LeCun calls auto-regressive LLMs “doomed” at NYU seminar

Meta Chief AI Scientist Yann LeCun delivered a pointy critique of enormous language fashions throughout a September 10, 2025 seminar at New York College’s Heart for Information Science, declaring that “Auto-Regressive LLMs are doomed” and presenting Joint Embedding Predictive Structure (JEPA) because the superior path ahead.

The presentation, documented in 67 slides, systematically dismantled present approaches to synthetic intelligence. “They can’t be made factual, non-toxic, and many others. They aren’t controllable,” LeCun acknowledged about massive language fashions, pointing to elementary mathematical limitations that he argued make present AI techniques unreliable for vital purposes.

In response to the seminar materials, LeCun outlined a particular mathematical drawback with token-by-token technology. The likelihood that any produced token takes techniques “outdoors of the set of right solutions” creates exponential degradation. “Chance that reply of size n is right (assuming independence of errors): P(right) = (1-e)^n. This diverges exponentially. It isn’t fixable (with no main redesign).”

The NYU presentation positioned this critique inside broader {industry} frustrations. Regardless of billions invested in autonomous driving expertise, self-driving automobiles stay inferior to human drivers who study primary abilities in roughly 20 hours. “The place are my Stage-5 self-driving automobile and my home robotic?” LeCun requested, highlighting what he termed Moravec’s paradox: “Issues which can be straightforward for people are tough for AI and vice versa.”

Present machine studying approaches face insurmountable limitations, in accordance with the presentation. Supervised studying requires huge labeled datasets. Reinforcement studying calls for “insane quantities of trials.” Even profitable self-supervised studying solely works successfully for discrete modalities like textual content, failing with high-dimensional steady domains similar to photos and video.

LeCun’s different framework facilities on techniques that study world fashions from sensory inputs moderately than statistical patterns from textual content. The presentation argued {that a} four-year-old youngster processes extra visible information than massive language fashions educated on textual content. “A 4 year-old youngster has seen extra information than an LLM,” one slide declared, calculating that kids obtain 1.1E14 bytes by way of optical processing in comparison with 0.9E14 bytes in textual content coaching.

The proposed Joint Embedding Predictive Structure addresses these limitations by way of summary illustration studying. In contrast to generative fashions that predict each pixel element, JEPA techniques study compressed representations that seize important world dynamics whereas discarding irrelevant info. “JEPA lifts the abstraction degree, generative architectures don’t,” the presentation defined.

Technical demonstrations showcased concrete progress past theoretical criticism. V-JEPA fashions detect bodily impossibilities in video sequences, exhibiting dramatically elevated prediction errors when objects violate primary physics rules. These techniques study intuitive physics from statement with out specific programming, much like how infants develop understanding of gravity and object permanence.

DINO-WM implementations reveal sensible robotics purposes the place world fashions allow planning by way of illustration area moderately than pixel manipulation. The system learns to foretell how visible options change beneath totally different actions, enabling model-predictive management for manipulation duties. Open-loop rollout experiments present superior efficiency in comparison with present strategies.

The seminar outlined hierarchical planning capabilities that mirror human cognitive processes. Utilizing the instance of touring from NYU to Paris, LeCun demonstrated how totally different abstraction ranges deal with planning at applicable scales. Excessive-level choices contain selecting transportation strategies and routes, whereas low-level planning manages instant obstacles and visitors situations.

Vitality-Based mostly Fashions present the mathematical basis for these approaches, assigning low power to suitable input-output pairs and better power to incompatible combos. This framework avoids collapse issues that plague conventional architectures whereas enabling versatile goal specification by way of power panorama shaping.

The presentation’s timing coincides with rising {industry} skepticism about massive language mannequin limitations. Current analysis has documented persistent hallucination issues, scaling challenges, and elementary constraints in reasoning capabilities. LeCun’s critique builds on these issues whereas providing particular technical options.

Meta’s analysis investments mirror dedication to those different approaches. The corporate has developed I-JEPA for picture evaluation, V-JEPA for video understanding, and a number of specialised implementations. V-JEPA 2 demonstrates large-scale coaching capabilities utilizing two-phase approaches with masked video prediction adopted by action-conditioning.

Coaching methodologies emphasize effectivity benefits over present approaches. VICReg regularization methods forestall mannequin collapse whereas maximizing info content material in realized representations. The presentation confirmed that JEPA fashions require considerably fewer coaching iterations than comparable techniques, with I-JEPA needing roughly 5x fewer iterations than different strategies.

Business purposes prolong past tutorial analysis into sensible deployment situations. Navigation World Fashions analysis, printed in December 2025, demonstrates MPC planning from pure motion-conditioned movies. These techniques generate coherent predictions about future states based mostly on motion sequences, enabling autonomous navigation by way of complicated environments.

