Yann LeCun Glossary

Deep-learning pioneer, FAIR founder, and AMI Labs executive chairman; for the Dictionary, LeCun is the leading advocate of world models as an alternative to treating larger language models as the whole road to machine intelligence.

Yann LeCun is a French-American computer scientist, a professor at New York University, the founding director of Meta’s Fundamental AI Research laboratory (FAIR), and the executive chairman of AMI Labs. He shared the 2018 A.M. Turing Award with Geoffrey Hinton and Yoshua Bengio for the conceptual and engineering breakthroughs that made deep neural networks a practical foundation of modern computing.

His first historical role is relatively settled. LeCun’s work on convolutional neural networks helped turn neural networks from an unfashionable research programme into machinery that could reliably recognise handwritten characters and, later, images. His second role is still being argued over: he is the most prominent internal dissenter from the idea that today’s large language models, made larger and given more compute, are the sufficient path to human-level machine intelligence.

Not anti-LLM

LeCun is often described as an LLM sceptic. That is accurate only if the noun matters more than the adjective. He does not claim that language models are useless. He disputes that fluent prediction over language is the same thing as understanding the world well enough to reason, plan, remember, and act within it.

Language is a compressed record of human experience. It contains an astonishing amount of knowledge, but not all the structure that humans and animals acquire by watching objects move, handling things, navigating space, pursuing goals, and discovering that actions have consequences. A cat learns a great deal about gravity without reading Newton. LeCun’s wager is that machines will also need a route to intelligence that begins in the world rather than only in descriptions of it.

Word models and world models

The centre of LeCun’s alternative is the world model: an internal representation that lets a system predict how a situation may change, including what is likely to happen if the system takes a particular action. A useful world model need not reproduce every pixel or every possible future. It needs to preserve the structure relevant to perception, prediction, planning, and control.

This is the point of JEPA, the Joint Embedding Predictive Architecture associated with LeCun’s research programme. Instead of generating an exact missing image region, video frame, or sensory future, a JEPA predicts in an abstract representation space. It tries to retain the important structure and ignore details that are inherently unpredictable. That is a technical proposal, not a finished theory of intelligence. But it is a genuine alternative to assuming that next-token prediction will simply grow into everything else.

LeCun set out the broader programme in his 2022 position paper, A Path Towards Autonomous Machine Intelligence. It combines learned world models, hierarchical planning, memory, intrinsic objectives, and self-supervised learning. Meta’s I-JEPA and V-JEPA projects were early implementations of pieces of that programme. In 2026, after leaving his chief-scientist role at Meta, he made the wager institutional by launching AMI Labs. Its compact thesis is unusually clear: “real intelligence does not start in language. It starts in the world.”

The counter-pole

For this Dictionary, LeCun is the clearest counter-pole to Dario Amodei’s Big Blob of Compute worldview.

That disagreement is easy to caricature. Amodei does not believe that racks of accelerators become intelligent through electrical enthusiasm alone; his blob also requires data, objectives, architecture, and numerical stability. LeCun is not opposed to scale; world models will also consume serious compute and data. The disagreement is about what is missing. The scaling camp expects broadly capable intelligence to continue emerging as general learning systems are enlarged and trained over richer tasks. LeCun expects a more fundamental architectural change: systems that learn abstractions from the physical world and use them to predict and plan.

This is an excellent argument to leave unsettled. If the scaling camp is right, LeCun may be building an elaborate detour around capabilities that language-centred systems will acquire anyway. If LeCun is right, the frontier laboratories may be spending fortunes polishing extraordinarily articulate machines that still lack a durable model of reality. The bill for discovering which view is closer to the truth will not be small.

Openness as infrastructure

LeCun has also been one of the most visible advocates for open AI research and downloadable model weights. FAIR’s research culture and Meta’s Llama releases helped make the modern Open Weights ecosystem possible, even though Llama’s licence is not fully open source in the traditional software sense.

His case for openness is scientific and political. Knowledge advances when researchers can inspect, reproduce, adapt, and challenge one another’s work. A world in which a handful of American corporations privately mediate the basic machinery of digital intelligence would also be a poor foundation for institutional or national autonomy. AMI Labs has carried this position forward by promising open publications, open-source work, and collaboration with the academic research community.

There is a useful tension here. LeCun spent twelve years at one of the largest platform companies in history while arguing that intelligence should not be controlled by a small number of companies. That does not invalidate the argument. It makes him a particularly good person through whom to examine it.

Why he belongs here

LeCun belongs in the Dictionary because he represents something rarer than prominence: a coherent rival theory.

He helped create the deep-learning settlement, then refused to treat its current dominant product as the end of the story. His world-model programme may fail, merge with language-model research, or help define the next architecture. Any of those outcomes would make the argument worth understanding now.

Sources

See also

Meta AI · Llama · Open Weights · Open-Weights Inversion · Scaling Laws · Big Blob of Compute · Dario Amodei · Sovereign Compute

Return to Dictionary All Entries (A–Z) For Students Other Writing Capstone 2.0