Turing Award Winner LeCun: AI Industry Fully Obsessed with LLMs, Leaves Meta to Found AMI Labs for Alternative World Models
Turing Award winner and former head of Meta AI research Yann LeCun pointed out that the entire AI industry has become completely "LLM-pilled," with all laboratories digging along the same path—relying on next-token prediction architectures, with additional reasoning, planning, and agents being temporary scaffolding. After leaving Meta in November 2025, he founded Advanced Machine Intelligence Labs (AMI Labs), which has completed $1.03 billion in seed funding, with a valuation of $3.5 billion, focusing on the JEPA world model to build AI systems that understand physical reality.
This viewpoint is based on the inherent limitations of LLMs in areas such as hallucination, planning, and robotics, as well as observations of diminishing returns in scaling. AMI Labs has the support of heavyweight investors like Bezos, Schmidt, and Cuban, aiming to break through the current architectural bottlenecks.
Source: Public Information
ABAB AI Insight
This statement and funding highlight the architectural-level divergence in AI development paths. LLMs achieve astonishing capabilities through massive text pattern completion but face fundamental limitations in causal reasoning, persistent world representation, and real-world interaction. This "same trench" phenomenon stems from short-term commercial incentives: existing architectures can quickly iterate benchmarks, attract funding, and users, while alternative paths like JEPA require longer validation, leading to a concentration of capital and talent in a few scalable routes.
Structurally, this reflects path dependence and power redistribution in technological innovation. Under LLM dominance, computing power, data, and distribution channels concentrate in leading laboratories, forming a strong feedback loop; LeCun's alternative bet attempts to shift AI from language statistics to physical world models, potentially reshaping productivity boundaries in robotics, industrial automation, and more. This fork could accelerate industry migration: if world models break through, pricing power will shift from language generation to embodied intelligence and long-term planning capabilities, further affecting wealth distribution in the AI supply chain.
The event is placed within the long-cycle evolution of AI. Historically, every paradigm (such as deep learning convolutional networks) has dominated for a time, ultimately being supplemented or replaced by new architectures due to inherent limitations. The current scaling competition both drives short-term capability leaps and exposes institutional inertia's constraints on diversified exploration. LeCun's independent action provides an external correction mechanism, with the ultimate outcome depending on whether world models can validate their theoretical advantages in real-world tasks, rather than merely remaining at the paper level.