What is Yann LeCun's new AI startup about?

Yann LeCun is launching a new AI startup, reportedly named AMI Labs, with the goal of building "world models." This is a different type of AI that aims to learn and simulate how the world works, rather than just predicting the next word in a sentence like current Large Language Models (LLMs). The company is seeking to raise €500 million at a valuation of around €3 billion, with plans to be based in Paris.

Why is Yann LeCun critical of AI models like ChatGPT?

Yann LeCun believes that current generative AI models, such as ChatGPT, are fundamentally limited because their sole objective is to predict the next word or token. He argues this prevents them from developing a true understanding of the world, the ability to plan, or to reason logically. He considers this approach a "dead end" for achieving human-level intelligence and advocates for a new architecture called "objective-driven AI."

What is the difference between "world models" and LLMs?

LLMs (Large Language Models) are trained on vast amounts of text data to become proficient at generating human-like language. Their core function is autoregressive prediction—guessing the next word. World models, on the other hand, are designed to learn an internal, predictive model of an environment. They learn cause and effect, allowing them to simulate future events and plan actions accordingly. The goal is to build a system with a foundational understanding of reality, not just language.

What does this new startup mean for the European AI scene?

LeCun's decision to base his startup in Paris is a major boost for the European AI ecosystem. It follows the success of other European companies like Mistral AI and strengthens the continent's position as a hub for top-tier AI research and development. LeCun has explicitly stated his desire to work with European talent, signaling a potential shift of focus and resources away from the traditional dominance of Silicon Valley.

Is Yann LeCun leaving Meta?

Yes, Yann LeCun has announced he will be departing from his role as Chief AI Scientist at Meta to launch this new venture. While at Meta (and Facebook before it), he led some of the most important breakthroughs in deep learning and AI. His departure to start an independent company marks a significant transition in his career and in the broader AI industry.

What is "Objective-Driven AI"?

Objective-Driven AI is a term used by Yann LeCun to describe a new paradigm for building intelligent systems. Instead of training a model on a simple objective (like predicting the next word), this approach involves giving the AI more complex, intrinsic goals. To achieve these goals, the AI must learn to perceive its environment, build an internal model of it, and use that model to predict the consequences of its actions and plan accordingly. This is seen as a more robust path to creating versatile and intelligent machines.

Yann LeCun's New Venture: A Bold Challenge to the Future of AI

Yann LeCun, one of the "godfathers of AI" and a towering figure in the machine learning community, is making waves once again. The French computer scientist, renowned for his pioneering work in deep learning and his role as Chief AI Scientist at Meta, is launching a new AI startup. The ambitious venture is already in talks to raise an estimated €500 million, targeting a pre-launch valuation of around €3 billion. This move signals a significant new chapter in the AI race, one that questions the very foundation of today's generative AI models and proposes a radically different path toward artificial intelligence.

A New European AI Contender

LeCun's new company, reportedly named AMI Labs, is poised to become a major player in the global AI landscape. While details are still emerging, the startup is expected to be headquartered in Paris, a move that would significantly bolster Europe's position in the AI industry, which has seen the rise of other major players like Mistral AI. LeCun has been vocal about his intention to tap into European talent, suggesting that Silicon Valley has become "hypnotised" by the current generative AI trend.

The initial funding target of €500 million is a clear statement of intent, placing the startup in the same league as major AI labs from day one. This significant capital injection will enable the company to attract top-tier talent and secure the immense computational resources required to build foundational AI models. LeCun has confirmed he will not serve as the CEO but will be the scientific leader, focusing on the core research and development that will define the company.

The Mission: Building "World Models"

The central goal of LeCun's new venture is to develop "world models." This approach represents a fundamental departure from the Large Language Models (LLMs) like those developed by OpenAI, Google, and even his current employer, Meta. Instead of simply predicting the next word in a sequence, world models aim to build an internal, predictive understanding of how the world works.

What Are World Models?

A world model is a type of AI system that learns to simulate the future. It builds an internal representation of a given environment and uses it to predict how that environment will change in response to certain actions. For example, a world model could learn the basic principles of physics by watching videos and then predict the trajectory of a ball if it were thrown. This approach is believed to be a crucial step toward creating AI systems that can reason, plan, and act with a level of understanding closer to that of humans and animals.

This concept is inspired by how living beings learn. A baby, for instance, develops an intuitive understanding of physics long before they can speak. LeCun argues that this is the missing piece in today's AI systems, which, despite their impressive linguistic abilities, lack a genuine grasp of reality.

LeCun's Critique of Generative AI

For years, Yann LeCun has been one of the most prominent and respected critics of the prevailing direction of AI research, particularly the over-reliance on autoregressive LLMs. He argues that the current approach is fundamentally limited and will never lead to Artificial General Intelligence (AGI).

The Limitations of Large Language Models

According to LeCun, LLMs are flawed because they are trained on a simple, singular objective: predicting the next token (a word or part of a word). This, he claims, is an incredibly inefficient and constrained way to learn. Here’s a breakdown of his main criticisms:

Limitation	Description
Lack of Grounding	LLMs learn from text alone, meaning their "understanding" is not connected to any real-world experience or sensory input. They can describe gravity but don't "know" it in the way a child does after dropping a toy.
Inability to Plan	Because they only predict the next step, LLMs cannot perform complex, multi-step reasoning or planning. They generate responses token by token, without a coherent long-term goal for the answer.
"Hallucinations"	The tendency of LLMs to generate false or nonsensical information is a direct result of their architecture. They are designed to produce plausible-sounding text, not factually accurate statements grounded in reality.
Inefficient Learning	LeCun often points out that a human child learns more about the world in their first few years of life than an LLM does from trillions of words. This is because humans learn through interaction and observation, not just text prediction.

The "Dead End" of Generative AI

Yann LeCun has controversially stated that the current path of scaling up LLMs is a "dead end" on the road to human-level intelligence. He warns that while these models are useful tools, they are not a viable path to creating machines that can reason, understand the world, and interact with it intelligently. He believes the industry’s focus on them is slowing progress toward more robust AI architectures.

The Future: Objective-Driven AI

LeCun's alternative vision is what he calls "Objective-Driven AI." This framework proposes that AI systems should be designed with more complex, intrinsic objectives than just predicting the next word. These objectives would drive the AI to build a rich internal model of the world to achieve its goals.

Learning from Multiple Modalities : Integrating video, audio, and other sensory data to build a richer understanding of the world, much like humans do.
Predictive Models : Creating systems that can simulate future outcomes of actions, allowing for true reasoning and planning.
Hierarchical Planning : Breaking down complex tasks into smaller, manageable sub-goals, a key feature of intelligent behavior.

By pursuing this path, LeCun's new startup aims to create AI that is more reliable, capable, and aligned with human-like intelligence. It's a high-stakes bet against the current generative AI boom, but one that could redefine the future of the field if successful.

FAQ

On the same topic

Marc-Antoine Lebrun

Editor in chief

Passionate about finance and new technologies for many years, I love exploring and delving deeper into these fascinating fields to better understand them. Curious and always eager to learn, I’m particularly interested in cryptocurrencies, blockchain, and artificial intelligence. My goal: to understand and share the innovations that are shaping our future.

Yann LeCun's Bold AI Venture: World Models Challenge Generative AI