Artificial intelligence researchers are increasingly exploring alternatives to today’s large language models (LLMs), arguing that current systems excel at language generation but remain fundamentally limited when interacting with the physical world.
Among those advocating a different direction is Yann LeCun, who said existing AI systems such as ChatGPT, Claude and Gemini are effective for tasks including coding, mathematics and text generation but are not designed to achieve human-like or even animal-level understanding of real-world environments.
Speaking during the VivaTech conference, LeCun said these systems primarily learn statistical relationships from vast datasets rather than developing an understanding of how the physical world behaves. He argued that this limitation makes them unsuitable for many robotics applications that require flexible reasoning and adaptation.
LeCun, who previously served as chief AI scientist at Meta before leaving in 2025, now leads Paris-based Advanced Machine Intelligence Labs (AMI Labs), where researchers are developing an alternative AI architecture known as Joint Embedding Predictive Architecture (JEPA).
Earlier this year, AMI Labs announced it had raised more than $1 billion in seed funding. The investment included backing from Nvidia and a fund managing the private wealth of Jeff Bezos, making it one of Europe’s largest seed funding rounds.
Why Researchers Want AI Beyond Language Models
LeCun argues that LLMs are highly capable within structured and predictable domains because they generate responses based on patterns learned during training. However, he said those systems do not possess an underlying model of physical reality.
To illustrate the difference, he described balancing a pen on its tip. While a person instinctively knows the pen will fall without needing to predict the exact direction, an LLM may attempt to generate a statistically likely outcome rather than reasoning about the uncertainty of the situation itself.
According to LeCun, JEPA seeks to address this challenge by creating abstract representations of the physical world. Rather than predicting every possible outcome, the system is intended to identify which information is meaningful while ignoring unnecessary detail, allowing it to reason more efficiently about real-world situations.
■ Related News
He said this capability could eventually make AI systems better suited for robotics, where machines must continually interpret changing environments instead of responding to fixed prompts.
Robotics Continues to Drive AI Research
Improving AI reasoning has become a significant objective for robotics developers, who have invested billions of dollars in humanoid machines capable of operating in human environments.
Although robotic hardware has advanced rapidly, teaching robots to perform everyday household activities such as loading dishwashers or ironing clothing safely remains technically difficult and expensive.
LeCun said current LLM-based approaches are unlikely to solve these challenges effectively because they are not built to interpret complex physical interactions.
World Models Gain Momentum Across the Industry
Other researchers share the view that future AI systems will require more sophisticated reasoning than today’s language models provide.
Ingmar Posner, who directs the Applied AI Lab at the University of Oxford and also serves as an Amazon Scholar, said future AI systems should be capable of explaining cause and effect, identifying what matters in a situation and evaluating alternative actions.
His research group has spent several years developing what he describes as a mechanistic world model, designed to organize knowledge so that information can be efficiently retrieved, combined and modified when solving problems.
World models have existed as a research concept for decades but received renewed attention following influential work published in 2018 by David Ha and Jürgen Schmidhuber. Their research proposed that AI systems could learn by building internal simulations of the world rather than relying solely on memorized patterns.
Since then, several organizations have expanded work in the field. Google has developed Dreamer world models, including a version that learned to collect diamonds in the video game Minecraft by imagining future scenarios during decision-making.
Additional research includes Genie from Google DeepMind, Gaia from Wayve, and work at World Labs, founded in 2023 by Fei-Fei Li.
Commercial Deployment Remains Ahead
Posner said it remains difficult to predict how quickly these newer AI architectures will mature, noting that the rapid arrival of generative AI systems surprised many researchers who had expected such capabilities to take decades longer.
LeCun said AMI Labs plans to continue refining its JEPA-based system through the remainder of this year, with the goal of introducing initial industrial deployments next year if development progresses as planned.
Looking further ahead, he said broader-purpose AI systems capable of performing many tasks with limited additional training remain the long-term objective. Even if those systems eventually exceed human capabilities in certain areas, LeCun argued that people will continue to play the central role in defining goals, asking questions and directing how AI is applied.
Tags: Artificial Intelligence, World Models, Yann LeCun, AMI Labs, Large Language Models, Robotics, Machine Learning, ChatGPT, Google DeepMind, Nvidia, JEPA, AI Research
