Current AI systems such as ChatGPT, Claude, and Gemini excel at coding, mathematics, and generating text, but they are fundamentally unable to handle the messy complexity of the physical world. According to Yann LeCun, one of the leading figures in artificial intelligence, these models will never achieve human-level — or even animal-level — intelligence because they cannot process real-world data. "They're not a path towards human level or human-like intelligence, or even animal-like intelligence, because they cannot deal with real world data, they just are not built for that," he said at VivaTech, France's premier technology conference.
To address this gap, LeCun left his role as chief AI scientist at Meta in 2025 and founded Advanced Machine Intelligence Labs (AMI Labs), based in Paris. The startup is developing a new type of artificial intelligence not based on the technology behind large language models (LLMs).
The limits of large language models
LeCun acknowledges that LLMs like ChatGPT are highly effective for well-defined, predictable tasks such as coding, mathematical problems, and text generation. However, he argues that these systems "basically just accumulate knowledge" and "can regurgitate something, you train them to regurgitate, but they're not particularly smart. They don't have an underlying understanding."
In the real world, actions have a bewildering array of possible outcomes. LeCun illustrates this with a simple example: holding a pen upright on its tip and letting it go. Even a toddler knows the pen will topple over, but no human would try to predict which direction it falls — there is no way to tell. An LLM, however, would attempt to generate a single prediction based on statistical patterns in its training data, which would almost certainly be wrong because the system is not reasoning about physical reality.
A new architecture: Joint Embedding Predictive Architecture
AMI Labs' solution is a system called Joint Embedding Predictive Architecture (JEPA). Unlike LLMs, JEPA creates abstractions of the real world that allow it to assess the outcomes of actions. These abstractions filter out useless information, leaving the AI with only useful pictures of the world. In the pen example, JEPA would recognize that there is no point in trying to predict the direction of the fall, saving computational effort and avoiding incorrect predictions.
LeCun stated that the robotics industry has invested billions of dollars in building humanoid robots, yet training them to safely perform household tasks like ironing or stacking dishes remains difficult and costly. He is blunt about the limitations of current AI in this domain: "LLMs are largely hopeless." JEPA is designed to provide the flexible intelligence that robots need to navigate and operate in the real world.
Funding and industry context
AMI Labs announced earlier this year that it had raised more than $1bn (£760m) in a seed funding round — one of the largest of its kind in Europe. Investors include US computer chip giant Nvidia and the fund that manages the private wealth of Amazon founder Jeff Bezos. The round underscores investor belief in LeCun's vision for a new AI paradigm.
The following table summarizes the key differences between conventional LLMs and JEPA as described by LeCun:
| Aspect | Large Language Models (LLMs) | Joint Embedding Predictive Architecture (JEPA) |
|---|---|---|
| Core ability | Statistical pattern matching on text/code | Abstractions of physical world outcomes |
| Real-world handling | Cannot deal with real-world data | Designed for real-world unpredictability |
| Example: falling pen | Tries to predict direction (likely wrong) | Knows direction is unpredictable, avoids prediction |
| Robotics application | “Largely hopeless” | Aims to enable safe, flexible physical tasks |
| Training data requirement | Massive text corpora | Learns from abstract representations |
LeCun's departure from Meta and the founding of AMI Labs signals a strategic shift in AI research, moving away from scaling LLMs toward architectures that can comprehend and interact with the physical world. While the immediate applications are in robotics, the technology could eventually influence any domain requiring real-world reasoning — from autonomous vehicles to manufacturing automation. For now, the race is on to build an AI that is at least as smart as a rat.
Sources: