NextFin News - Recorded during the AI Action Summit "Science Days" organized by Institut Polytechnique de Paris, the following interview with Yann LeCun was conducted at the IP Paris conference campus in Saclay on or around 06 February 2025. The AI Experts interview series was produced by IP Paris and carried out by a team of PhD students and researchers from its member schools.
The interview was conducted as part of IP Paris's "AI Experts" series; the conversations were organized and coordinated by the university and the AI Action Summit program. The interviews in this series were produced by Xiaoxuan Hei, Paul Krzakala, Olivier Laurent, Louise Davy, Marie Reinbigler and Rajaa El Hamdani, under the coordination of Christine Nayagam.
How LeCun defines intelligence
LeCun began by reframing what it means to be intelligent. He emphasized that intelligence is not a single number or a simple scale but a multi‑faceted set of abilities grounded in knowledge and rapid learning. As he put it, intelligence is not a linear thing
, and people may be "very smart about certain things and very stupid about others." He said intelligence consists of "a collection of skills" supported by knowledge and an ability to learn new skills quickly or sometimes to solve problems without prior training.
Difference between human intelligence and today's AI
LeCun contrasted human flexibility with current machine learning systems. Humans can face a new situation, plan a sequence of actions and achieve a goal without prior training; today's AI systems, by contrast, must be trained separately for every task. He summarized the practical consequence plainly: trained tasks are handled well, and untrained tasks are often handled terribly. This gap, he argued, explains why chatbots can appear smart in domains with abundant text data but fail in others.
The limits of large language models
On large language models, LeCun was unequivocal: while LLMs are useful and have many applications, they are not the path to human‑level intelligence. We're never going to get to human level intelligence by scaling up large language models,
he said. He explained this with a comparison of training data bandwidth: the enormous token counts used to train LLMs may match the raw bytes a young child has received, but that child’s data is rich, multimodal and grounded in continuous sensory experience. From that LeCun concluded that mere text scaling is insufficient.
Why multimodal, world‑based learning matters
LeCun argued that real progress requires systems that learn from high‑bandwidth sensory streams such as video. He noted that a four‑year‑old’s visual experience—measured in raw bytes—compares with the scale of LLM training corpora, but the nature of the information is fundamentally different. That observation led him to call for "world‑based models" that learn from the dynamics of the physical world rather than only from static text: systems that can form representations, predict outcomes of actions, and abstract away irrelevant details.
Four essential traits AI currently lacks
LeCun listed four traits he sees as essential to intelligent behavior and largely absent from present systems: understanding the physical world, persistent memory, reasoning, and hierarchical planning. He described each briefly and stressed that current approaches often tack these capabilities onto LLMs as add‑ons—vision modules, retrieval systems, or enlarged parameter counts—which he characterized as temporary "hacks." Instead, he called for new techniques that integrate these traits at a fundamental level.
On learning abstractions and V‑JEPA
LeCun illustrated the world‑based approach with Meta’s V‑JEPA work: rather than predict pixels, the model learns abstract representations of masked parts of video and makes predictions in that representation space. He likened the process to scientific abstraction—building hierarchical representations that discard irrelevant details at each layer so higher‑level reasoning becomes tractable.
Research culture, peer review and fashions in the field
The conversation turned to the research ecosystem. LeCun reflected on peer review and the pressure of fashions in machine learning: because the field is expanding rapidly, many researchers are young and inexperienced, which can lead to antagonistic reviewing practices and a tendency to follow crowded topics. He observed that highly innovative work can be rejected if it falls outside prevailing interests, and he lamented that the reviewing process can be unreliable and overly negative.
Advice to students and early‑career researchers
LeCun offered practical guidance for students deciding research directions. He advised against working on problems dominated by industry scale and resources—explicitly mentioning large language models—because companies already command vast engineering teams and compute. Instead, he urged students to pursue less crowded domains where they have a better chance to make original contributions, such as building systems that understand the physical world, develop persistent memory mechanisms, plan hierarchically, or improve reasoning.
On production systems, ranking algorithms and content moderation
LeCun also described how production systems differ from research prototypes. Ranking algorithms used by platforms are lightweight, executed billions of times per day with strict latency and power constraints, and are deployed on specialized chips. He explained the trade‑offs these systems embody—between responsiveness, content moderation, legal obligations and the desire to allow broad public discussion—and noted that moderation policies and ranking choices have evolved over time to strike different balances.
Closing points
Throughout the interview LeCun emphasized pragmatic research goals: build models that learn from the real world, create persistent memory structures, and develop reasoning and planning capabilities intrinsic to the models rather than piecemeal add‑ons. His remarks closed with a call for new techniques and a reminder that linguistic fluency alone does not equal understanding of the physical world.
References
IP Paris — AI Experts interview series page: Global AI Insights: IP Paris Presents “AI Experts Interviews Series”.
AI Action Summit (IP Paris): AI Action Summit Conference: AI, Science, and Society by IP Paris (06–07 Feb 2025).
Reporting on LeCun’s remarks at the AI Action Summit: Business Insider — "Meta chief AI scientist Yann LeCun says current AI models lack 4 key human traits".
Photo and event reporting noting LeCun at Saclay on 06 Feb 2025: Channels Television — "Meta Chief AI Scientist Yann LeCun Leaving For Startup".
Explore more exclusive insights at nextfin.ai.
