Yann Le Cun : où va l’intelligence artificielle ?

Overview

This video is a lecture and discussion at Sciences Po (France) on Yann LeCun’s view of “where AI is going.” It covers Europe’s strategic position, the risks/security debate, AI governance, and—most importantly—how AI systems should be trained and deployed.

1) AI’s trajectory: LLMs work, but won’t deliver “human-like” intelligence alone

LeCun argues that the recent breakthrough (LLMs such as GPT-style models) is real and surprising, but insists there are important limits:

• Language prediction scales well, because predicting the next token is comparatively “easy.”
• The physical world is far more complex than text. Systems trained only on language don’t automatically gain physical competence—such as:
  • robotics
  • self-driving
  • household robots

He also emphasizes compute and data costs:

• Frontier LLMs require massive training, especially memory/infrastructure (e.g., large GPU clusters).
• This creates an uneven global landscape where big tech (US)—and potentially China—dominates, leaving other regions (including Europe) at a disadvantage.

2) A counter-proposal: “world models” and predictive learning beyond text (JEPA/JPA)

LeCun links intelligence to learning via prediction and adaptation, drawing inspiration from neuroscience ideas like predictive coding.

His proposal is that future progress depends on systems that learn abstract representations of reality—called world models—so they can predict the consequences of actions.

JEPA/JPA (Joint Embedding Predictive Architecture)

He describes his concept as:

• Training models to predict in a learned abstract space
• Removing details that can’t be predicted directly while preserving causal structure needed for planning

Claimed benefits

LeCun argues this approach could enable:

• More efficient models (less reliance on massive parameter counts/memory)
• Better generalization to physical tasks
• A path toward robotics and planning-capable AI

3) Europe’s “sovereignty” problem—and how to respond

The discussion frames Europe’s concern: if AI is dominated by the US and China, Europe risks a “digital Thucydides trap” (an arms-race dynamic) and loss of sovereignty.

LeCun’s response: sovereignty should be pursued through two main routes:

1. New technical directions
   • World-model-based systems rather than only scaling LLMs.
2. Open models/weights
   • Prevent information/control from being monopolized by a few Western and/or Chinese firms.

He also points to a broader initiative:

• “Project Tapestry”: an open, collaborative effort (across multiple countries/regions) to build and train an open repository of human knowledge, spanning many languages and cultural datasets—including some non-public cultural collections.

4) The “existential risk” debate: LeCun is skeptical of AI catastrophe narratives

In the discussion, LeCun (and others’ framing) pushes back against the claim that today’s AI is an existential threat to humanity.

His core stance:

• Regulation should focus on deployment/use, not speculative assumptions about near-term superintelligence or inevitable doom.
• He argues that “dangerous” scenarios often cited (e.g., AI enabling catastrophic cyber/bioweapon capabilities) are overhyped because:
  • real-world execution requires expertise, resources, and controlled environments
  • defense capabilities and automation for protection can keep pace

Instead of banning development, he favors realistic governance targeted at where harm can occur.

5) Governance: regulate applications, test deployments, and consider energy realities

A question (from an ambassador/UNESCO contributor) asks how governance can still work given rapid progress, especially around issues such as:

• protection of children
• cognitive rights
• copyright
• cultural diversity

LeCun replies:

• Many regulatory frameworks already apply by domain, such as:
  • medical authorization
  • testing for driver-assistance systems
  • rules governing private data
• Regulating AI research itself is unlikely to be workable—similar to trying to regulate a technology before it exists.
• He also notes that incentives already push for efficiency (e.g., energy/carbon), reducing the need for heavy regulation for that dimension.

6) Training data for physical AI: what’s missing

Asked what data is missing to move from text models to physical world-model AI, LeCun says the issue isn’t primarily raw data availability.

The harder problem is data shaped like:

• state → action → next state

This is especially difficult for robotic manipulation, where outcomes must be measured precisely.

He also highlights a robotics bottleneck:

• Humanoid robots exist, but making them practically useful is still extremely difficult.
• Collecting enough manipulation experience is economically challenging.

7) Social learning and cognition beyond the physical

A question challenges that LeCun’s focus on physical learning may understate the social dimension—learning with and through others.

LeCun responds that:

• Even if superintelligent systems emerge, they would likely remain under human control
• They would function as powerful teams/assistants enabling human projects
• Intelligence and learning can accelerate through social mechanisms (while noting that other intelligent species may develop without the same dynamics)

He also reiterates a “grounding problem” viewpoint:

Symbols/language must be grounded in reality to support genuine understanding.

8) Concluding tone

The moderator/closing remarks emphasize hope: catastrophic outcomes are not immediate, and dialogue across technical sciences and humanities/social sciences is important for education and for shaping AI’s future responsibly.

Presenters / contributors (as named in the subtitles)

• Yann LeCun
• Louis Vassie (speaker/moderator)
• Éric Azan
• Julie Clin (mentioned as dean of the law school; co-chair role for an AI committee)
• Ivan Lequin (entrepreneur associated with “Ami Ami” / Advanced Machine Intelligence)
• Alam Garby (French ambassador to UNESCO; UNESCO-related question)
• Dominique Boulier (speaker who asks a question)