LeCun calls auto-regressive LLMs "doomed" at NYU seminar

Meta Chief AI Scientist Yann LeCun delivered a sharp critique of large language models during a September 10, 2025 seminar at New York University's Center for Data Science, declaring that "Auto-Regressive LLMs are doomed" and presenting Joint Embedding Predictive Architecture (JEPA) as the superior path forward.

The presentation, documented in 67 slides, systematically dismantled current approaches to artificial intelligence. "They cannot be made factual, non-toxic, etc. They are not controllable," LeCun stated about large language models, pointing to fundamental mathematical limitations that he argued make current AI systems unreliable for critical applications.

According to the seminar materials, LeCun outlined a specific mathematical problem with token-by-token generation. The probability that any produced token takes systems "outside of the set of correct answers" creates exponential degradation. "Probability that answer of length n is correct (assuming independence of errors): P(correct) = (1-e)^n. This diverges exponentially. It's not fixable (without a major redesign)."

The NYU presentation positioned this critique within broader industry frustrations. Despite billions invested in autonomous driving technology, self-driving cars remain inferior to human drivers who learn basic skills in approximately 20 hours. "Where are my Level-5 self-driving car and my domestic robot?" LeCun asked, highlighting what he termed Moravec's paradox: "Things that are easy for humans are difficult for AI and vice versa."

Current machine learning approaches face insurmountable limitations, according to the presentation. Supervised learning requires massive labeled datasets. Reinforcement learning demands "insane amounts of trials." Even successful self-supervised learning only works effectively for discrete modalities like text, failing with high-dimensional continuous domains such as images and video.

LeCun's alternative framework centers on systems that learn world models from sensory inputs rather than statistical patterns from text. The presentation argued that a four-year-old child processes more visual data than large language models trained on text. "A four year-old child has seen more data than an LLM," one slide declared, calculating that children receive 1.1E14 bytes through optical processing compared to 0.9E14 bytes in text training.

The proposed Joint Embedding Predictive Architecture addresses these limitations through abstract representation learning. Unlike generative models that predict every pixel detail, JEPA systems learn compressed representations that capture essential world dynamics while discarding irrelevant information. "JEPA lifts the abstraction level, generative architectures do not," the presentation explained.

Technical demonstrations showcased concrete progress beyond theoretical criticism. V-JEPA models detect physical impossibilities in video sequences, showing dramatically increased prediction errors when objects violate basic physics principles. These systems learn intuitive physics from observation without explicit programming, similar to how infants develop understanding of gravity and object permanence.

DINO-WM implementations demonstrate practical robotics applications where world models enable planning through representation space rather than pixel manipulation. The system learns to predict how visual features change under different actions, enabling model-predictive control for manipulation tasks. Open-loop rollout experiments show superior performance compared to existing methods.

The seminar outlined hierarchical planning capabilities that mirror human cognitive processes. Using the example of traveling from NYU to Paris, LeCun demonstrated how different abstraction levels handle planning at appropriate scales. High-level decisions involve choosing transportation methods and routes, while low-level planning manages immediate obstacles and traffic conditions.

Energy-Based Models provide the mathematical foundation for these approaches, assigning low energy to compatible input-output pairs and higher energy to incompatible combinations. This framework avoids collapse problems that plague traditional architectures while enabling flexible objective specification through energy landscape shaping.

The presentation's timing coincides with growing industry skepticism about large language model limitations. Recent research has documented persistent hallucination problems, scaling challenges, and fundamental constraints in reasoning capabilities. LeCun's critique builds on these concerns while offering specific technical alternatives.

Meta's research investments reflect commitment to these alternative approaches. The company has developed I-JEPA for image analysis, V-JEPA for video understanding, and multiple specialized implementations. V-JEPA 2 demonstrates large-scale training capabilities using two-phase approaches with masked video prediction followed by action-conditioning.

Training methodologies emphasize efficiency advantages over current approaches. VICReg regularization techniques prevent model collapse while maximizing information content in learned representations. The presentation showed that JEPA models require significantly fewer training iterations than comparable systems, with I-JEPA needing approximately 5x fewer iterations than alternative methods.

Industry applications extend beyond academic research into practical deployment scenarios. Navigation World Models research, published in December 2025, demonstrates MPC planning from natural motion-conditioned videos. These systems generate coherent predictions about future states based on action sequences, enabling autonomous navigation through complex environments.

The seminar's broader message challenged current AI development priorities. "IF YOU ARE INTERESTED IN HUMAN-LEVEL AI, DON'T WORK ON LLMs," LeCun concluded, advocating for fundamental redirections in research focus toward world model development, energy-based learning, and objective-driven systems.

Future research directions outlined specific technical challenges requiring resolution. Large-scale world-model training from multiple modalities, gradient-based planning algorithms, and hierarchical architecture development represent priority areas. Mathematical foundations of energy-based learning need continued theoretical work alongside practical implementations.

The September 10 presentation drew significant academic and industry attention, reflecting broader debates about AI development trajectories. LeCun's position as Meta's Chief AI Scientist lends institutional weight to these critiques, potentially influencing research funding and strategic priorities across the artificial intelligence community.

According to social media documentation, the seminar attracted capacity attendance from NYU students, postdocs, and faculty. The event represents ongoing academic-industry collaboration in fundamental AI research, particularly focused on alternatives to dominant generative modeling paradigms.

Why this matters for marketing

LeCun's systematic critique of large language models carries significant implications for marketing technology development. Current AI-powered advertising tools rely heavily on generative approaches that LeCun characterized as fundamentally flawed, suggesting potential disruptions to existing automation infrastructure.

The mathematical limitations LeCun identified affect campaign optimization systems that use language models for creative generation and audience targeting. Exponential error accumulation in token-by-token generation could explain persistent issues with AI-generated advertising copy that appears coherent initially but contains subtle inaccuracies or inappropriate messaging.

World model approaches could address current limitations in programmatic advertising systems that struggle to understand complex customer journeys and cross-channel attribution. JEPA architectures that learn abstract representations of consumer behavior could enable more sophisticated planning and optimization than statistical pattern matching.

Video advertising applications appear particularly relevant given demonstrated V-JEPA capabilities in understanding temporal dynamics and physical interactions. As advertising industry adoption of AI-generated video approaches 90%, world models could generate more contextually appropriate content based on understanding of visual narratives rather than text-to-video generation.

The shift toward objective-driven systems aligns with marketing needs for controllable AI that respects brand guidelines and compliance requirements. Unlike current black-box approaches, JEPA frameworks enable explicit objective specification and constraint handling, potentially addressing concerns about AI-generated content that violates brand standards.

However, the timeline for practical marketing applications remains uncertain. LeCun's presentation focused on foundational research rather than immediate commercial deployment. The technology transition could require several years of development before reaching production readiness for marketing automation platforms.

Current marketing technology investments in language model-based tools may face obsolescence if JEPA approaches prove superior. Companies like Meta have invested heavily in AI infrastructure, but the shift away from generative models could require fundamental architectural changes in advertising technology stacks.

Timeline

• September 10, 2025: Yann LeCun presents critique of LLMs at NYU Center for Data Science, calling auto-regressive models "doomed" and presenting JEPA alternatives
• December 2025: Navigation World Models research published, demonstrating practical applications of JEPA approaches in autonomous systems (arxiv:2412.03572)
• August 2025: Meta CEO outlines personal superintelligence vision emphasizing individual empowerment over centralized AI control
• July 2025: McKinsey analysis documents agentic AI trends showing $1.1 billion investment and 985% increase in related job postings
• 2024: Meta releases V-JEPA model under Creative Commons license, demonstrating video understanding capabilities superior to generative approaches
• 2023: I-JEPA paper published at CVPR, introducing first practical implementation of Joint Embedding Predictive Architecture
• 2022: LeCun publishes "A Path Towards Autonomous Machine Intelligence," proposing JEPA as fundamental alternative to generative models
• Related: IAB Europe releases AI policy whitepaper addressing European digital advertising transformation (July 2025)

Summary

Who: Yann LeCun, Meta Chief AI Scientist and NYU professor, delivering critique to NYU Center for Data Science audience including students, postdocs, and faculty

What: Systematic critique of large language models as "doomed" due to mathematical limitations, presenting Joint Embedding Predictive Architecture as superior alternative for achieving human-level artificial intelligence

When: September 10, 2025, during NYU Center for Data Science seminar series, documented in comprehensive 67-slide presentation

Where: New York University Center for Data Science, with research spanning Meta's AI labs and academic collaborations

Why: To challenge current AI development priorities focused on generative models, arguing that world model approaches offer more promising paths toward controllable, reliable artificial intelligence systems