Artificial intelligence has made remarkable progress over the past few years, but according to Demis Hassabis, CEO of Google DeepMind, today’s AI systems are still far from achieving true intelligence. Despite the rapid evolution of large language models (LLMs) and multimodal AI systems, Hassabis believes that modern models lack several critical capabilities needed to understand and interact with the real world effectively.
Speaking about the limitations of current AI, the Nobel Prize–winning researcher highlighted gaps in long-term planning, continual learning, and advanced reasoning skills—areas that remain largely unresolved even in the most advanced AI models available today.
Why Current AI Models Fall Short
Modern foundational AI models, such as Google’s Gemini 3, can process text, images, audio, and video with impressive accuracy. However, Hassabis argues that these systems do not truly understand the world. Instead, they rely on pattern recognition rather than genuine comprehension.
According to him, today’s LLMs struggle with:
- Understanding physics and causality
- Predicting how actions affect outcomes over time
- Learning continuously without retraining
- Planning across long time horizons
In practical terms, this means that while AI can generate convincing responses, it often fails when asked to reason about real-world scenarios that involve cause-and-effect relationships or long-term consequences.
The Case for ‘World Models’ in AI
To overcome these shortcomings, Hassabis advocates for the development of world models—a new class of AI systems designed to understand how the world works at a deeper level. World models aim to simulate reality by learning the underlying rules of physics, environment dynamics, and causality.
Unlike traditional LLMs that are trained primarily on vast datasets of human-generated content, world models would:
- Build internal representations of the physical world
- Predict outcomes based on actions and environmental changes
- Enable better decision-making and long-term planning
- Learn continuously as they interact with new data
Hassabis believes that world models are essential for creating AI systems that can move beyond surface-level intelligence and achieve more robust reasoning abilities.
Industry Leaders Agree: World Models Are the Future
Demis Hassabis is not alone in this belief. Yann LeCun, former Chief AI Scientist at Meta and one of the most influential figures in AI research, has also emphasized the importance of world models as the next frontier in artificial intelligence.
In December 2025, LeCun announced the launch of his new startup, Advanced Machine Intelligence (AMI), which is dedicated to developing world models. The goal of AMI is to create AI systems that can learn like humans and animals—through observation, prediction, and interaction with the environment.
This shared vision among AI pioneers suggests a broader industry shift away from purely language-based models toward systems that can reason about the real world.
A Philosophical Divide on General Intelligence
Despite their agreement on world models, Hassabis and LeCun differ significantly in their views on general intelligence.
LeCun has publicly argued that the concept of general intelligence—as it is commonly used to describe human-level intelligence—is flawed. According to him, human intelligence is not truly general but highly specialized, shaped by biology, evolution, and experience. From this perspective, building a single, general-purpose AI that matches human intelligence may be neither realistic nor necessary.
Hassabis, on the other hand, continues to explore broader definitions of intelligence, focusing on systems that can adapt, reason, and plan across a wide range of domains.
Implications for the Future of AI
The debate highlights a crucial moment in AI development. While current models excel at language generation and pattern recognition, experts agree that true intelligence requires deeper understanding. World models could be the missing piece that enables AI to operate reliably in complex, real-world environments—such as robotics, scientific discovery, and autonomous systems.
If successful, this approach could redefine how AI systems are built, shifting focus from massive datasets to learning mechanisms inspired by human cognition and the physical world.
Final Thoughts
Demis Hassabis’ comments serve as a reminder that despite impressive advancements, AI still has fundamental limitations. The push toward world models represents a bold attempt to address these gaps and move closer to more capable, adaptable, and intelligent systems.
As leading researchers like Hassabis and LeCun invest in this vision, the next phase of AI development may not be about bigger language models—but about smarter systems that truly understand the world they operate in.
