In a recent podcast organized by CNBC, Google DeepMind CEO Demis Hassabis pointed out three major limitations of modern AI, including: long-term planning capabilities, continuous learning and deeper reasoning.
Mr. Hassabis emphasized that multi-modal platform models like Google's Gemini 3 can process text, images and videos, but still mainly rely on identifying and synthesizing data samples, instead of understanding the rules that govern the world.
According to Hassabis, LLM (big language model) currently does not grasp visual physics, causality, as well as how an action can lead to different results over time. This makes it difficult for AI to predict, invent, or explain completely new phenomena.
Mr. Hassabis said that, to date, AI systems still do not have the ability to self-create and run simulations within themselves to test hypotheses, just like how excellent scientists still do.
From there, Hassabis bet on a development direction that he calls "world models". These are AI systems designed to build a relatively accurate demonstration of how the world works, starting from visual physics, but not stopping there.
According to him, to truly understand and be creative, AI needs to integrate knowledge from many fields such as biology, economics and social sciences, instead of just learning from language data.
This view is not only pursued by Hassabis. Yann LeCun, former Director of AI science at Meta and one of the big names in the machine learning industry, also considers the world model as the next frontier of artificial intelligence.
In December 2025, LeCun announced the establishment of Advanced Machine Intelligence (AMI) startup, focusing on developing AI systems capable of understanding and simulating the real world.
However, disagreements still exist between Hassabis and LeCun, especially surrounding the concept of general intelligence. LeCun argues that general intelligence does not actually exist, because human intelligence is the result of high specialization. According to him, people considering themselves general intelligence is just an illusion.
In response, Hassabis refuted this view and argued that LeCun had confused "general wisdom" and "universal wisdom".
In a post on social network X, Mr. Hassabis emphasized that the human brain is the most sophisticated and complex phenomenon ever known, with the ability to apply knowledge flexibly in many different fields.
The debate between the two leading scientists shows that AI is still in the stage of finding the optimal development path.
Although there are still many opposing views, both Hassabis and LeCun agree on one point: to get closer to human-level intelligence, AI needs to go beyond the limits of current LLM and learn to understand the world, not just simulate language.
