The Great Divide in AI: Will Superintelligence Come from Language Models or "World Models"?

Artificial intelligence research has accelerated rapidly in recent years, leading to a growing debate among leading scientists in the field. This discussion is not merely technical; it also raises a fundamental question about the future direction of technological development.

It is believed that AI could one day reach human-level or even super-human intelligence. However, the scientific community is divided into two distinct approaches regarding which technological architecture will lead to this point.

This debate was recently reignited by a significant development. Advanced Machine Intelligence (AMI), a venture founded by Meta's former Chief AI Scientist Yann LeCun, announced it had raised $1.03 billion in funding.

The company's funding round reportedly valued it at $3.5 billion. LeCun's new venture is seen not just as another tech company, but as a harbinger of a new direction in AI research.

LeCun's goal is not to simply scale up the large language models that have dominated the tech world in recent years. Instead, he aims to work on next-generation AI architectures based on reasoning, planning, and understanding the environment. At the heart of this approach is a concept researchers call the “world model.

Two Different Paths in AI

Today, two main approaches stand out in AI research.

The first approach revolves around large language models, which have achieved significant success in recent years. Developed by companies like OpenAI, Anthropic, and Google, these systems are trained on massive text datasets and learn probabilistic relationships between words to generate text.

Systems like ChatGPT, Claude, and Gemini not only generate text but can also write software code, perform data analysis, and exhibit human-like performance in many different tasks. These developments have led some researchers to ask a crucial question: If these models continue to be trained with more data and more processing power, could they eventually lead to human-level AI?

The statement in OpenAI's GPT-4 technical report, "GPT-4 exhibits early signs of artificial general intelligence," highlights how seriously this approach is taken. According to this view, progress in AI largely relies on scaling; that is, larger models, larger datasets, and more powerful processing infrastructure can lead to the emergence of new capabilities.

Language is not intelligence

However, there is also a significant group of researchers who believe this approach has limitations. Yann LeCun, who served as Meta's Chief AI Scientist for many years, argues that despite their impressive appearance, large language models are fundamentally systems that predict text patterns.

According to LeCun, these systems often appear to reason but actually operate solely on probability calculations. He believes that human intelligence is not just about processing language; it is also about understanding how the world works.

Therefore, to develop truly powerful AI, machines need to learn not just text, but also the world. At the core of LeCun's approach is the concept of a “world model.”

What is a World Model?

A world model means an AI system can create an internal simulation of its environment. Such a system would not only make predictions from past data but could also simulate future scenarios and evaluate the consequences of different actions.

This approach is based on how the human brain works. Neuroscience research shows that the brain is a system that constantly predicts its environment and simulates possible outcomes.

Researchers advocating the world model approach believe that the next major leap in AI will occur in this direction.

Expert Opinion: Language Models Have Limits

Technology entrepreneur and Enquire AI CEO Cenk Sidar is one of those who believe large language models are limited for real-world analysis.

According to Sidar, these systems often appear smarter than they are.

“LLMs are, as the name suggests, language models. They predict word sequences and give the impression of reasoning.

It is not possible for them to provide guidance, especially in areas like politics, economics, and geopolitics. World models will be decisive in these areas, as well as in fields requiring physical senses like robotics.

According to Sidar, for AI systems to truly become decision support tools, they must not only generate text but also be able to model cause-and-effect relationships in the world.

DeepMind: The Future Could Be Hybrid

Google DeepMind CEO Demis Hassabis believes that the future of AI will emerge from a combination of both approaches. According to Hassabis, a powerful AI system must be able to both understand language and model the world.

Systems like AlphaGo, AlphaZero, and MuZero developed by DeepMind are seen as early examples of this approach. MuZero, in particular, attracted attention for its ability to learn its environment and develop strategies without prior knowledge of game rules.

Such systems do not just make predictions by looking at data but learn a model of the environment and run simulations on this model.

The Names Behind World Model Research

The world model approach is not just an idea advocated by Yann LeCun. Fei-Fei Li from Stanford University argues that AI needs to develop three-dimensional spatial representations to understand the world.

David Ha's 2018 paper titled “World Models,” published while he was at Sakana AI, is also considered one of the most important references in this field.

Jeff Hawkins, founder of Numenta, develops theories arguing that the primary function of the human brain is to predict the world. Turing Award winner Yoshua Bengio has also been working on causal learning in recent years, emphasizing that AI should learn not only data patterns but also cause-and-effect relationships.

How Will It Affect Professions?

The direction of AI development will directly impact the business world. Large language models are rapidly transforming professions that require information generation and analysis.

Many tasks in areas such as content creation, software development, marketing, data analysis, and customer service have begun to be automated. World model-based AI systems, on the other hand, could transform sectors more connected to the physical world.

In areas such as robotics, autonomous vehicles, logistics, manufacturing technologies, and defense systems, the impact of such systems could be much greater. In other words, while language models change the digital business world, world model-based AI could transform the physical economy.

The Next Stage of AI

Today, many researchers believe that future AI systems will not rely on a single architecture. Instead, more complex structures combining language models, world models, and autonomous agent systems may emerge.

Such an architecture could enable machines to both communicate with humans and make independent decisions by understanding their environment.

In the world of AI, the big question that still awaits an answer is: Will the path to superintelligence pass through larger language models, or through systems that can truly understand the world?

Yann LeCun's new billion-dollar venture indicates that this debate will remain at the center of the technology world in the coming years.

Because the outcome of this debate could determine the direction not only of artificial intelligence but also of the future economy and business world.