Meta’s Chief AI Scientist, Yann LeCun, on the Future of Artificial Intelligence

Meta’s chief AI scientist, Yann LeCun, recently shared his thoughts on the future of artificial intelligence, asserting that AI is on the brink of matching or even surpassing human intelligence. However, LeCun emphasized that current approaches, particularly those centered around Large Language Models (LLMs), are reaching their limitations.

Speaking at the NVIDIA GTC 2025, he explained, “I’m not so interested in LLMs anymore. They are reaching their limits by adding more data, increasing computing power, and using synthetic data.”

Shifting Focus Towards Fundamental AI Aspects

LeCun highlighted four key areas he believes are crucial for machine intelligence:

• Understanding the physical world
• Persistent memory
• Reasoning
• Planning

“There is some effort to get LLMs to reason, but in my opinion, it’s a very simplistic way of viewing reasoning. I think there are probably better ways of doing this,” he stated.

The Importance of World Models and Limitations of Token Prediction

LeCun introduced the concept of “world models,” systems capable of forming internal representations of the physical environment, which are crucial for reasoning and prediction. “We all have world models in our minds. This is what allows us to manipulate thoughts, essentially,” he noted.

He criticized the current reliance on token prediction in LLMs, arguing that it is insufficient for understanding complex data such as video. “Tokens are discrete… When you train a system to predict tokens, you can never train it to predict the exact token that will follow,” he explained, adding that current methods fail to grasp high-dimensional data.

Instead, LeCun advocated for joint embedding predictive architectures, which predict in an abstract representation space rather than the raw input space. He described a method where a system observes the current state, envisions an action, and predicts the next state as part of its planning process. “We don’t do [reasoning and planning] in token space. That’s the real way we all do planning and reasoning,” he emphasized.

Criticism of Agentic AI Systems and AGI Hype

LeCun also critiqued agentic AI systems that rely on generating numerous token sequences and then selecting the best outcome. “It’s sort of like writing a program without knowing how to write it. You write a random program and then test them all… It’s completely hopeless,” he remarked.

With growing discussions about the imminent arrival of artificial general intelligence (AGI) or advanced machine intelligence (AMI), LeCun remains skeptical. He places his trust in gradual, fundamental improvements in AI capabilities rather than breakthrough solutions.

About the Journalist

Siddharth Jindal, the article’s author, is a media graduate passionate about technology and its impact on society. He endeavors to bring thoughtful insights to readers through his journalism.

For more on AI advancements and expert opinions, make sure to subscribe to aitechtrend.com.