Yann LeCun on AGI, AI Agents & the Future of AI Research

The evolving landscape of artificial intelligence was a central topic of discussion at the recent Imagination in Action summit in Davos, Switzerland. A key conversation, led by interviewer John Werner, centered on the current state and future trajectory of AI with leading researcher Yann LeCun. The discussion revealed a critical reassessment of current approaches and a potential shift in the field’s focus.

Getting Realistic About AGI

LeCun challenged the conventional understanding of Artificial General Intelligence (AGI), suggesting the term itself is a misnomer. He prefers “human-level intelligence,” and while acknowledging progress, he doesn’t anticipate achieving this level of AI in the immediate future. “We need a few conceptual breakthroughs,” he stated.

A core argument presented was the disparity between the intellectual capabilities of Large Language Models (LLMs) and the practical, real-world knowledge humans possess. LLMs, while capable of complex intellectual tasks, lack the “street smarts” necessary to navigate the physical world effectively. LeCun emphasized that intelligent behavior requires anticipating consequences and planning actions – a capability he termed a “world model,” which is currently missing in existing AI systems.

Did You Know? Despite millions of hours of training data, fully autonomous vehicles (Level 5 capability) remain elusive, highlighting fundamental architectural limitations in current AI approaches.

This perceived deficiency is driving a predicted “physical AI revolution,” focused on systems that can understand and interact with the messy, high-dimensional data of the real world – video, sensor data, and more. These systems would need to build predictive models of their environment and plan actions accordingly.

But What About AI Agents?

LeCun expressed skepticism about the current trend of building AI agents on top of LLMs, calling the approach “a disaster.” He questioned how a system can plan effectively without the ability to predict the outcomes of its actions. The need for a “world model” remains central to achieving true intelligence, regardless of the application.

Advanced Machine Intelligence

LeCun recently launched Advanced Machine Intelligence, a company focused on building systems leveraging these “world models.” The goal is to create AI capable of intelligent action based on a comprehensive understanding of its environment. Prototypes already demonstrate an ability to understand and even identify impossibilities within video data – for example, recognizing when a ball defies gravity.

The foundation of this work is JEPA (Joint Embedding Predictive Architecture), a methodology still under development. The aim is to generalize this approach to various data types and sensors, enabling the creation of models for complex systems ranging from industrial processes to living cells.

Expert Insight: LeCun’s emphasis on “world models” represents a significant departure from the current focus on LLMs, suggesting a need for fundamentally different architectural approaches to achieve more robust and adaptable AI.

Digital Twinning, and the LaPlace Demon

LeCun distinguished his approach from simple digital twinning, suggesting a need for abstraction rather than complete simulation. He likened understanding a complex system to utilizing psychology and science, rather than attempting to model it at the level of quantum physics.

Is Alignment the Right Frame?

The discussion also touched on AI alignment – the effort to ensure AI systems act in accordance with human values. LeCun suggested that the framing of this issue may be misguided if we assume future AI will resemble current LLMs. He posited that objective-driven, world-responding systems may present fewer alignment challenges.

The Digital Commons

LeCun advocated for open systems and open research in AI development, emphasizing the importance of a “consortium” approach over proprietary walled gardens. He noted the innovation occurring within open-source models, even citing examples from China. He argued for a “bottom up” approach, contrasting it with traditional industrial hierarchies.

Some Challenges, and Solutions

LeCun identified the concentration of power as a key risk, warning that a handful of companies controlling AI systems could threaten democracy, cultural diversity, and linguistic diversity. He reiterated the need for a diverse AI landscape, achievable through open-source development. Examples of this collaborative spirit are already emerging at institutions like Drexel University and MIT.

Frequently Asked Questions

What is a “world model” in the context of AI?

According to LeCun, a “world model” is a system’s ability to anticipate what will happen in the world and predict the consequences of its actions, enabling it to plan and achieve objectives. It’s a capability currently lacking in most AI systems.

Why is LeCun critical of building AI agents on top of LLMs?

LeCun believes that without the ability to predict consequences, AI agents built on LLMs cannot effectively plan sequences of actions and therefore will not achieve human-level intelligence. He considers this approach “a disaster.”

What is Advanced Machine Intelligence aiming to achieve?

Advanced Machine Intelligence is focused on building systems that can work intelligently using “world models,” allowing them to understand and interact with the real world in a more sophisticated and adaptable way.

As AI research continues to evolve, will a shift towards “physical AI” and open-source collaboration prove to be the key to unlocking truly intelligent systems?