Yann LeCun Just Left Meta to Bet $3.5 Billion That LLMs Are a Dead End
One of the three "godfathers of AI" just walked away from the world's largest social media company to bet $3.5 billion that everything you know about artificial intelligence is about to become obsolete. Yann LeCun — Turing Award winner, inventor of convolutional neural networks, and Meta's chief AI scientist for over a decade — has left to build something he believes will make ChatGPT, Claude, and Gemini look like calculators compared to computers.
His new startup is called AMI Labs — Advanced Machine Intelligence. It has a $3.5 billion valuation, a $1.03 billion seed round (Europe's largest ever), and a mission statement that directly challenges the foundation of every major AI company in the world.
This is not a side project. This is the most consequential departure in the history of artificial intelligence.
The Biggest Exit in AI History
Let us start with the numbers, because they are extraordinary.
$3.5 billion pre-money valuation — before the company has shipped a single product. This makes AMI Labs one of the most valuable AI startups at inception, rivaling the early valuations of companies that took years to build.
$1.03 billion seed round — the largest seed round in European history. The round was led by a consortium of investors who are betting that LeCun's vision of AI's future is correct.
Headquartered in Paris — "You pronounce it 'ami' — it means 'friend' in French," LeCun explained. The company will also have offices in Montreal, New York, and Singapore.
The CEO is Alex LeBrun, while LeCun took the Executive Chair role. His explanation was characteristically blunt: "I am too disorganized for this, and also too old" for day-to-day management. His job, he said, is to inspire.
Why LeCun Left: The Zuckerberg-Wang Conflict
The departure was not planned. It was provoked.
In a Financial Times interview, LeCun laid out the chain of events that pushed him out of Meta. The core issue was Mark Zuckerberg's decision to hire Alexandr Wang — the 27-year-old founder of Scale AI — and give him authority over Meta's new Superintelligence division.
Zuckerberg paid $14.3 billion for a 49 percent stake in Scale AI and appointed Wang to lead the division. LeCun, who had been Meta's top AI researcher for over a decade, was suddenly reporting to someone nearly four decades younger with a fundamentally different vision for AI.
LeCun's response was direct: "You don't tell a researcher what to do. You certainly don't tell a researcher like me what to do."
But the conflict went deeper than organizational hierarchy. LeCun and Zuckerberg had a fundamental disagreement about the direction of AI research.
Zuckerberg was pouring billions into large language models — the same technology behind ChatGPT, Claude, and Gemini. He wanted Meta to compete directly with OpenAI and Anthropic in the LLM race.
LeCun believed this was a mistake. He had been publicly arguing for years that LLMs are "useful but fundamentally limited" in their ability to reason and plan like humans. He wanted Meta to invest in a radically different approach.
When Wang arrived and doubled down on Zuckerberg's LLM-first strategy, LeCun decided it was time to build what he believed in — on his own terms.
What Are "World Models"? The Idea That Could Change Everything
LeCun's core argument is simple but radical: real intelligence does not start in language.
Humans do not learn about the world by reading text. Babies learn by observing, touching, moving, and experimenting. They build mental models of how the physical world works — gravity, object permanence, cause and effect — long before they learn their first word.
Current AI models — ChatGPT, Claude, Gemini — learn almost exclusively from text. They process written language and generate written language. They can appear intelligent because they are very good at predicting what word comes next. But they do not actually understand the physical world.
A world model is an AI that learns the way humans do — from video, spatial data, and physical interaction. It builds an internal simulation of how the world works and uses that simulation to plan, reason, and predict.
The technical architecture is called V-JEPA (Video Joint Embedding Predictive Architecture), which LeCun developed at Meta. Instead of predicting the next word in a sequence, V-JEPA predicts what will happen next in a video — how objects will move, interact, and change over time.
