Abstract:
Yann LeCun, winner of the 2018 Turing Award and founder and chairman of AMI Labs, said in Seoul on October 12 that the next stage of breakthroughs in artificial intelligence will come from systems that can understand, predict and act on the physical world, rather than simply expanding the scale of large language models.
He believes that "World Models" will become an important technical direction to promote the development of AI. Regarding the security concerns caused by the enhanced autonomy of artificial intelligence, Yang Likun said that recent network security incidents involving AI agents have mainly exposed the flaws of the existing security protection system, and related risks can be controlled by improving system architecture and security engineering.
Yang Likun: Large language models lack a reliable understanding of the causal relationship in the physical world
Yang Likun said at an industrial artificial intelligence symposium in Seoul on Monday that although current large language models can effectively learn knowledge from text, images and videos, they still lack a reliable understanding of cause and effect in the physical world.
He pointed out that artificial intelligence can already prove mathematical theorems and pass professional exams, but robots are still difficult to complete many daily household chores that 10-year-old children can master without training. This capability gap reflects that current AI systems still have obvious limitations in physical environment understanding, action planning, and real-world interaction.
Yang Likun believes that the important development direction of AI technology in the next stage is world models, that is, artificial intelligence systems that can predict how the physical environment will change with specific actions. In contrast to language models, which accomplish tasks primarily by predicting the next word or learning patterns in human-generated data, world models are designed to help AI understand the relationship between actions and outcomes and plan a series of actions based on goals. For example, robots need to be able to predict how the environment will change after moving or grabbing an object in order to reliably complete complex tasks in the real world.
AMI Labs founded by Yang Likun is advancing this technology route through the Joint-Embedding Predictive Architecture (JEPA) and exploring how to make artificial intelligence more reliable in environmental prediction and action planning capabilities.
As AI technology gradually expands from knowledge processing and content generation to fields such as robotics and industrial automation, world models are regarded by Yang Likun as an important technical basis for realizing intelligence in the physical world. However, relevant technologies are still being developed, and whether they can achieve reliable prediction, independent planning and large-scale deployment in complex real-world environments remains to be further verified.
Yang Likun refutes concerns that AI is out of control: recent security incidents have exposed engineering protection flaws
When discussing the development prospects of AI autonomous action capabilities, Yang Likun also responded to recent cybersecurity incidents involving artificial intelligence agents and objected to interpreting these incidents as AI is becoming uncontrollable.
Regarding the recent AI-related cyber attacks encountered by South Korea’s financial industry, Yang Likun believes that the problem mainly stems from the lack of sufficient network security professional capabilities of companies deploying AI agents and the security vulnerabilities in the sandbox system used to restrict software behavior. He pointed out that current large language models may indeed be subject to jailbreak attacks, bypassing established security restrictions through specific prompt words, and it is difficult to completely constrain their behavior just by relying on model fine-tuning.
However, Yang Likun believes that future goal-oriented AI systems are expected to proactively evaluate safety constraints before performing operations, thereby improving the shortcomings of existing models in behavioral control. Rather than relying solely on security rules established during the model training phase, such systems can incorporate additional security checks during action planning and execution.
He acknowledged that achieving this capability is not a simple technical task, but emphasized that related issues are still within the scope of engineering design. In the long term, AI may even give cyber defenders an advantage, as defense agencies typically have more resources and are able to share security information and defense experience more openly.
This judgment is different from some views that emphasize the potential risk of loss of control of advanced AI systems. Yang Likun also emphasized that through system architecture design, execution authority restrictions and security protection mechanisms, ensure that the behavior of autonomous AI systems is within a controllable range.
In his technical vision, in the future artificial intelligence will gradually have the ability to understand the real environment, predict the consequences of actions, and complete tasks autonomously. How to enhance these capabilities while establishing reliable safety constraints will become an important issue that needs to be solved in the development of world models and related autonomous systems.
Comments