基础要素，而有待解决的问题是：如何组合这些基础要素，来构建具有人类同等智能水平的智能系统？

在科学、服务和工业的诸多应用领域中，人工智能的使用范围仍然有限。这些领域对自动化和可靠性的要求更高，而且需要对话式以外的交互模式。

在科学应用方面，无论医学还是生命科学，无论气候科学还是宇宙科学，都需要能够分析复杂现象并理解其动态与特性的系统。而人工智能将深刻地改变人类创造和应用知识的方式。众所周知，面对依赖大量独立参数的现象时，人类的观测能力和分析能力有限。因此，科学理论往往相对简化且存在明显局限。相比之下，人工智能可以分析多维数据，从而揭示复杂的生物、物理和社会现象。

我们必须着力推动人工智能在科学领域的发展，这不仅需要资金投入，还要在各科学领域建设相应的数据基础设施。人类专家与专用人工智能系统通力合作，必将发现并最终构建那些人类心智所无法掌握的关联，从而应对复杂现象。

已经有实例表明，相较于传统的预测方法，人工智能可以助力改进天气预报甚至地震预测，AlphaFold1也在预测蛋白质三维结构方面取得了成功。

当然，这样得到的知识仍属经验性成果，并不具备传统科学方法的有效性。为此，我们应该组织并推动人机协作，掌控知识生产过程，以免机器在不受控的情况下自主生成和应用“准科学（parascience）”。

在工业应用方面，我们亟需发展自主系统——以物联网构想下的应用场景为例，包括自主交通系统、智能电网、智能工厂和农场，以及自主通信系统等。在所有智能系统中，自主系统的实现难度最大，因为其设计初衷是取代复杂组织中的操作人员。自主系统由多个智能体组成，每个智能体既要独立达成自己的特定目标，也要协作达成系统的整体目标。可以说，自主系统是人工智能发展的终极阶段。

当前的人工智能技术还远不足以实现完全自主的愿景，这一点从自动驾驶汽车产业遭受的挫折可见一斑。虽然注入了大量资金，但 2020 年实现完全自动驾驶的承诺至今仍未兑现。

自主系统愿景正在冲击传统系统工程。这一愿景要将信息通信技术（ICT）所使用的传统研发方法与数据驱动的人工智能技术结合起来，而后者在可解释性和可靠性上存在与生俱来的局限。在可靠的智能系统中集成不可靠的人工智能组件，必须采用可扩展性和可适应性兼具的混合解决方案，尤其要将符号化知识和非符号化知识（如感知信息及决策模型）关联起来。在开发过程中，则需要在设计的精确性和运行的韧性之间寻求平衡，系统本身也需要通过常规更新或特殊更新的方式持续迭代。

此外，目前尚无评估人工智能系统可靠性的技术。基于模型的传统验证技术虽然能够保障较高的可靠性，但是考虑到自主系统本身的复杂性和异质性，这些方法并不适用。我们亟需超越现有随机测试的方法，以测试和仿真为基础，研发严谨的、全新的验证技术。最终，通过运用基于知识的监测技术，来弥补可靠性保障的缺失。

人工智能当前的局限性源于多方面因素，其中包括科技巨头战略性地偏向对话式技术，以便促使该技术被公众广泛采纳。这些大型科技公司正在以一种蛮力式的、颇具争议的方式展开疯狂博弈——它们认为，只要堆叠扩大现有人工智能模型的规模，就足以使其达到人类智能水平。因此，它们极力推动规模化，对高能耗的基础设施投入巨额资金。显而易见，这种战略意在通过“闪电战”迅速占领市场，进而淘汰无法达到同等投资体量的竞争公司，甚至国家。

人工智能创造了新的可能性，而如何用它造福人类则取决于我们自己。因此，我们必须明智地使用人工智能，并管控其带来的风险。不幸的是，目前尚无评估此类风险影响程度的技术框架。我们必须辨别哪些是“超智能机器必将统治人类”的迷思，哪些才是真实存在的风险。

在知识的生产和应用领域，人类正在与机器——尤其是与计算机和人工智能展开竞争。在我看来，这是人类历史的转折点。与人工智能的交互将带来新的分工格局，因此我们必须对其加以理解和掌控。原始人类的生存依赖体力，但随着杠杆与机械的应用，体力已不再是必需。人工智能的应用也将带来类似变化。然而，某些方面正在发生质的改变：体力的减弱并不会带来灾难性后果，而心智能力的削弱和异化将深刻地改变人类最重要的特征之一——自主选择和自主行动的能力，即通过责任感和理性判断践行人类的自由意志。

正因如此，规范人工智能的使用就显得尤为重要，尤其当年轻人使用人工智能的时候。年轻人需要学会思考，需要培养创造力和负责任的行动能力。教育的目的在于锻炼心智。如果你想锻炼身体，难道派机器人去健身房替你运动就够了吗？同样，我们应当规范人工智能在教育中的使用，鼓励发展那类不直接给出现成答案，而是像教练之于运动员那样的辅导型智能体。

针对人工智能产品相关风险，目前尚未形成全球统一的监管框架，这一点已成为人工智能在关键决策过程应用中的重大障碍。安全人工智能是全球共同关切的话题，联合国活动和多场世界峰会均将其列为核心议题。遗憾的是，目前主要利益相关方仍未就一个切实可行的全球监管框架达成共识，他们的分歧体现在以下方面：使用人工智能所带来的风险、哪些风险可以且应当受到监管，以及如何落实监管。这些分歧在欧盟与美国在人工智能监管上的重大对立中尤为凸显。

