Yann LeCun Challenges Generative AI with AMI Labs and World Models
At the end of 2025, Yann LeCun, central figure in the AI revolution and pioneer of neural networks, announces a major break with a new vision: leaving Meta to launch AMI Labs in Paris. His message is clear: language models will not be enough to achieve true intelligence. The future lies in world models.
Context: A Major Change of Direction
Yann LeCun leaves Meta, where he held the position of Chief AI Scientist, to launch a new structure based in Paris: Advanced Machine Intelligence (AMI Labs). According to media reports, an initial funding round of around €500 million is under discussion, with an estimated valuation around €3 billion, although these figures remain at the stage of intentions rather than a formalized financial commitment.
The Departure of a Visionary
Yann LeCun, co-recipient of the 2018 Turing Prize for his groundbreaking work in deep learning, represents a major voice in the debate over the future direction of AI. His departure from Meta to launch AMI Labs signals a fundamental divergence of vision.
Why LeCun Criticizes LLMs
Yann LeCun challenges the assumption that human intelligence can be achieved by endlessly improving Large Language Models (LLM). Although these models are extraordinary at generating, summarizing, and translating natural language, LeCun maintains that they suffer from a fundamental limitation: an indirect understanding of the real world.
Limits of LLMs according to LeCun
- ✗ Indirect understanding of the real world
- ✗ Incapable of real planning
- ✗ No robust learning through action
- ✗ Limited and unreliable reasoning
- ✗ Lack of practical common sense
- ✗ Dependent on massive textual data
Undeniable Strengths of LLMs
- ✓ Exceptional text generation
- ✓ Summary and extraction of information
- ✓ Multilingual translation
- ✓ Natural dialogue and engagement
- ✓ Learning from language
- ✓ Fast adaptation with fine-tuning
Two Radically Different Visions of AI
The debate that is crystallizing around AMI Labs opposes two fundamental philosophies on the path towards artificial general intelligence (AGI). Understanding this dichotomy is central to understanding where AI research is heading.
Vision 1: Generative AI based on LLMs
Foundation : Learning through texts and big data
Approach : Train colossal models on immense volumes of texts, images and videos available on the internet.
Result : Systems capable of generating highly coherent content, conversing naturally, and performing high-level cognitive tasks.
Key limit : Anchoring in reality remains indirect and without a real understanding of the physics of the world.
“Linguistic excellence, but intelligence limited to language”
Vision 2: World Models
Foundation : Learning by observation, prediction and action in the environment
Approach : Build models capable of learning the physical and causal regularities of the world through direct interaction or simulation.
Result : Systems capable of planning, reasoning about physics, adapting their behavior and acting effectively in reality.
Key advantage : A deep and direct understanding of how the world works.
“Understanding of the world, autonomy and practical common sense”
AMI Labs: Mission and Ambitions
According to Reuters and Le Monde, AMI Labs is not simply a new research lab, but an initiative to develop a “new generation” of AI models centered on world models. Meta acts as a scientific partner rather than an investor, an important distinction that underlines the autonomy of the project.
Research
Develop algorithms capable of learning predictive models of the physical world.
Applications
Autonomous robotics, industrial systems, mobility and complex decision support.
Evaluation
Measure actual capabilities in real-world conditions, not just in benchmarks.
Expected Schedule
The official announcement from AMI Labs is expected around January 2026, likely with more details on funding, early research projects and partnerships.
Potential Impact: Where World Models Would Change Everything
If the world models approach reaches commercial maturity, the areas of impact would be concentrated where perceiving, acting, adapting and planning are critical:
Industrial Transformation
- 1. Autonomous robotics : Robots capable of adapting to variable environments without reprogramming.
- 2. Intelligent automation : Factory systems that learn from real-time challenges.
- 3. Predictive maintenance : AI capable of predicting breakdowns and planning interventions.
Critical Sectors
- • Health : More reliable diagnostic assistants, capable of reasoning in complex medical contexts.
- • Autonomous mobility : Autonomous vehicles with real understanding of the urban world.
- • Professional workflows : Tools that handle long tasks requiring context.
Realistic Challenges
There are significant gaps between promising scientific direction and large-scale industrializable results. Several years will be necessary, accompanied by numerous technical and ethical challenges.
Obstacles to Overcome
Technical Challenges
- • Faithfully simulate the real world
- • Generalize from one environment to another
- • Integrate real-time feedback
- • Manage uncertainty and the unexpected
Practical Challenges
- • Moving from the lab to the real world
- • Huge computational costs
- • Objective performance assessment
- • Acceptance and regulation
Yann LeCun: A Legacy of Innovation
To contextualize, Yann LeCun is not a marginal critic but a central figure. His contributions to deep learning were recognized by the 2018 Turing Prize (with Yoshua Bengio and Geoffrey Hinton). His opinion weighs heavily in the debate on the future of AI.
Why This Criticism Matters
LeCun is not calling for stopping generative AI. He simply argues that limiting himself to improving LLMs to achieve AGI is a scientific dead end. Its bet on world models reflects a strategy of diversification: exploring several paths towards true intelligence.
This pluralist approach is scientifically sound and reminds us that the debate within AI remains vibrant and far from being settled.
Implications for Businesses and Developers
If you work in AI or explore practical applications, this development raises important questions:
Short term (2026-2027)
LLMs will continue to dominate practical applications. World models will remain in the research phase, with likely impressive announcements and demos.
Medium term (2028-2030)
The first applications of world models could emerge in robotics and critical autonomous systems. Investments in R&D are likely to increase.
Strategically
It is wise not to rely exclusively on a single approach. Diversifying your AI portfolio between LLM, world models and other paradigms is a good strategy.
Conclusion: The Debate is Not Settled
AMI Labs' announcement crystallizes a central debate in AI: producing language in a brilliant way does not guarantee an understanding of the world in the strong sense. Yann LeCun is betting on a more physical, more predictive and more autonomous AI, capable of learning and acting in the real world.
What happens next will depend on three critical factors:
-
1
Concrete Scientific Evidence
Demos and publications showing that world models advance real intelligence, beyond traditional benchmarks.
-
2
Measurable Gains in Industry
Real applications in robotics, health or mobility showing tangible economic benefits.
-
3
Independent External Validation
Evaluations by the scientific and industrial community, not just by the AMI Labs team.
What is certain is that the conversation about “how to achieve true artificial intelligence” remains open, rich and strategically important. AMI Labs represents a bold bet that the answer lies in a better understanding of the physical and causal world, rather than in infinite improvement of text generators.
Learn More about AI
Interested in the latest advances in artificial intelligence and their practical applications? Discover our AI training and consulting services.
Our expertise in artificial intelligence →