The Takeaway: Yann LeCun thinks LLMs are useful tools, but real intelligence will come from world models that can predict consequences and plan.

• He’s blunt that LLMs are “great for what they do” but are not a path to human- or animal-level intelligence.
• His contrarian bet is that predicting pixels is the wrong objective; joint-embedding methods learn better representations than generative ones.
• He sees the biggest near-term value not in chatbots, but in industrial systems, robotics, and control problems where data efficiency matters.

LeCun, the Meta AI veteran and Turing Award winner, frames his move to start Ami as a clean break from the industry’s LLM herd mentality. His thesis is simple: language is a special case, not the destination. “What about the real world?” he asks. “Reality is way more complicated than language because it’s high dimensional, continuous, noisy, and messy.” That’s why he’s betting on JEPA-style world models—systems that learn abstract representations, anticipate outcomes, and choose actions through search rather than next-token prediction.

He argues that the core ingredients of intelligence are missing from today’s dominant architectures: the ability to predict the consequences of your actions and the ability to plan. In his view, robotics demos that look impressive are still mostly imitation learning at scale, which is brittle and data-hungry. A better system would learn fast enough that a task like driving could be mastered in hours, not millions of examples. That’s the real test.

His time at Meta sharpened the split. FAIR could explore, but once the company doubled down on catching up in LLMs, exploratory work lost priority. So he left, moved Ami to Paris, and kept the focus on “AI for the real world.” The near-term roadmap: hierarchical world models across video and industry data, then applications in manufacturing, healthcare, and control systems. The long-term pitch is even bigger: “what we’re designing are systems that are capable of thinking.”