历史证明，开放的生态系统与国际合作是推动技术进步的重要加速器。人工智能工具和平台的开放，将有力地推动系统互操作性和集成方面的突破。而推动国际合作，对于建立人工智能产品相关风险的监管框架同样不可或缺。

在此背景下，中国已然具备开发更可靠人工智能的良好条件，这样的人工智能更加符合实体经济的需要一一尤其是在盼望已久的自主系统转型方面。中国拥有坚实的工业和科研基础，以及统一的国内市场，能够提供海量的特定领域数据，为工业和科学领域研发前沿的人工智能应用提供支撑。

中国正在构建自己的人工智能愿景，这与大型科技公司截然不同。这一愿景以全球视角构建智能系统和基础设施，覆盖经济生活的各个领域，尤其体现在“粤港澳大湾区”的生态系统建设上一一这或许将是下一经济发展阶段最重要的项目。通过整合资本、人才、制造实力和政治决心，粤港澳大湾区有潜力成为全球领先的创新中心。

中国的人工智能发展理念能够产生协同效应，动员相关领域的关键力量，来建设可信自主人工智能所需的特定基础设施和数据平台。这一蓝图的实现，将有助于中国在全球人工智能的战略博弈中发挥平衡作用，并与志同道合的国家携手，从社会利益出发，推动人工智能的安全与发展。

AI: Understanding the challenges, seizing the opportunities

Since the beginning of  this century, AI has marked a new era in the scientific and technological revolution, as well as a turning point in human history. Machines can now rival humans in the production and application of knowledge, opening up new perspectives in scientific and technological progress. However, it should be noted that despite impressive achievements, which culminated with the arrival of generative AI, AI is still in its infancy. An analysis of effective AI applications in the real economy leads us to conclude that they currently focus on assistants, such as ChatGPT and DeepSeek, which operate in question/answer mode. AI only provides us with basic elements, and the question that remains unanswered is how to assemble them to create intelligent systems that reach the level of human intelligence. 

There are many fields of application in science, services and industry where the use of AI remains marginal. These areas have requirements for increased automation and reliability and modes of interaction other than conversational. 

For scientific applications, we need  systems capable of analyzing complex phenomena and understanding their dynamics and properties in various scientific fields, from medicine and life sciences, to climate and universe sciences. Artificial intelligence will radically change the way we create and apply knowledge. It is well known that humans have a limited capacity to observe and analyze phenomena that depend on a large number of independent parameters. Thus, scientific theories are relatively simple, with well-known limitations. Unlike humans, artificial intelligence is capable of analyzing multidimensional data that characterize complex biological, physical, and social phenomena.

We must strive to develop artificial intelligence for science, which requires investment and the development of an appropriate data infrastructure in different scientific fields. Human experts, working in collaboration with specialized AI systems, will be able to tackle complexity by discovering and, ultimately, formulating relationships that are inaccessible to the human mind.

We already have examples of how artificial intelligence is helping to improve weather forecasting and even predict earthquakes compared to forecasts using traditional methods, or the success of AlphaFold1 in predicting the three-dimensional structure of proteins.

Of course, the knowledge thus produced will only be empirical and will not have the validity of traditional scientific methods. Here, we must organize and support human-machine collaboration in order to maintain control over the processes of knowledge production, avoiding the creation of a “parascience”that would be produced and applied autonomously and uncontrollably by machines.

For industrial applications, we need to develop autonomous systems, such as those envisaged by the Internet of Things, for example autonomous transportation systems, smart grids, smart factories and farms, and autonomous communication systems. Autonomous systems are the most difficult intelligent systems to implement, as they are designed to replace operators in complex organizations. They are composed of agents, each pursuing their own specific goals while cooperating to achieve the overall goals of the systems of which they are a part. They can be considered the ultimate stage in the development of AI.

Current AI technologies are far from sufficient for achieving the autonomy vision, as demonstrated by the setbacks suffered by the autonomous car industry. The promise of fully autonomous cars by 2020, despite substantial investment, has not materialized.

The vision of autonomous systems is disrupting traditional systems engineering. It requires combining the traditional development methods that we have applied to ICT with data-driven AI techniques. The latter have inherent limitations in terms of their explainability and reliability. Building reliable intelligent systems that integrate unreliable AI components leads to hybrid solutions that must be scalable and adaptable. In particular, these require linking symbolic and non-symbolic knowledge, such as sensory information and models used for decision-making. Their development must seek a compromise between accuracy at design time and resilience at run time. Systems must be able to evolve through regular or exceptional updates.

Furthermore, we currently have no techniques for assessing the reliability of AI systems. Traditional model-based validation techniques, which can guarantee high reliability, are no longer applicable due to the inherent complexity and heterogeneity of these systems. We need to develop new rigorous validation techniques based on testing and simulation that go far beyond current random testing. Finally, the lack of reliability guarantees must be compensated for by the use of knowledge-based monitoring techniques.

The current limitations of AI are  due to several factors, including the biased emphasis on conversational technologies fostered by the strategy of tech giants, which promotes their adoption by the general public. Big Tech companies are engaged in a frantic race, relying on a brute-force and questionable approach that considers current solutions to be sufficient to achieve human-level intelligence and that it is mainly a question of scale. They are thus pushing for gigantism, which requires colossal investments in energy-intensive infrastructure. It is clear that this strategy aims to occupy the field through a blitzkrieg and eliminate competing companies, or even countries, that do not have the same investment capacity.