YANN LE CUN RÉVÈLE LES PLUS GROS MENSONGES SUR L’IA

Overview

Yann Le Cun (invited guest) argues that today’s AI—especially large language models (LLMs)—is revolutionary, but not a “royal road” to human-level intelligence or common-sense understanding of the physical world. He contrasts language-based systems with animal/human intelligence and emphasizes the key missing capability: fast learning of new skills and reliable understanding of consequences in the real world.

1) Leaving Meta / critique of the LLM-only direction

• Le Cun says he left Meta officially effective Dec 31, 2025, after Meta redirected priorities toward LLMs and short-term objectives.
• He frames this as a “change of era,” despite the fact that he helped build/lead a research lab that later shifted direction.
• He argues the industry repeatedly tells investors that scaling LLMs (more data, more compute, bigger models) will reach human-level intelligence—an explanation he calls false.

2) What intelligence is (and why LLMs fall short)

• Le Cun defines intelligence not as “a collection of skills” or “stored declarative knowledge,” but as the ability to acquire new skills quickly.
• He claims animals (e.g., cats) show intelligence that today’s AI systems cannot replicate; similarly, humans can learn tasks rapidly—even in novel situations.
• Core criticism:
  • LLMs are strong at language patterns and can be useful tools.
  • However, they do not understand the physical world and lack “common sense” reasoning in real environments.

3) LLM limitations: hallucinations and economic/engineering realities

• Hallucinations: he notes that the more you “back-and-forth” with an LLM, the higher the drift toward incorrect answers.
• Engineering constraints: he discusses the industry shift from “research” toward engineering optimization, focusing on inference efficiency and cost.
• Market economics: many AI services are expensive at inference time (e.g., usage/tokens). Some features (such as certain video generation capabilities) are removed because they aren’t profitable enough.

4) Specialization, tool-use, and “LLM systems” vs pure LLMs

• Le Cun argues that what people often call “LLMs” in products are frequently LLM systems combined with:
  • calculators for arithmetic,
  • databases for factual lookups,
  • external/regression tools for analysis,
  • orchestration among specialized components.
• These systems can be more useful than raw LLMs, but they still don’t solve the core issue: reliable real-world understanding and planning.

5) “World models” as the missing ingredient

• He proposes world models as the path toward real-world intelligence:
  • models that predict how the world changes in response to actions,
  • enabling planning (“what happens if I do X?”), closer to how humans decide.
• He describes training approaches that avoid pixel-level future prediction (too hard/impossible) and instead learn abstract predictive representations.
• He references a concept related to Joint Embedding Predictive Architecture (JEPA/JPA):
  • not generative in the sense of reproducing raw inputs (like pixel-perfect video),
  • but learning an internal abstraction where prediction becomes feasible.

6) Why video/video prediction won’t work like language modeling

He argues full-detail (pixel-level) video prediction fails because:

• too many relevant details are not observable,
• the future is combinatorially complex,
• representing a full probability distribution over all possible futures is infeasible.

Instead, world-model approaches should predict at an abstract level (e.g., “a bottle will fall and break”) rather than forecasting every pixel trajectory.

7) Robotics: LLMs won’t fix humanoid robots by scaling

• He claims humanoid robots fail because they don’t reliably predict consequences during physical interaction.
• He compares LLMs to hammers: applying them broadly to robotics leads to failure because the core challenge is dynamics and physical causality.
• Progress, he suggests, depends more on:
  • perception modules trained separately (e.g., convolutional nets / vision models),
  • simulation-based training,
  • real-world predictive control rather than language-only intelligence.

8) Tesla/Waymo as examples (and why they’re not “intelligent” in the AGI sense)

• He discusses autonomous driving capability levels (Level 2/3/4/5).
• He argues success in favorable conditions comes from:
  • engineering,
  • sensors,
  • accurate maps,
  • long development cycles, rather than human-level intelligence emerging from LLM scaling.

9) Investment and roadmap for “AI for the real world” (Amy Labs)

• He says their startup (Amy Labs, discussed as “AI for the real world”) raised about ~$1B+ (with approximate euro figures also mentioned).
• Funding supports:
  • solidifying the research methodology,
  • then infrastructure/software engineering to run training.
• Expected timeline:
  • within 1 year: solidify training methodology for world models and enable planning/agentic systems,
  • within 3–5 years: aim for more universal intelligent systems that learn from relatively limited data and handle real-world problems.
• He claims data scarcity is less of an issue if models learn from large-scale video/sensor streams (e.g., training JEPA video models on “about a century of video” is referenced).

10) Societal/labor impact view

• He argues AI-driven automation won’t necessarily cause mass unemployment.
• He cites economists and analogies to past technological revolutions (e.g., smartphone era) that eventually created new jobs.

---

Presenters / contributors

• Yann Le Cun (guest; interviewed)
• Host / presenter: (unnamed in subtitles; interviewer during “Génération… / podcast” segment)
• Alexandre Lebrin (discussed as a contributor/entrepreneur)
• Laurent Soli (mentioned as a former Meta colleague in Europe operations)
• Delphine Groll (mentioned as co-founder at Nabla)
• Martin Raison (mentioned as CTO at Nabla)
• Mark Zuckerberg (mentioned; support in past research)
• Alex Wang / “Alex” (mentioned; Meta leadership shift)
• Jean-Louis Constanza (mentioned; robotics/industrial viewpoint discussed)
• Mathieu Stéphanie (mentioned as an illustrative example name in a video prediction anecdote)
• Adrien Caner (mentioned as founder of a defense/drone-related company)
• Jensen (mentioned; e.g., as a commentator on AI/GPU/strategy)
• Dario Modi (mentioned)
• Elon Musk (mentioned; X/Tesla discussion)
• Jules (spoken as “Julien Guéron”) (mentioned as go-kart instructor)
• Philippe Mangin (mentioned as surgeon/doctor in personal example)