"Large language models (LLMs) will become useless within five years. If you are interested in advancing AI to the human level, you should study what LLMs cannot do."
World-renowned artificial intelligence (AI) scholar Yann LeCun said this on the morning of the 27th at the AI Frontier International Symposium 2025 held at Dragon City in Yongsan District, Seoul. Wearing Meta's smart glasses as he took the podium, he stressed that additional innovation is needed for AI to reach the level of human intelligence.
LeCun is a professor at the Courant Institute of Mathematical Sciences and chief AI scientist at Meta, and is regarded as a leading authority in AI. He received the Turing Award, known as the Nobel Prize of computer science, in 2018. LeCun serves as co-director of the Global AI Frontier Lab, launched at New York University in Sept. last year as a joint Korea-U.S. AI research hub.
At the symposium, LeCun delivered a keynote speech on world model learning. A "world model" is a technology that enables AI to understand how the real world works so it can predict and prepare for future situations like humans. Unlike LLMs that simply learn text to generate sentences, a world model learns movements in the real environment to predict the next action. Meta, OpenAI, and Google are developing world models that will surpass LLM-based models.
LeCun said, "We need systems that can handle complex sequences of actions, can reason over long chains of logic, and above all are controllable and safe," criticizing that "current AI architectures essentially have none of these elements or only to a very small degree." He emphasized that text alone cannot achieve human-level AI, and that AI systems must understand the physical world and be able to learn on their own through sensory inputs such as video.
LeCun proposed JEPA (Joint-Embedding Predictive Architecture) as a new world model that will surpass LLMs. JEPA is a non-generative AI model that understands and learns about the world through images and video. As Meta's chief AI scientist, he unveiled a new world model, V-JEPA2, in June. Focused on video learning, V-JEPA2 is a successor to the V-JEPA model first introduced last year, and was pre-trained on more than 1 million hours of footage and 1 million images.
He stressed that to make progress in AI, the field should, in the long run, abandon generative models and choose the non-generative JEPA architecture. LeCun predicted, "If JEPA-style architectures become widespread, there will be a shift from inference to optimization."
Following LeCun, Choi Yejin of Stanford University took the stage with a talk titled "Democratizing Generative AI: Beyond scaling laws." Choi works at Stanford's Human-centered AI (HAI) institute and was named to Time's list of 100 most influential people in AI.
Choi cited OpenAI's ChatGPT as an example and said the current global AI market being led by a handful of big companies is like "David and Goliath." Choi argued that the bias of "the bigger the model, the better" is reaching its limit, and that smaller models can outperform large ones through data learning.
Choi said, "It is meaningless to train on data already on the internet," explaining, "because every company has already downloaded that data." Choi added, "It is important to train on data that is not on the internet and is qualitatively different."
The symposium marked the first joint event by the National AI Research Hub established by the government in Korea and the Global AI Frontier Lab. Along with LeCun, fellow AI authorities Geoffrey Hinton of the University of Toronto and Yoshua Bengio of the University of Montreal sent congratulatory messages via video, drawing attention.
Bae Kyung-hoon, vice minister, said, "We are at a critical inflection point where AI is reshaping the landscape of every field of science, technology, industry, and society," adding, "Today's symposium is the first international event co-hosted by the two AI research institutes, providing a venue for researchers to explore the present and future of AI."
Bae added, "The Ministry of Science and ICT will continue to strengthen international cooperation in AI research and actively support coordination of a global ecosystem for AI research collaboration."