2018年图灵奖得主、AMI Labs创始人兼董事长杨立昆（Yann LeCun）10月12日在首尔表示，人工智能的下一阶段突破将来自能够理解、预测并作用于物理世界的系统，而非单纯扩大大语言模型规模。
他认为，“世界模型”（World Models）将成为推动AI发展的重要技术方向。针对人工智能自主性增强所引发的安全担忧，杨立昆表示，近期涉及AI智能体的网络安全事件主要暴露了现有安全防护体系的缺陷，相关风险可以通过改进系统架构和安全工程加以控制。
杨立昆：大语言模型缺乏对物理世界因果关系的可靠理解
杨立昆周一在首尔举行的工业人工智能讨论会上表示，当前大语言模型虽然能够有效学习文本、图像和视频中的知识，但仍然缺乏对物理世界因果关系的可靠理解。
他指出，人工智能已经能够证明数学定理、通过专业考试，但机器人仍难以完成许多10岁儿童无需训练就能掌握的日常家务。这种能力差距反映出当前AI系统在物理环境理解、行动规划及现实世界交互方面仍存在明显局限。
杨立昆认为，下一阶段AI技术的重要发展方向是世界模型，即能够预测物理环境如何随着特定行动发生变化的人工智能系统。与主要通过预测下一个词或学习人类生成数据中的模式来完成任务的语言模型相比，世界模型旨在帮助AI理解行动与结果之间的关系，并根据目标规划一系列操作。例如，机器人需要能够预测移动或抓取物体后环境将发生什么变化，才能可靠地完成现实世界中的复杂任务。
杨立昆创立的AMI Labs正在通过联合嵌入预测架构（Joint-Embedding Predictive Architecture，JEPA）推进这一技术路线，探索如何使人工智能具备更可靠的环境预测和行动规划能力。
随着AI技术逐渐从知识处理和内容生成扩展至机器人及工业自动化等领域，世界模型被杨立昆视为实现物理世界智能的重要技术基础。不过，相关技术仍在持续研发，能否在复杂现实环境中实现可靠预测、自主规划及规模化部署，仍有待进一步验证。
杨立昆反驳AI失控担忧：近期安全事件暴露的是工程防护缺陷
在讨论AI自主行动能力的发展前景时，杨立昆还回应了近期涉及人工智能智能体的网络安全事件，并反对将这些事件解读为AI正在变得无法控制。
针对韩国金融行业近期遭遇的AI相关网络攻击，杨立昆认为，问题主要源于部署AI智能体的企业缺乏足够的网络安全专业能力，以及用于限制软件行为的沙箱系统存在安全漏洞。他指出，当前大语言模型确实可能受到越狱攻击，通过特定提示词绕过既定安全限制，而且仅依靠模型微调也很难完全约束其行为。
不过，杨立昆认为，未来以目标为导向的AI系统有望在执行操作前主动评估安全约束，从而改善现有模型在行为控制方面的不足。相比仅依赖模型训练阶段建立的安全规则，这类系统可以在行动规划与执行过程中纳入额外的安全检查。
他承认，实现这种能力并非简单的技术任务，但强调相关问题仍属于可以通过工程设计解决的范畴。从长远来看，人工智能甚至可能使网络防御方获得优势，因为防御机构通常拥有更多资源，而且能够更加开放地共享安全信息和防御经验。
这一判断与部分强调先进AI系统潜在失控风险的观点存在差异。杨立昆更强调通过系统架构设计、执行权限限制和安全防护机制，确保自主AI系统的行为处于可控范围。
在他的技术设想中，未来人工智能将逐步具备理解现实环境、预测行动后果以及自主完成任务的能力，而如何在增强这些能力的同时建立可靠的安全约束，将成为世界模型及相关自主系统研发需要解决的重要问题。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
In Seoul on 12 October, Yann LeCun, the winner of the 2018 Turing Prize, founder and chairman of AMI Labs, stated that the next phase of the breakthrough in artificial intelligence would come from systems that understand, predict and work in the physical world, rather than simply scaling up large language models.
In his view, the World Models would be an important technological orientation for AI. In response to security concerns raised by increased artificial intelligence autonomy, Yang Liqun indicated that recent cyber-security incidents involving AZAI agent had mainly exposed deficiencies in the existing security protection system and that the associated risks could be controlled through improvements to the system architecture and security engineering.
Yang Liquon: Large language models lack a reliable understanding of causality in the physical world
At the Industrial Artificial Intelligence Seminar held in Seoul on Monday, Yang Liqun stated that the current large language model, while effectively learning about text, images and video, still lacks a reliable understanding of causality in the physical world.
He noted that artificial intelligence had proved the mathematical theorem and had passed a professional examination, but that robots still had difficulty in completing the daily household that many 10-year-olds could master without training. This capability gap reflects the clear limitations of the current AI system in terms of physical environment understanding, operational planning and real-world interaction.
In Yang Liqun ' s view, an important development in the next phase of AI technology is the world model, an artificial intelligence system that can predict how the physical environment changes with specific actions. Compared to the linguistic model that accomplishes the mission primarily by predicting the next word or learning the pattern in human-generated data, the World Model is intended to help AI understand the relationship between action and results and to plan a range of operations according to the objective. For example, robots need to be able to predict changes in the environment after moving or capturing objects in order to reliably perform complex tasks in the real world.
The AMI Labs, created by Yang Liquon, is advancing this technological path through the Joint-Embeding Predicative System, JEPA, to explore ways to equip artificial intelligence with more reliable environmental forecasting and operational planning capabilities.
As AI technology gradually expands from knowledge processing and content generation to robotics and industrial automation, the world model is seen by Yang Liquon as an important technological foundation for the realization of physical world intelligence. However, the technology continues to be developed and the ability to achieve reliable forecasting, autonomous planning and scale deployment in complex realities remains to be further tested.
Yang Liqun rebuted AI's fear of losing control: recent security incidents have exposed engineering protection deficiencies
In discussing the prospects for the development of AI ' s ability to act autonomously, Yang Liqun also responded to recent cyber-security incidents involving artificial intelligence AI agent and opposed interpreting them as being beyond the control of AI.
In response to recent AI-related cyberattacks in the financial industry in Korea, Yang Liqun argued that the problem arose mainly from the lack of sufficient professional capabilities for cyber-security in companies deploying AIAI agent and security gaps in sandbox systems used to limit software behaviour. He pointed out that the current large-language model might indeed be subject to escape attacks, bypassing established security restrictions through a specific prompt, and that it would be difficult to regulate its behaviour entirely by simply fine-tuning the model.
However, Yang Liqun believes that future target-oriented AI systems are expected to proactively assess safety constraints prior to implementation, thereby improving the behavioural control deficiencies of existing models. Such systems can incorporate additional security checks in the course of operational planning and implementation, as compared to security rules that rely only on those established during the training phase of the model.
He acknowledged that the realization of that capability was not a simple technical task, but stressed that the relevant issues remained within the scope of engineering solutions. In the long run, artificial intelligence may even gain advantage over cyberdeficit, as defence agencies usually have more resources and are able to share security information and defence experience more openly.
This judgement differs from the one that partially emphasizes the risk of potential uncontrolledness of the advanced AI system. Yang Liqun placed greater emphasis on ensuring that the conduct of the autonomous AI system was manageable through the design of the system architecture and the implementation of permission restrictions and safety protection mechanisms.
In his technology scenario, the ability to understand the real environment, predict the consequences of action and carry out tasks on its own will gradually become available in the future, and how to build reliable safety constraints while enhancing those capabilities will be an important issue to be addressed in the development of models and associated autonomous systems worldwide.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.