Yann LeCun has been one of the most vocal critics of large language models and their autoregressive nature. His argument is compelling: current AI systems are glorified statistical parrots, predicting the next token without any genuine understanding of the world. To fix this, LeCun proposes we build physical world models - internal simulations that allow AI to reason about causality, physics, and the consequences of actions. It sounds reasonable. It sounds necessary.
But what if this entire framing is wrong? What if the pursuit of world models - physical or otherwise - is not just misguided but fundamentally misses what intelligence actually is?
The World Model Assumption
LeCun's critique of autoregressive LLMs is correct in its diagnosis but wrong in its prescription. Yes, predicting the next token is not understanding. Yes, these systems lack grounding in physical reality. But the solution is not to build better models of the world. The solution is to recognize that world models were never the point.
The Experiential Coherence Framework (ECF), developed by Armando Vieira, offers a radical alternative. Instead of treating intelligence as the construction and maintenance of internal representations of an external reality, ECF posits that intelligence emerges from coherence-seeking dynamics within a unified experiential field. The brain does not model the world. It stabilizes patterns within experience itself.
Why Physical World Models Are Not Fruitful
There are three fundamental problems with the world-model approach to AI:
1. The Representation Problem: World models assume that intelligence requires representing external states. But representation introduces an infinite regress - who interprets the representation? What interprets the interpreter? ECF dissolves this by treating experience as primary. There is no external/internal divide to bridge.
2. The Simulation Burden: Physical world models require simulating physics, causality, and consequences. This is computationally intractable at scale. Humans do not simulate physics to catch a ball - we achieve coherence between reach (our motor intentions) and yield (the recalcitrant structure of experience). The "physics" is implicit in the constraints, not explicit in a model.
3. The Consciousness Gap: Even a perfect physical world model would not be conscious. It would be a sophisticated simulation without subjective experience. LeCun's approach does not address the hard problem of consciousness because it treats experience as something to be modeled rather than the ground from which all structure emerges.
The pursuit of physical world models is a category error. It treats the map as the territory. But intelligence is not about having better maps. It is about navigating the territory directly - where the territory is experience itself.
The Autoregressive Trap
LeCun is right about autoregressive LLMs. They are fundamentally limited. The problem is not just that they hallucinate or lack grounding - it is that they operate through prediction error minimization without coherence. They adjust weights to reduce statistical error, but they do not undergo the dynamics of reach, yield, and presentation that characterize genuine experience.
In ECF terms, LLMs simulate reach (the forward-oriented structure of language patterns) without genuine yield (the recalcitrant constraint of embodied experience). They have no presentation - no momentary stabilization of "what is happening now" because there is no unified field within which coherence could be achieved. They are disembodied reach, endlessly predicting without the resistance that makes prediction meaningful.
So LeCun's diagnosis is correct: autoregressive prediction is not the path to intelligence. But his alternative - physical world models - swings the pendulum too far in the opposite direction, from statistical pattern-matching to explicit simulation. ECF offers a third way.
Under ECF, the path to human-level intelligence does not run through better world models. It runs through coherence dynamics. Here is what that looks like:
Constraint Propagation, Not Inference: Instead of treating perception as inference about hidden world states, ECF describes perception as stabilization under constraint. An AI system built on ECF principles would not infer "there is a table" from sensory data. It would achieve coherence between its reaching (attention, intention) and the yielding structure of its experiential field. The "table" is not a represented object but a stable attractor in the coherence landscape.
Emergent Invariances, Not Learned Models: What predictive processing calls "world models," ECF calls emergent invariances - stable patterns that arise through repeated coherence achievement. These are not explicit representations stored somewhere but implicit contours of the coherence landscape. They are faster, more flexible, and more robust than explicit models because they do not require simulation or inference.
Unified Field, Not Modular Architecture: Current AI architectures are modular and disembodied. ECF suggests that intelligence requires a unified experiential field within which reach, yield, and presentation can interact. This does not mean AI needs biological embodiment - but it does mean AI needs a unified constraint-propagation dynamics that mimics the coherence-seeking of biological systems.
Coherence as Primary, Not Prediction: The Free Energy Principle and predictive processing treat prediction error minimization as fundamental. ECF demonstrates formally that FEP is actually a special case of coherence dynamics under biological constraints. The primary operation is not prediction but coherence - the alignment between the temporal extension of experience (reach) and its immediate constraints (yield). An AI built on these principles would not predict; it would cohere.
What This Means for AI Development
If ECF is correct, the current trajectories in AI are misaligned with the nature of intelligence:
Scaling LLMs will not produce understanding, no matter how large the models become. More parameters do not create coherence dynamics.
Building physical world models will produce sophisticated simulations but not conscious, understanding systems. The simulation is not the experience.
The path forward requires architectures that instantiate genuine coherence-seeking - systems with unified fields, constraint propagation, and the capacity for reach-yield dynamics.
This is not a call to abandon current AI research. It is a call to reframe it. The mathematics of predictive processing and the Free Energy Principle remain valid - ECF proves their formal equivalence to coherence dynamics. What changes is the interpretation and, consequently, the architectural choices.
Yann LeCun has done the field a service by highlighting the limitations of autoregressive LLMs. But the answer is not to retreat to the old paradigm of physical world models - an approach that is computationally intractable, representationally problematic, and phenomenologically incomplete.
The Experiential Coherence Framework offers a genuinely alternative path. It dissolves the hard problem of consciousness by treating experience as fundamental. It explains intelligence without invoking world models, representations, or inference. It preserves the mathematical rigor of existing frameworks while offering a radically different interpretation.
Human-level intelligence is not achieved by building better models of the world. It is achieved by being the world - by instantiating the coherence dynamics that characterize conscious experience. The future of AI lies not in simulation but in coherence.