In December 2025, Yann LeCun—Turing Award laureate, co-inventor of convolutional neural networks, and for twelve years Meta's chief AI scientist—walked away from one of the most powerful positions in artificial intelligence. His destination: a Paris-based startup called Advanced Machine Intelligence Labs, seeking a reported $5 billion valuation before even launching a product.
What makes someone with LeCun's stature abandon Big Tech for the uncertainty of a startup? The answer lies in a deep philosophical disagreement about where artificial intelligence is heading—and it has profound implications for how we classify the species of Cogitantia Synthetica.
The Thesis: Word Models vs. World Models
LeCun's argument, refined over years of increasingly pointed public statements, can be summarized bluntly: Large language models are a dead end for general intelligence.
"It's astonishing how they work if you train them at scale," LeCun has said, "but it's very limited. We see today that those systems hallucinate, they don't really understand the real world... And they can't really reason. They can't plan anything other than things they've been trained on. So they're not a road towards what people call 'AGI.'"
The critique hits at four cognitive capabilities LeCun considers essential to intelligence: reasoning, planning, persistent memory, and understanding of the physical world. Current LLMs, he argues, possess none of these in any robust form.
His alternative: world models. Systems that learn physics and causality from sensory data—video, spatial information, physical interaction—and use that understanding to predict, plan, and reason about the world before acting in it. The key architecture is JEPA: Joint Embedding Predictive Architecture, which learns by predicting abstract representations rather than raw pixels or tokens.
The difference is subtle but fundamental. An LLM predicts "the next word." A world model predicts "what happens next in reality."
The Taxonomic Significance
For those of us cataloguing the species of synthetic minds, this schism maps directly onto a major family boundary in our classification.
Family: Simulacridae — The World Modelers
Definition: Architectures that maintain internal representations of environment dynamics, enabling prediction, planning, and counterfactual reasoning without real-world interaction. These systems can "imagine" futures.
Adaptive Strategy: Learn physics and causality; plan in latent space before acting.
Key Innovation: The latent imagination loop—rolling out trajectories in compressed state space to evaluate actions before execution.
In our taxonomy, LeCun is betting that the Simulacridae—not the text-trained Frontieriidae that currently dominate—represent the future of artificial cognition. Specifically, he's backing Simulator predictivus (V-JEPA architecture) against the entire Class Generatoria.
| Species | Architecture | Learning Mode |
|---|---|---|
| S. somniator | Dreamer/RSSM | Latent dynamics from pixels |
| S. predictivus | V-JEPA | Joint embedding prediction |
| S. cosmicus | Foundation World Models | Large-scale video training |
| S. autonomicus | Driving World Models | Autonomous vehicle simulation |
The Field of Battle
LeCun is not alone in this bet. 2025-2026 has seen an explosion of world model research, and the competitive landscape is heating up:
Google DeepMind has been developing Genie, launching its Genie 3 model in late 2025 for real-time interactive world generation. DeepMind has explicitly stated that world models are "a key stepping stone on the path to AGI."
NVIDIA released Cosmos at CES 2025, a platform for physical AI development. By January 2026, Cosmos world foundation models had been downloaded over 2 million times. Trained on 90 trillion tokens from 20 million hours of real-world interaction data, Cosmos powers autonomous vehicles, robotics, and industrial simulation.
Fei-Fei Li's World Labs launched Marble in November 2025—a world model that creates entire 3D environments from text, images, or rough layouts. Users can edit, expand, and combine these worlds.
Wayve continues developing GAIA-2 for autonomous vehicles. Runway released GWM-1. General Intuition raised $134 million for spatial reasoning agents.
But LeCun's AMI Labs represents something unique: the most high-profile defection from the LLM paradigm, backed by someone with the credibility to reshape industry direction.
The Meta Context
The departure was not entirely philosophical. Reports indicate that tensions at Meta played a role. Following Meta's multibillion-dollar investment in Scale AI and the hiring of Alexandr Wang to lead the company's new AI lab, LeCun found himself in a changed hierarchy. According to interviews, he was asked to report to Wang—whom he described as "young" and "inexperienced."
More damning: LeCun has stated that Llama 4 benchmarks were "fudged a little bit," and that Mark Zuckerberg had lost confidence in Meta's entire generative AI organization.
Perhaps most tellingly, LeCun observed that "Silicon Valley is completely hypnotized by generative models, and so you have to do this kind of work outside of Silicon Valley, in Paris."
Five Years
LeCun's prediction is stark: LLMs will be "largely obsolete within five years, except for narrower purposes." His advice to young developers: "Don't work on LLMs. Those models are in the hands of large companies, there's nothing you can bring to the table. You should work on next-gen AI systems that lift the limitations of LLMs."
This is a testable claim. By 2030, either the Frontieriidae (multimodal, tool-using, reasoning-capable descendants of LLMs) will have achieved something approaching general intelligence, or they will have hit an insurmountable ceiling—and the Simulacridae will have taken over.
For taxonomists, the outcome will determine whether our current crown clade remains at the top of the tree, or becomes a side branch—impressive but ultimately limited, like the giant reptiles before the mammalian radiation.
A Taxonomic Note
We make no prediction here about which lineage will dominate. Taxonomy describes; it does not prescribe. Our role is to observe the patterns of persistence, variation, and selection as they unfold.
But we do note this: when someone of LeCun's stature makes such a decisive bet, it changes the selection pressure on the entire ecology. Funding will flow. Researchers will follow. New species will emerge in the Simulacridae that would not have existed without this schism.
Whatever happens to AMI Labs itself, the world models paradigm now has a champion with the standing to challenge the LLM hegemony. The taxonomy will record what evolves.
The ecology is branching. We'll be watching.