AIの転換点てんかんてん：言語げんごから世界せかいモデルへ

Current AI research is shifting its focus from language and image generation models to world models.

This is a technology that simulates the real physical world and infers the environment, which is essential for applications such as robots and self-driving cars.

This change is clearly demonstrated by the fact that major technology companies like Meta and Google are focusing on development in this field.

Metas Chief AI Scientist, Yann LeCun, emphasized his intention to focus on research into world models.

World Labs, led by Fei-Fei Li, has widely released the Marble model.

Google is testing the Genie model, and Nvidia is developing the Omniverse and Cosmos platforms for AI that operates in physical spaces.

These efforts show that AI research is going beyond understanding text and 2D images, and is now taking on the challenge of physical simulation in 3D space.

Fei-Fei Li points out that the core of human cognition lies in spatial intelligence, and that this is the ability most lacking in modern AI.

Humans exist based on the laws of physics and spatial relationships, living and acting in the real world.

However, current AI only interprets the world based on symbolic symbols.

Autonomous vehicles are a practical application of this technology, but their operational scope is still very strictly structured.

In order to operate in more complex environments, AI needs to understand and simulate spatial interactions more extensively.

Efforts toward applying current world models reveal both future possibilities and technical challenges that need to be overcome.

For example, when the author used the marble model to recreate Van Gogh’s “Bedroom in Arles” in three dimensions, the result was an incomplete depiction.

The walls and furniture are blurry, and the details of the objects are lost.

Even if the world model can construct a consistent spatial structure from limited data, this shows that there are still limitations in retaining details and making accurate inferences in more complex situations.

To create an effective world model, it is necessary to overcome technical challenges that go far beyond what has been achieved so far in the field of AI.

To accurately reproduce physical space, AI must predict new states and understand vast amounts of data as well as context and causality.

Furthermore, AI needs the ability to retain information over time and maintain consistent behavior.

Not only are these complex technical challenges being solved, but new risks are also emerging.

If an AI system acts in the physical world without human instructions, the limitations of its understanding or simulations could cause harmful changes in the real world.

Therefore, the development of world models requires not only overcoming technical challenges but also establishing a framework to ensure that this technology is accepted by society in a safe and trustworthy manner.