Are World Models the Future of AI? Insights from Distinguished Speaker Yann Lecun and CCDS Faculty

On 3 November 2026, Advanced Machine Intelligence (AMI) Executive Chairman Yann Lecun gave a well-attended talk, “World Models: Enabling the Next AI Revolution,” as part of the College of Computing and Data Science and IAS@NTU Lee Kong Chian Distinguished Professor Public Lecture series.

Lecun, who is also a Jacob T. Schwartz Professor at NYU, discussed the current challenges facing AI systems, particularly LLMs, in attaining human-like intelligence. To overcome those challenges, Lecun proposed that world models, instead, offer a promising path for developing truly “intelligent” systems.

World models, which Lecun seeks to develop at AMI, have various definitions. Generally speaking, they are systems that create a model of a given environment, for instance, the physical world that we live in. They aim to “learn” the rules of the environment, whether they are the laws of physics or the complex dynamics of human social interactions. World models can then reason using these rules to predict outcomes or to propose actions that will lead to desired outcome.

This is very different from LLMs, which are trained using vast databases of human language, and “learn” the world through human descriptions. In contrast, a world model “learns” the world through its own sensors and inputs.

To emphasize his point, Lecun closed his talk with a slide titled “Recommendations to my fellow AI scientists” in which he suggested that generative models, probabilistic models, contrastive methods, and reinforcement learning should all be “abandoned”.

“If you are interested in human-level AI, don’t work on LLMs,” he concluded, setting the stage for a robust discussion.

Biologically Inspired AI Offers Additional Insights: Zhang Mengmi

CCDS Nanyang Assistant Professor Zhang Mengmi studies the intersection of artificial intelligence and computational neuroscience. Her research group investigates how brain computations inspire new paths in AI and how AI can help elucidate brain computations.

Zhang agreed with Lecun’s proposal and suggested that there are additional possible avenues for world models.

“I strongly agree with Yann LeCun that language cannot be the foundation of human-like intelligence,” said Zhang. “Developmental science offers important insights into how intelligence emerges: infants learn about the world through perception, exploration, and interaction long before they master language. Learning predictive representations in a joint embedding space is a promising first step, but I believe we must go further by developing AI agents that actively seek information, decide where to look, and learn through their own experiences, rather than passively observing the world. This is where biologically inspired attention, memory, and active learning become essential, as we investigate in my lab. While Yann describes reinforcement learning as the ‘cherry on the cake,’ I believe that cherry is equally worth studying—after all, someone has to figure out how to make the cherry!”

Reinforcement Learning Is a Fundamental Principle: Stefano Albrecht

CCDS Associate Professor Stefano V. Albrecht is leader of the Autonomous Agents Research Group at NTU. His research develops algorithms for autonomous systems control, decision making, and multi-agent interaction, with a focus on deep reinforcement learning and foundation model-based agents.

“I would have liked to see a more nuanced and insightful treatment of reinforcement learning,” Albrecht said. “World models play an important role and require further development, but they do not remove the complexity of optimal decision-making in uncertain and noisy environments, which is what reinforcement learning algorithms are designed to address. Reinforcement learning is one of the most fundamental principles of learning in natural and artificial systems, and it should be studied rigorously to advance modern AI.”

Language Is Powerful: Jordan Boyd-Graber

CCDS Professor Jordan Boyd-Graber, whose research focuses on the evaluation of language models and on building efficient, interactive AI systems, offered a different perspective on the importance of language when constructing AI models.

First, I should preface this by saying that there is a long history of Yann being right when I was wrong,” Boyd-Graber said. “However, while I agree that just scaling up and tweaking LLMs is a scientific (but probably not economic) dead end, I disagree about the importance of language. One of the reasons that LLMs have become so ubiquitous and powerful is that language is a natural language of interaction with people and that most of humanity’s knowledge has been stored as text. While it would be incredible if world models could recover language from first principles, the example of intelligent life we have is humans, and civilization is founded on communication. I think the evidence shows that AI progress is grounded in language, and it seems premature to abandon that scaffolding.”

Further Thoughts

Following the lecture, Lecun had a fireside chat with CCDS Professor Bo An, who is also head of the Division of Artificial Intelligence in CCDS, the President’s Chair in Computer Science and Engineering, and director of the AI-for-X Centre. In the chat, Lecun gave more detailed thoughts on world systems compared to other methods of training and building AIs. Afterwards, Lecun took questions from students and gave an interview.

Looking to the future, Lecun said that it will be important for future computer scientists to have a vision. He recommended that students take courses outside of computer science—such as physics and engineering—that can help inform and improve their ability to be a successful computer scientist in the future.

In addition, Lecun offered reassurance to today’s computer science students, despite the widespread and growing availability of AI with coding capabilities.

“The idea that AI and computer science is not a good professional avenue is completely false,” Lecun said. “Basic programming is dead. But computer science is more alive than ever and more important than ever. There is a difference between programming, which is writing code, and designing programs at the conceptual level, the architecture level. This is what humans still need to do.”