If it works, a world model could:
- Plan multi-step tasks by simulating outcomes before acting
- Understand physics without being explicitly taught
- Reason about spatial relationships that LLMs struggle with
- Learn from much less data because visual/physical learning is more information-dense than text
- Generalize to new situations more naturally than pattern-matching on text
Why LeCun Thinks LLMs Are a Dead End
LeCun has been making this argument for years, but the AI industry has largely ignored him because LLMs keep getting better. His counter-argument has three pillars:
Pillar 1: LLMs cannot plan. Ask ChatGPT to plan a complex multi-step task — like moving house, or designing a building, or navigating a city — and it will give you a plausible-sounding list. But it is not actually simulating the task. It is generating text that looks like a plan based on patterns in its training data. Real planning requires predicting consequences, and LLMs cannot do this reliably.
Pillar 2: LLMs have no persistent memory of the physical world. They know that water flows downhill because they have read about it. But they do not have an internal model of fluid dynamics. This means they fail unpredictably on physical reasoning tasks that a three-year-old could solve.
Pillar 3: Scaling LLMs has diminishing returns. The industry has been scaling models by making them bigger and feeding them more data. LeCun argues this approach is hitting a wall — the improvements from each doubling of compute are getting smaller, while the costs are getting larger. A fundamentally different approach is needed.
Critics argue that LLMs are improving fast enough that these limitations will be overcome through scale and clever engineering. The debate is genuinely unresolved. But LeCun is backing his conviction with $3.5 billion.
AMI Labs: What We Know So Far
The company is still in early stages, but here is what has been revealed:
Focus areas: Industrial process control, automation, wearable devices, robotics, healthcare. These are all domains where understanding the physical world is essential — exactly where LLMs struggle and world models should excel.
Core technology: V-JEPA and its successors. The models will train on video and spatial data rather than text, building internal representations of physical reality.
Team: LeCun is recruiting from the top tier of AI research. Several key hires have come from Meta's own AI research division, as well as from academic institutions in Europe and North America.
Meta partnership: Despite leaving, LeCun says Meta will partner with AMI Labs. He thanked Zuckerberg and other executives for supporting the transition. This suggests Meta may use AMI Labs' world model technology to complement its own LLM efforts.
Timeline: No products have been announced. LeCun has indicated that the research phase will take several years. This is a long-term bet, not a startup racing to market.
Could World Models Actually Replace ChatGPT?
The honest answer: not soon, and maybe not entirely.
LLMs are extraordinarily good at language tasks — writing, coding, analysis, conversation. A world model that understands physics but cannot write a coherent paragraph would be useless for 90 percent of what people currently use AI for.
The more likely future is a merger of both approaches. An AI that has both a world model (for understanding physical reality, planning, and reasoning) and a language model (for communicating with humans) would be far more capable than either alone.
LeCun himself has acknowledged this: "LLMs are useful. They are just not sufficient." AMI Labs is not trying to eliminate language models. It is trying to build the missing piece — the part of intelligence that language alone cannot provide.
If successful, the impact would be transformative:
- Robots that can navigate unfamiliar environments without specific programming
- Autonomous vehicles that understand traffic scenarios they have never seen before
- Medical AI that can reason about how drugs interact with the body's physical systems
- Manufacturing AI that can predict equipment failures before they happen
What This Means for India's AI Ecosystem
India's AI industry is heavily focused on LLM applications — chatbots, content generation, coding assistance, and business automation. If world models prove to be the next frontier, India needs to start building capacity now.
Academic opportunity: IITs and IISc are well-positioned to contribute to world model research. The intersection of computer vision (where India has strong talent) and physical AI is exactly where world models live.
Industrial applications: India's manufacturing sector, which contributes 17 percent of GDP, could be transformed by AI that understands physical processes. World models could optimize everything from steel production to pharmaceutical manufacturing.
Startup opportunity: The world model space is wide open. Unlike the LLM market, where OpenAI, Anthropic, and Google dominate, world models have no established leaders beyond AMI Labs. Indian AI startups that move early could claim significant ground.
At Brandomize, we watch where AI is heading — not just where it is today. Whether the future belongs to LLMs, world models, or a fusion of both, we build with the best tools available and adapt as the technology evolves. Explore our work at brandomize.in.