Artificial intelligence has made incredible strides in recent years, from chatbots that can hold conversations to systems capable of solving complex mathematical problems. Yet, despite these achievements, one of the most prominent figures in AI research, Yann LeCun, argues that today’s models are still far from being truly intelligent. His work at Meta and now his new venture, Advanced Machine Intelligence Labs (AMI Labs), is focused on building a more flexible form of artificial intelligence—one that can understand and interact with the physical world as effectively as humans or even animals.

The Limitations of Large Language Models

LeCun has long been critical of the current state of AI development. While large language models like ChatGPT, Claude, and Gemini have become household names, he believes they are not truly intelligent. These systems excel at tasks such as coding, answering questions, and generating text, but they lack a deeper understanding of the world around them.

“LLMs are not smart,” LeCun explains. “They accumulate knowledge, but they don’t understand it. They can regurgitate information, but they don’t reason about the real world.”

This is a key distinction. Large language models operate on statistical patterns derived from their training data. When asked to predict what happens when a pen is dropped, for example, an LLM might generate a statistically plausible answer based on previous text it has seen. However, this prediction lacks any true understanding of physics or cause and effect.

A New Approach: Joint Embedding Predictive Architecture

To address these limitations, LeCun’s team at AMI Labs is developing a new type of AI architecture called Joint Embedding Predictive Architecture (JEPA). This system is designed to create abstractions of the real world that allow it to assess the outcomes of actions in a more flexible and intelligent way.

“JEPA is built to understand how things work in the physical world,” LeCun says. “It creates models of the environment that let the AI reason about what will happen next.”

This approach involves complex mathematical operations, but essentially, JEPA filters out irrelevant information and focuses on key elements of the world. For instance, when presented with a scenario like a pen falling, the system would recognize that predicting the exact direction is unnecessary—what matters is understanding that the pen will fall.

The Robotics Industry’s Need for Flexible AI

The demand for more flexible AI is not just theoretical. It has real-world applications in fields such as robotics, where current systems are struggling to perform everyday tasks like ironing clothes or stacking dishes.

“LLMs are largely hopeless for robotics,” LeCun says. “They can’t understand the physical world, so they can’t safely navigate it.”

This is a major challenge for the robotics industry, which has invested billions in developing humanoid robots capable of performing increasingly complex tasks. However, training these systems to operate in real-world environments remains difficult and expensive.

LeCun argues that without a more flexible form of AI, progress in this area will be limited. “We need models that can reason about the world, not just memorize patterns,” he says.

The Rise of World Models

Other researchers are also exploring alternative approaches to building more intelligent AI systems. One promising direction is the development of World Models, a concept that has been around for decades but has gained renewed interest in recent years.

A World Model is essentially an AI system that builds a mental simulation of its environment, allowing it to predict what will happen next and make decisions based on that understanding. This approach is fundamentally different from large language models, which rely on statistical patterns rather than causal reasoning.

One notable example is the Dreamer World Model, developed by Google. This system was able to learn how to collect diamonds in the video game Minecraft by imagining future scenarios and using them to guide its actions.

Professor Ingmar Posner of Oxford University, who leads a team working on similar systems, describes this as a “mechanistic world model.” His team is developing AI that can organize knowledge in a way that allows it to be recalled, combined, and modified when needed.

“World Models are the next step in AI development,” Posner says. “They allow systems to understand cause and effect, which is essential for real-world applications.”

The Road Ahead: Challenges and Opportunities

Developing these new forms of AI is a long-term endeavor. While LeCun’s team at AMI Labs plans to refine their model this year and begin testing it in industrial settings next year, the path to full deployment remains uncertain.

“Building flexible AI is not just about scaling up existing models,” LeCun says. “It’s about rethinking how we design these systems from the ground up.”

Posner agrees that progress will be slow but significant. He notes that even a system like ChatGPT took years of research and development to reach its current capabilities.

“Predicting when we’ll see truly intelligent AI is difficult,” he says. “But what’s clear is that the next decade will be about building systems that can reason, explain, and interact with the world in meaningful ways.”

Conclusion

Yann LeCun’s work at AMI Labs represents a significant shift in the direction of artificial intelligence research. By moving beyond large language models and focusing on more flexible, physically aware AI systems, he is challenging the status quo and opening new possibilities for how machines can interact with the real world.

The development of World Models and other alternative approaches highlights the growing recognition that current AI systems are limited in their ability to understand and navigate complex environments. As researchers like LeCun and Posner continue their work, the future of AI is likely to be shaped by a new generation of models that can reason, adapt, and operate with greater autonomy.

For readers interested in the evolution of artificial intelligence, the coming years will offer many opportunities to watch how these innovations unfold—and what they mean for our relationship with technology.

---

Original Source

This article is based on publicly available reporting. For the complete original story, visit the publisher’s article.