In my previous post, I explored the tremendous energy consumption of Large Language Models (LLMs), emphasising that the root cause of their inefficiency lies in their structure and fundamental mathematics, not in the incompetence of data centre engineers to prioritise sustainability.
The core problem with these powerful LLMs is their incapacity to create and sustain meaningful, understandable, and dynamically evolving models of the world. Apple acknowledged this in June 2025 in their paper titled “The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models through the Lens of Problem Complexity.”
A world model, also referred to as a cognitive model, is an information-processing framework utilised by a system—whether that be a machine, a human, or another animal—to monitor events occurring in its environment. While world models may not always be entirely accurate or comprehensive, they are crucial for cognition in both humans and animals. Alan Turing created a dynamic world model for his chess program, known as Turochamp, which we have previously discussed in this blog. This pioneering chess-playing AI was developed in 1949, even before Turing had the necessary hardware to implement it.
The idea of (world) models was no less central to the thinking of Nobel Laureate and one of AI’s Papst Herb Simon. His memoir, “Models of My Life”, illustrates how the creation and manipulation of internal models is at the heart of intelligent behaviour—whether in people or in machines.
In the 1950s, Simon, together with Allen Newell, designed the General Problem Solver (GPS), one of the first computer programs to simulate human problem-solving. GPS worked by creating an abstract model of the task (for instance, the Tower of Hanoi puzzle), representing all possible moves and the rules connecting them. The GPS could then “think ahead,” visualising or simulating the steps needed to reach a solution—much like how a person mentally works through steps before acting.
Ernest Davis and Gary Marcus stressed the central importance of world (cognitive) models in their 2019 book “Rebooting AI”, in an example of what happens in the human mind as one understands a simple children’s story.
Yann LeCun, Chief AI scientist of Meta/ Facebook, has stressed the need for world models and implied that prior to building human-like AGI, we must shoot for the intelligence of a mouse and a cat. Jürgen Schmidhuber was perhaps the first to emphasise the need for world models in the context of neural networks.
Lack of reasoning and cognition in LLMs makes it even more incredible that some AI builders and media talk about the necessity to give LLMs a moral status and rights. I have already reported about an encounter with Artificial Consciousness GED – a Swiss-based nonsensical technology aiming at luring struggling coaches and consultants – and harming their clients – with a scam about ‘not built but emerged” conscious machine.
The pure scaling of LLMs implies adding more data and more parameters – won’t improve AI reliability and safety. It won’t solve hallucinations, remove mistakes. Scaling does not bring reasoning and understanding into these models, and it does not incorporate a world model into existing architectures.
Let’s review who is currently going beyond the traditional LLMs:
- LeCun has been an outspoken advocate for going “beyond LLMs,” promoting cognitive architectures that build internal world models (I-JEPA: Joint Embedding Predictive Architecture; focus on vision, perception, and cause-effect learning). There is currently funding from Meta, though LeCun is leaving the company to build technology based on a world-model approach. He will rely on angel investing and Venture Capital to move forward;
- Demis Hassabis (CEO), Oriol Vinyals, and the Genie project team at DeepMind are trying to implement world model agents that learn interactive environments from video—not just still images or text—and can generate/simulate physics-rich 3D worlds based on prompts. The funding is from Alphabet, DeepMind’s parent company.
- Yoshua Bengio is working on LawZero, a non-agentic world model architecture for safe AGI. He received $30M in philanthropic backing (Schmidt Sciences, Open Philanthropy, Future of Life Institute). Besides, Bengio is leading/participating in large international consortia (including UK ARIA—£59M over four years for “safeguarded AI” world modelling and safety initiatives);
- National University of Singapore (NUS) is doing research on Brain-JEPA, a world model brain dynamics foundation model (fMRI/vision data);
- Astrus (Canada) is a startup building physics-aware world models for chip design, recently secured $8M in Series A;
- Swiss AI Initiative is working on a world-model-based AI, and is funded with grants (up to €1M+ per research group plus application domains) from Swiss and EU support bodies.