Yann LeCun: LLMs Are ‘Not a Path to Human-Level Intelligence’

PARIS — After more than a decade of helping shape artificial intelligence research at Meta, Turing Award winner Yann LeCun is betting that the industry’s dominant approach to AI has reached its limits.

At the VivaTech conference today in Paris, which sponsored this journalist’s trip, LeCun outlined the vision for his new startup, AMI (Advanced Machine Intelligence) Labs. The startup focuses on developing what he believes could become the next major breakthrough in AI: world models capable of understanding and reasoning about reality rather than simply predicting words.

LeCun, one of the pioneers of modern deep learning and former chief AI scientist at Meta, said the industry’s fascination with large language models has produced useful tools but has also created a technological monoculture that may not lead to human-level intelligence.

“There’s nothing wrong with LLMs. They’re useful,” LeCun said. However, “there is simply not a path to human-level intelligence.”

LLMs “are useful, they’re powerful and they’re super-human in some domains, but they have limits,” he continued. “The idea that somehow those limits are going to be expanded to cover all the things that humans do is just false.”

How could so many brilliant people in Silicon Valley be wrong?

“They took a pill, and now they’re hypnotized, and to some extent, a lot of the culture in Silicon Valley, or in the AI industry, is pilled,” LeCun said. “And the funny thing is that because all the companies are poaching each other’s engineers, and they’re all working on the same thing, they can’t really afford to deviate from the mainstream because they run the risk of falling behind. So that creates this monoculture, but I think it’s a bit of a delusion.”

Meta joined the bandwagon with a well-publicized spending spree to hire top AI researchers. LeCun said “what happened last year was that the company chose to refocus its efforts, including some of its research efforts, towards LLM and catching up with the rest of the industry.” While leadership was still willing to support his project, he said, “the general ambience was not as favorable.”

Why Meta pivoted away from open source

At Meta, LeCun often occupied a unique position. While the company aggressively pursued generative AI and large language models, he remained one of the field’s most vocal skeptics of the idea that scaling LLMs alone would produce artificial general intelligence, or AGI.

Although LeCun said he continued to receive support from Meta CEO Mark Zuckerberg and other executives, he described the overall atmosphere as becoming less favorable for the kind of long-horizon research that AMI required. He also pointed to growing discussions around limiting publications and becoming less open with research findings.

LeCun left Meta last November to start a new company built around research that began inside its Fundamental AI Research (FAIR) lab under a project known internally as Advanced Machine Intelligence, or AMI. According to LeCun, the project had been his primary focus for years and enjoyed support from Meta leadership, including Zuckerberg.

The corporate shuffle comes at a pivotal moment for the AI industry. OpenAI, Anthropic, Google DeepMind, Meta and xAI are spending billions of dollars building ever-larger language models in pursuit of AGI. Many executives argue that scaling today’s systems will eventually produce machines capable of performing most intellectual tasks at human levels.

LeCun disagrees.

He argues that while language models have already achieved superhuman performance in certain narrow domains such as translation and portions of software coding, they fundamentally lack the ability to understand how the physical world works.

An alternative approach

Instead, LeCun has spent years developing an alternative approach known as Joint Embedding Predictive Architecture, or JEPA.

Unlike generative AI systems that attempt to predict every detail of future data, JEPA seeks to learn abstract representations of the world and predict how those representations evolve over time. The goal is to enable machines to anticipate the consequences of actions, a capability LeCun considers essential for true intelligence.

He contends that “you cannot build a reliable agentic system without this system having the ability to anticipate the outcome and the consequences of its own actions. … We (humans) are certainly capable of anticipating the outcome resulting from our actions, and that’s why it allows us to plan a sequence of actions to accomplish a task.”

The concept stems from LeCun’s long-standing criticism of generative AI. He argues that predicting every pixel in a video frame or every word in a sentence becomes mathematically intractable when dealing with complex real-world environments. Instead, machines should focus on learning higher-level abstractions that capture the underlying structure of reality.

The approach draws from decades of LeCun’s research. Before the current AI boom, he helped develop convolutional neural networks, which became foundational to modern computer vision. In 2018, he shared the Turing Award — often called the Nobel Prize of computing — with fellow AI pioneers Geoffrey Hinton and Yoshua Bengio for their contributions to deep learning.

Meta’s AI strategy has evolved significantly in recent years. While the company initially became a champion of open-weight models through its Llama releases, competition from OpenAI, Anthropic and Google intensified pressure to prioritize commercial products and rapid deployment.

LeCun has long advocated for open-source AI and played a role in internal debates surrounding the release of Meta’s Llama models. He credited the open release of Llama 2 with helping launch an ecosystem of startups and developers around the world.

Advocating for Project Tapestry

Today, he warns that concentrating advanced AI capabilities in the hands of a few companies risks limiting innovation and cultural diversity.

That concern has expanded into another initiative he discussed at VivaTech: Project Tapestry. The effort aims to create a globally distributed and open AI foundation model through federated training, letting nations, universities and organizations contribute knowledge without sharing underlying data.

LeCun argues that AI sovereignty will become increasingly important as digital assistants mediate more of the world’s information. If future AI systems are controlled by only a handful of companies or nations, he said, cultural diversity and democratic access to information could suffer.

For now, however, his primary focus is proving that a different path to advanced intelligence exists.

The challenge remains enormous. LeCun acknowledged that developing systems capable of genuine reasoning, planning and understanding will take years. Despite rapid advances in AI, he dismissed predictions that superintelligent machines are imminent.

Instead, LeCun is attempting something increasingly rare in today’s AI industry: building a fundamentally new paradigm rather than scaling the existing one. Whether AMI Labs succeeds could help determine if the future of AI belongs to ever-larger language models — or to world models that learn how reality itself works.