The seminar’s broader message challenged present AI growth priorities. “IF YOU ARE INTERESTED IN HUMAN-LEVEL AI, DON’T WORK ON LLMs,” LeCun concluded, advocating for elementary redirections in analysis focus towards world mannequin growth, energy-based studying, and objective-driven techniques.

Future analysis instructions outlined particular technical challenges requiring decision. Massive-scale world-model coaching from a number of modalities, gradient-based planning algorithms, and hierarchical structure growth signify precedence areas. Mathematical foundations of energy-based studying want continued theoretical work alongside sensible implementations.

The September 10 presentation drew vital tutorial and {industry} consideration, reflecting broader debates about AI growth trajectories. LeCun’s place as Meta’s Chief AI Scientist lends institutional weight to those critiques, probably influencing analysis funding and strategic priorities throughout the bogus intelligence group.

In response to social media documentation, the seminar attracted capability attendance from NYU college students, postdocs, and school. The occasion represents ongoing academic-industry collaboration in elementary AI analysis, notably centered on options to dominant generative modeling paradigms.

Why this issues for advertising

LeCun’s systematic critique of enormous language fashions carries vital implications for advertising expertise growth. Present AI-powered promoting instruments rely closely on generative approaches that LeCun characterised as essentially flawed, suggesting potential disruptions to present automation infrastructure.

The mathematical limitations LeCun recognized have an effect on marketing campaign optimization techniques that use language fashions for artistic technology and viewers focusing on. Exponential error accumulation in token-by-token technology might clarify persistent points with AI-generated promoting copy that seems coherent initially however incorporates delicate inaccuracies or inappropriate messaging.

World mannequin approaches might deal with present limitations in programmatic advertising systems that wrestle to grasp complicated buyer journeys and cross-channel attribution. JEPA architectures that study summary representations of shopper habits might allow extra refined planning and optimization than statistical sample matching.

Video promoting purposes seem notably related given demonstrated V-JEPA capabilities in understanding temporal dynamics and bodily interactions. As advertising industry adoption of AI-generated video approaches 90%, world fashions might generate extra contextually applicable content material based mostly on understanding of visible narratives moderately than text-to-video technology.

The shift towards objective-driven techniques aligns with advertising wants for controllable AI that respects model pointers and compliance necessities. In contrast to present black-box approaches, JEPA frameworks allow specific goal specification and constraint dealing with, probably addressing issues about AI-generated content material that violates model requirements.

Nevertheless, the timeline for sensible advertising purposes stays unsure. LeCun’s presentation centered on foundational analysis moderately than instant business deployment. The expertise transition might require a number of years of growth earlier than reaching manufacturing readiness for advertising automation platforms.

Present advertising expertise investments in language model-based instruments could face obsolescence if JEPA approaches show superior. Corporations like Meta have invested heavily in AI infrastructure, however the shift away from generative fashions might require elementary architectural adjustments in promoting expertise stacks.

Timeline

• September 10, 2025: Yann LeCun presents critique of LLMs at NYU Heart for Information Science, calling auto-regressive fashions “doomed” and presenting JEPA options
• December 2025: Navigation World Fashions analysis printed, demonstrating sensible purposes of JEPA approaches in autonomous techniques (arxiv:2412.03572)
• August 2025: Meta CEO outlines personal superintelligence vision emphasizing particular person empowerment over centralized AI management
• July 2025: McKinsey analysis documents agentic AI trends exhibiting $1.1 billion funding and 985% improve in associated job postings
• 2024: Meta releases V-JEPA mannequin beneath Artistic Commons license, demonstrating video understanding capabilities superior to generative approaches
• 2023: I-JEPA paper printed at CVPR, introducing first sensible implementation of Joint Embedding Predictive Structure
• 2022: LeCun publishes “A Path In direction of Autonomous Machine Intelligence,” proposing JEPA as elementary different to generative fashions
• Associated: IAB Europe releases AI policy whitepaper addressing European digital promoting transformation (July 2025)

Abstract

Who: Yann LeCun, Meta Chief AI Scientist and NYU professor, delivering critique to NYU Heart for Information Science viewers together with college students, postdocs, and school

What: Systematic critique of enormous language fashions as “doomed” as a result of mathematical limitations, presenting Joint Embedding Predictive Structure as superior different for attaining human-level synthetic intelligence

When: September 10, 2025, throughout NYU Heart for Information Science seminar sequence, documented in complete 67-slide presentation

The place: New York College Heart for Information Science, with analysis spanning Meta’s AI labs and tutorial collaborations

Why: To problem present AI growth priorities centered on generative fashions, arguing that world mannequin approaches supply extra promising paths towards controllable, dependable synthetic intelligence techniques