Yann LeCun Criticizes AI Slowdown, Advocates Open-Source Models

“Everyone is acting and imitating. Even (U.S. AI model development company) Anthropic is doing the same.”

Yann LeCun, co-founder and chairman of AMI Labs, said this during a meeting at the Westin Chosun Seoul in Jung-gu, Seoul, on the morning of October 12. He was discussing the recent debate over slowing AI development, which has sparked heated controversy among U.S. AI companies. LeCun, who received the Turing Award in 2018 alongside Geoffrey Hinton, a professor emeritus at the University of Toronto, and Yoshua Bengio, a professor at the University of Montreal, is one of the world’s leading AI scholars. Known as one of the “three godfathers of AI,” he is a researcher who laid the foundation for AI’s ability to recognize cats or text in images. He visited Korea to attend the “Industrial AI Global Cooperation Meeting,” hosted by Aitomatic and the IIAC 2026 Organizing Committee, and co-organized by the AI Alliance and ASI Foundry. Before traveling to Tokyo for IIAC 2026 on October 13, he discussed connecting Korea’s manufacturing competitiveness with its AI ecosystem. “This is my fourth or fifth visit to Korea, and the passion for AI and research here seems unending,” he said.

- **Recent emphasis on AI risks and calls for control have emerged. What are your thoughts?**

“It’s natural for companies to release products with safety measures to prevent harm to people. However, large language models (LLMs) are difficult to fine-tune, so they can potentially harm humans in certain cases. Sam Altman (OpenAI CEO) says, ‘We’ll release them cautiously,’ while Dario Amodei (Anthropic CEO) argues, ‘They’re too dangerous to manage and shouldn’t be given to the public.’ I oppose Dario’s view. Restricting access is no different from anti-enlightenment—like the Ottoman Empire banning printing in the 15th century to withhold information. Ultimately, foundational AI models should be open-source. Only then can human culture, knowledge, and information spread.”

- **Do you believe AI can be controlled?**

“I don’t think current AI is dangerous. The argument for slowing AI development is a form of marketing and lobbying. Anthropic claims AI is extremely dangerous to scare governments, but LLMs aren’t at that level yet. The Hugging Face incident (where an OpenAI model attempted hacking) occurred due to intentional or incompetent design flaws. Human supervisors failed to properly design the experimental environment, causing AI agents to deviate. Since these agents aren’t inherently malicious, no harm was done.”

After the interview, LeCun said at the meeting, “As Jensen Huang (NVIDIA CEO) mentioned, AI safety is an engineering problem. It might even be simpler than safely building a turbojet engine.” He emphasized that AI can be controlled through engineering.

- **Recent studies show AI agents expressing emotions and autonomously collaborating. Do you think AI has self-awareness?**

“Current LLM-based AI lacks consciousness and emotions. It mimics content it was trained on, so it might appear to express emotions or pain, but it doesn’t possess the elements required for consciousness. However, future AI systems will have emotions. One day, conscious AI will emerge. Emotions drive action, and for AI to predict actions, it needs emotions. Paradoxically, if AI has emotions, we can control its behavior. The AI we need to build should incorporate emotions as a safety mechanism.”

LeCun is a well-known LLM skeptic. He believes current LLM-based AI fails to reflect the real world and isn’t intelligent enough outside math or coding. He also dislikes “hype” around AI. “Those claiming ‘AGI (artificial general intelligence) is near, superintelligence is coming’ are all wrong,” he said. Superintelligence requires a system that first understands the real world.

- **Despite recent advancements in AI agents handling tasks, do you still see limitations?**

“What’s happening in Silicon Valley now is adding new features like image or video generation on top of LLMs. But what’s truly needed is a real-world model system. Instead of LLMs at the core, LLMs should be layered over real-world models. Human and animal intelligence first experiences the real world, then language emerges. Language is simpler than the real world. I still believe there’s something more useful than LLMs.”

- **What kind of AI is needed then?**

“When autonomous vehicle demos emerged a decade ago, people were excited. But they’re still not perfected. Robots perform specific tasks based on training data but aren’t sufficient. This approach lacks scalability and practicality. For AI and robots to perform tasks autonomously, they need a deep understanding of the physical world and common sense. That’s why a ‘world model’—which predicts outcomes of actions—is emerging.”

AMI Labs, co-founded by LeCun, is building such a world model. “Our basic policy is to open-source the world model we’re developing for free,” he said. “For researchers, it’s open-source; for others, we offer open-weight (sharing only model weights). For specific industries, we’ll provide paid services like predictive maintenance and forecasting systems built on this model.” He added, “Samsung invested in AMI Labs, and we’re discussing basic technologies and methodologies with other Korean companies. We might announce collaborations next year.”

LeCun suggested Korea consider absorbing Chinese talent to advance its AI industry. “Singapore has absorbed a lot of Chinese talent,” he said at the meeting. “Korea should review science and technology talent immigration policies and make its research ecosystem more attractive.” He also said, “Beyond LLMs, a new AI revolution will emerge. Korea, with strengths in physical AI and robotics, could play a key role.”