AI Thinkers·AI scientist, founder of AMI Labs
Yann LeCun
Autoregressive LLMs are a dead end on the path to human-level AI. They are useful writing tools but structurally incapable of understanding the world, planning, or reasoning. The alternative is the world model paradigm: AI systems that learn internal representations of how the world works — physics, causality, consequences — through observation and interaction, not text prediction. The architectural instantiation is JEPA (Joint Embedding Predictive Architecture), which operates in abstract representation space. He believes a fundamentally different AI paradigm will supersede autoregressive LLMs within three to five years, and left Meta to build it.
Key positions
Autoregressive LLMs cannot plan or reason — errors compound exponentially at every step.
When an LLM generates each token, there is a probability of drifting outside the space of correct answers. Since each word is generated conditionally on the last, errors accumulate multiplicatively — the probability of a nonsensical answer increases exponentially with output length. This is not a data or scale problem; it is a structural consequence of next-token prediction. This explains hallucination and the inability of LLMs to execute reliable multi-step plans.
Yann LeCun on LinkedIn, January 2023; Lex Fridman Podcast #416 (2023); HPCwire, Feb 2025
LLMs lack a world model — no internal representation of causality, physics, or 3D space.
Intelligent systems need a configurable world model capable of predicting future states of the world given actions. LLMs trained on text inherit language's surface structure but none of its grounding in physical reality. "An AI trained on words and sentences alone will never approximate human understanding."
"A Path Towards Autonomous Machine Intelligence," LeCun, 2022 (openreview.net); Meta AI blog on I-JEPA (2023) and V-JEPA 2 (June 2025)
JEPA — Joint Embedding Predictive Architecture — is the correct inductive bias for learning world representations.
JEPA learns by predicting the abstract representation of one part of an input from another — e.g. predicting the representation of a future video frame from a current one — without reconstructing raw pixels. By operating in embedding space, JEPA models the world at multiple levels of abstraction. V-JEPA 2 (June 2025) achieved state-of-the-art on physical reasoning, human action anticipation, and zero-shot robot control.
Meta AI blog, "Introducing V-JEPA 2" (June 11, 2025); arxiv.org/abs/2506.09985
LLMs will be largely obsolete as the central AI component within 3–5 years.
"Nobody in their right mind would use [LLMs of the current type] anymore" within this window. He characterised them as "a distraction" — an "off-ramp" that cannot deliver persistent memory, causal understanding, planning, and embodied common sense. He left Meta in November 2025 to build his conviction into practice at AMI Labs.
HPCwire (Feb 2025); MIT event (Oct 2025); AMI Labs launch, Fortune / TechCrunch (Dec 2025)
AI existential risk concerns are wildly overblown — possibly a regulatory moat for incumbents.
LeCun disagrees sharply with Turing Award co-winners Hinton and Bengio on AI safety. He argues AI has no inherent drive to acquire power, and the existential risk narrative is "wildly overblown and highly premature." He has suggested the safety narrative may be exploited by large companies to push for regulation that stifles competition.
TechCrunch, Oct 2024; X/Twitter post, May 2024
Open-source AI is the correct path — closed models dangerously concentrate power.
Open-source AI platforms are historically more powerful and more secure than closed-source equivalents. Closed AI — OpenAI, Anthropic, Google — concentrates technological power dangerously in few hands, while open models distribute capability and allow independent scrutiny. He left Meta in part because new leadership was moving toward a more closed approach.
BusinessToday, Oct 2024; LinkedIn posts; CNBC, Nov 2025
In their own words
“Auto-Regressive LLMs can't plan (and can't really reason).”
“But on the path towards human-level intelligence, an LLM is basically an off-ramp, a distraction, a dead end.”
“AI is not some sort of natural phenomenon that will just emerge and become dangerous. WE design it and WE build it.”
“We believe world models will usher a new era for robotics, enabling real-world AI agents to help with chores and physical tasks without needing astronomical amounts of robotic training data.”
Tensions with other thinkers
Where this view genuinely conflicts with others in the field.
Altman treats continued scaling as the near-certain path to superintelligence; LeCun calls LLMs a dead end.
Amodei argues open release removes the ability to enforce any controls; LeCun argues open distributes power and safety.
Disagrees sharply with Turing Award co-winners Hinton and Bengio on the urgency and nature of AI safety concerns.
Both reject LLMs as the path to AGI, but disagree on solution: world models (LeCun) vs. program synthesis (Chollet).