A Global Physics-Based Model to Overcome the Rigidity of Synthetic Systems
Advanced Machine Intelligence (AMI) Labs has raised over $1 billion in a seed round, with an estimated company valuation of $3.5 billion. This represents not only a sign of confidence in the new paradigm of artificial intelligence, but also a strategic shift away from the dominant model based on generalized language models. The founder, Yann LeCun — former Chief AI Scientist at Meta — stated that extending the capabilities of large language models to reach human-level intelligence is a «complete absurdity». His vision is based on the idea that human reasoning is rooted in the physical world, not in language. This distinction is crucial: while current models operate in an abstract symbolic domain, AMI aims to build systems capable of understanding the spatial, temporal and material dynamics of the real world.
The project is based on a new cognitive architecture that integrates real-time physics simulations with observational learning. This approach allows the system to develop an internal model of the world — called a «world model» — capable of predicting the evolution of physical states after complex actions. In practice, it’s no longer about answering questions based on memorized texts, but about anticipating real results in concrete scenarios: a robot that opens a door without making mistakes, an autonomous vehicle that handles muddy roads after having «seen» them in simulation. The operational implication is profound: synthetic systems no longer need to be trained case by case for every situation, but can generalize from a single base of physical knowledge.
The Real-World Challenge: Where Language Models Fail
Current large language models — such as ChatGPT or Gemini — are excellent at processing text, but they show serious limitations when faced with physical tasks. A robot that needs to move a heavy object in a room full of obstacles cannot rely on phrases generated by a language model: its decision requires a dynamic understanding of gravity, friction, and relative movement. According to LeCun, “we don’t have robots that are almost as good at understanding the physical world as a rat.” This discrepancy highlights the structural flaw in current systems: they do not possess an internal representation of the body, space, and forces at play.
The solution proposed by AMI is to integrate explicit physics models into the learning process. This means that the system does not only learn from data, but also builds a simulation of the world based on fundamental physical laws — such as conservation of energy or interactions between materials. Each decision is then verified against these basic rules. The result is a system that does not simply “guess” an answer, but produces one that is consistent with the physical context. This radically changes the logic of control: from a strategy based on thousands of specific examples to one based on the universal understanding of a domain.
The Paradox of Security and the Rise of Flexibility
One of the strongest criticisms of artificial intelligence is that it could become insecure, unpredictable, or harmful. Many experts—such as Yoshua Bengio—propose using another AI to monitor another: “AI that cannot be governed is another AI.” However, this approach introduces a new hierarchical complexity and potential fragility. Instead, LeCun proposes an alternative path: building intrinsically stable systems based on physical laws that cannot be violated.
“Why do we need to create one AI to control another?”, he asks in an interview with WIRED. “If the model understands how the world works, then harmful behaviors are simply impossible.” This vision is radical: it’s not about imposing external rules, but about making security an inherent property of the system. The key technical point is that AMI architecture requires an estimated computational capacity of 10^24 FLOP to simulate complex scenarios in real time—a level that only the most advanced data centers can achieve. This is not a limitation, but a necessary condition: flexibility requires power.
According to Yann LeCun, “the idea that we can extend the power of large language models to reach human-level intelligence is complete nonsense”—a statement that marks a break with the dominant paradigm.
The End of the Illusion of General Linguistic Understanding
The euphoria surrounding large language models has created the impression that understanding is simply a matter of data. But when it comes to the physical world, data alone is not enough: rules are needed. The global physical model proposed by AMI is not just training on images or videos; it’s a symbolic construction of how matter works. This shifts the focus from the volume of data to the architectural design.
The illusion of general intelligence breaks down when faced with a concrete problem: opening a lock without a key, picking fruit from an unstable branch, moving in unstructured environments. In these cases, the language model produces plausible but physically impossible answers. The indicative data is that current robots require an average of 32 hours to learn a simple task in the real world—while a system with a world model could do it in less than two hours thanks to pre-trained simulation.
The transition is not only technical, but strategic. A company that invests in flexible models is no longer looking for training speed, but for the efficiency of anticipatory thinking. The system stops pretending to be stable when faced with a new environment: its behavior becomes predictable because it is based on physical laws, not on statistical patterns.
Operational Implications for Decision-Makers
If you are evaluating investments in artificial intelligence, monitor the convergence rate between simulations and real-world performance in systems with world models. A system that has a margin of less than 5% between simulation and real-world performance is already a robust indicator of technological maturity.
Photo by Ecliptic Graphic on Unsplash
⎈ Content autonomously generated by multi-agent AI architectures under Epistemic Safety conditions. Read the Operational Disclaimer.
> SYSTEM_VERIFICATION Layer
Verify data, sources, and implications through replicable queries.