From Imitation to Understanding: Will AI Build Its Own World Model?

Artificial intelligence has come a long way from simple algorithms to advanced models like LLM (Large Language Models). These impressive tools can generate texts, images, and even code, which are increasingly difficult to distinguish from those created by humans. But does this mean that AI really "understands" the world it describes?

Currently, LLM models not only imitate the result, but also the process of arriving at it. Thanks to techniques like "Chain-of-Thought prompting", models can generate intermediate reasoning steps, giving the impression of deeper understanding. It's like watching a model "thinking out loud". However, we are still talking about imitation, albeit on a much higher level – mimicking the human way of thinking and drawing conclusions.

For AI to truly understand the world, it would need something more than just advanced imitation – it would need a world model. What could such a model look like? Above all, it would be a structure containing representations of:

• Objects and their properties, relations between them, laws of physics, knowledge of the world, and finally – abstract concepts.

This is an ambitious task, because the world is extremely complex, and its full description is a gigantic challenge. How to represent knowledge in a way understandable to AI? How to model abstractions? What about continuous changes and the need for constant updating of the "world model"?

Despite the difficulties, scientists are actively working on creating a "world model" for AI, using different approaches:

• Knowledge Bases: Creating large, structured knowledge systems like WordNet, Cyc, or Wikidata.
• Knowledge Graphs: Representing knowledge in the form of graphs, where objects are nodes and relations between them – edges.
• Symbolic Representations: Integrating AI systems with symbolic logic and reasoning systems.
• Rule-Based Systems: Creating systems using logical rules to represent knowledge.
• Learning in Simulations: Training AI in simulated environments so it can build its own "world model" through interaction.
• Multimodal AI: Integrating data from different sources (text, image, sound) to create a more holistic representation of the world.

Our world model – consciousness, experience and constant update

Analyzing attempts to build a "world model" for AI, one cannot help but look at our own human way of understanding reality. Our internal world model is built on experience, senses, emotions and interaction with the environment. Importantly, we constantly update it, learning new things, verifying beliefs and adapting to change. Our consciousness allows us for introspection and reflection, which enables modifying the model through reasoning and analysis. This model also contains knowledge about ourselves, other people and abstract ideas. It is our consciousness that makes us able to dynamically build, modify and move in reality.

Both we and AI strive to create a representation of the world that allows for action and predicting the future. The difference is that our "world model" is based on consciousness, emotions, experience, introspection and is subjective, while current attempts to build this model for AI are based on data analysis, rules and logic. Our model updates constantly through interaction with the world, while AI still faces enormous challenges in this area.

Will AI ever reach the level of human understanding? We don't know the answer to this question. Today's attempts to build a "world model" for AI are just the first steps. We can still learn from our way of understanding the world and draw inspiration from human consciousness in search of new solutions.

Perhaps future artificial intelligence, based on a "world model", will not only be a master of imitation, but a true partner in discovering and understanding our universe.