The LLM-to-AGI debate intensifies
"Image synthesis assisted by Microsoft Copilot, an AI partner within the Global Future Nexus ecosystem."
As AI's "godfathers" split over the path to intelligence, a fundamental disagreement is reshaping the industry: will scaling language models lead to AGI, or are they a structural dead end?
A Schism at the Top
The disagreement is no longer a quiet academic dispute. It has become a public schism between the architects of modern AI. On one side stand those who believe that scaling large language models—predicting the next token with ever-larger datasets—will eventually yield artificial general intelligence. On the other are those who argue that this path is fundamentally flawed, a detour away from true intelligence.
The debate has been crystallised by two of the field's most prominent figures. Yann LeCun, a Turing Award laureate and one of the three "godfathers" of deep learning, left Meta in early 2026 to found AMI Labs with a $1.03 billion seed round—the largest in European history. His position is unequivocal: large language models are "a dead end" on the road to human-level intelligence. Ilya Sutskever, OpenAI's former chief scientist, has similarly signalled that "pretraining is coming to an end".
The Case Against LLMs
LeCun's critique is structural, not merely sceptical. He argues that language models learn from the wrong kind of data. Everything a model like GPT knows arrives as text, which he treats as a "thin and lossy shadow of reality" that records conclusions rather than raw observation. A four-year-old child absorbs more sensory information through watching and touching the world than the entire text corpus used to train a frontier model. Text can state that an unsupported object falls, but it does not carry the lived experience of watching a thousand things drop.
The core technical objection concerns the architecture itself. LeCun argues that autoregressive mechanisms—predicting one token at a time—are fundamentally incapable of modelling causal relationships or planning actions. The system is merely calculating the most statistically likely next character, not building a durable internal model of how the world is put together. This structural limitation, he contends, cannot be solved by adding more parameters or data.
The Missing Capabilities
Recent research has provided empirical support for this critique. A 2026 study in the International Journal of Computational Intelligence Systems identified a persistent "grounding gap" across contemporary models, "where surface-level linguistic fluency masks failures in mechanical plausibility, geometric transformation, and multi-entity relational consistency". The researchers found that models default to high-probability training templates even when presented with explicit counterfactual constraints—a computational analog of the "Einstellung effect," where familiar solutions block novel ones.
The "strawberry problem" has become emblematic of this limitation. When asked how many times the letter "r" appears in the word "strawberry," leading models have repeatedly answered "two"—even when they can correctly spell the word. The problem is not the mistake itself. The problem is that the system can make the mistake, produce the correct spelling, and still fail to recognise the contradiction. There is no intrinsic mechanism requiring these outputs to agree, and no internal alarm that a contradiction has occurred. As one analysis concluded, "a system that fails silently on simple tasks may pose more risk than one that struggles visibly on difficult ones".
The World Model Alternative
LeCun's alternative is the world model—a system that learns an internal representation of how the physical world behaves and uses it to predict consequences and plan actions. The technical core is JEPA (Joint Embedding Predictive Architecture), which makes its predictions not in the raw space of pixels or tokens, but in an abstract representation space.
The key distinction is that JEPA is explicitly not generative. Instead of asking what the next video frame will look like, a JEPA asks what the abstract representation of the next state will be—forcing it to capture high-level structure and discard unpredictable detail. LeCun's vision is a modular cognitive architecture with six components: a configurator that acts as executive control, a perception module, a world model module (where JEPA lives), a cost module, a short-term memory, and an actor that proposes and refines action sequences.
AMI Labs CEO Alexandre LeBrun draws a useful parallel: a factory robot repeating the same motion works well enough today, but "when you take your robot outside into a more open environment, in your household, or in the street, it must understand its surroundings and operate safely." He adds: "Robots are not safe right now—there's no solution for that today".
A Disagreement, Not a Verdict
The honest summary is that nobody yet knows who is right—and that uncertainty is the point. The other leaders of the field—Sam Altman, Dario Amodei, Demis Hassabis—largely think LeCun is wrong. This is a real and unresolved scientific dispute, not a settled verdict.
What is clear is that the disagreement has moved from academic journals to industrial investment. AMI Labs's $1.03 billion seed round is a wager on a scientific idea, not on demonstrated traction. The field has split, and the next few years will determine which path—or which synthesis—leads to general intelligence.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
