The industry applauds LLMs. Yann LeCun is already betting on what comes next.
While much of the market is still confusing verbal fluency with general intelligence, Yann LeCun is reopening a debate many people were ready to close far too early.
His startup AMI, founded after leaving Meta, has raised $1.03 billion with an explicit goal: challenge the dominance of the LLM paradigm. The ambition is not to build a slightly sharper chatbot. It is far more radical: to build systems able to reason, plan, and understand the world — in other words, world models rather than systems locked into statistical next-token prediction. That is how both Reuters and Wired describe AMI’s promise.
That is exactly why this moment matters.
Because what is happening here is not just a technology battle.
It is a battle over how reality itself should be modeled.
LLMs are impressive. That does not make them the final destination.
LLMs have transformed the interface between humans and machines. They summarize, write, translate, code, assist, and accelerate. Their impact is real. So is their adoption. Stanford’s AI Index 2025 shows that enterprise AI usage rose sharply in 2024 and that private investment in AI, especially generative AI, remains extremely strong.
But the history of technology is full of moments when market success was mistaken for the endpoint of progress.
A product can dominate a cycle
without being the final architecture of intelligence.
That is the value of LeCun’s move. He is not simply saying, “I want to do better.”
He is saying something much more unsettling: “You may be optimizing the wrong path.”
This critique is not new for him. As early as 2016, he was already arguing that truly robust AI systems would need an internal model of the world capable of prediction, reasoning, and planning. In 2022, in A Path Towards Autonomous Machine Intelligence, he formalized that direction around predictive world models and agents able to act across multiple time horizons.
In other words:
LLMs speak about the world very well.
LeCun’s bet is to build systems that capture its structure more deeply.
It is a powerful intuition. It is not yet a completed innovation.
This is where enthusiasm needs discipline.
In my book, chapter 3, I stress that creativity, invention, and innovation are not the same thing, and above all that “without implementation, we cannot speak of innovation.” The same chapter also makes clear that innovation requires real implementation, often through a team, not just a compelling idea or a brilliant narrative. P001-304-9782100876556_ep05 – f… P001-304-9782100876556_ep05 – f…
That distinction matters.
AMI is, at this stage:
a strong vision,
a bold thesis,
a spectacular funding story,
a very powerful strategic signal.
But it is not yet, at this point, a validated market innovation or a demonstrated large-scale shift. Reuters reports a startup targeting automotive, aerospace, biomedical, and eventually consumer robotics. Wired describes ambitions around world models and persistent memory. All of that signals direction. Not yet proof of a new dominant architecture.
And that is fine.
Industries also move forward because some people reopen questions the market thought were already settled.
The real signal may not be technical. It may be intellectual.
What I find most valuable here is not that a famous scientist raised a lot of money.
It is that he is reintroducing intellectual tension into a sector that loves herd behavior.
When a paradigm dominates, it attracts everything:
capital,
talent,
imitators,
narratives,
conferences,
demos.
Then the danger appears: everyone starts improving the same thing, with surface-level variation, and calls that a revolution.
LeCun forces the ecosystem to ask an uncomfortable question:
what if the next major leap is not about making machines speak even better, but about making them model reality more deeply?
That question does not erase the value of LLMs.
It simply puts them back into perspective.
A dominant tool is not always the final horizon.
It can be a highly profitable, highly visible, highly useful… transitional stage.
Recent AI history gives weight to that caution
Even in domains where models are improving quickly, their limits show up as soon as we expect more than high-probability phrasing. A 2025 Scientific Reports paper, for example, found clear limitations in several LLMs on clinical reasoning tasks, including weak medical commonsense, hallucinations, and overconfidence.
That does not mean LLMs are “finished.”
It means they should not be confused with general intelligence already in motion.
Markets love shortcuts.
Serious research is much less patient with them.
Where many see a model war, I see a strategy lesson
The most interesting point for leaders may not even be technical.
It is the strategic discipline behind this kind of bet.
Challenging a dominant paradigm requires three things:
1. Refusing the seduction of consensus
When everyone runs in the same direction, the temptation is to treat that as proof.
Often it is just social validation at scale.
2. Accepting that a market can over-optimize a partial solution
A technology can create massive value while still being incomplete.
Economic history is full of profitable intermediate stages.
3. Understanding that innovation is not repetition
In my book, chapter 9, I underline that an organization’s ability to innovate rests on a culture that must be deliberately maintained, and that psychological safety is essential to innovative team performance. Bets like this do not emerge from cultures obsessed with safe repetition. P001-304-9782100876556_ep05 – f…
Any sector that stops tolerating contradiction eventually refines its habits instead of preparing its next rupture.
The biggest mistake would be to caricature the debate
Framing this as “LLMs versus world models” as if one must permanently defeat the other would be simplistic.
The most credible future may not be replacement.
It may be hybridization.
Very powerful language systems for interface, abstraction, and knowledge access.
World models for prediction, planning, action, robustness, and autonomy.
So the useful debate is not:
“Are LLMs dead?”
The useful debate is:
“What is still missing for current systems to move from convincing conversation to operational understanding?”
On that front, LeCun has already done something important: he has forced the industry to look beyond its own reflection.
What this moment really teaches us
Markets love technologies that are easy to see.
Deep innovation often starts with architectures that few people fully understand at first glance.
AMI may not yet have proved that LLMs have hit their ceiling.
But it has already achieved something important: it reminds us that a dominant paradigm is never safe from a more ambitious successor.
In periods of euphoria, lucidity becomes a competitive advantage.
And in AI, as elsewhere, the decisive question is never only:
“What works today?”
The useful question is:
“What will truly prepare tomorrow?”
References
(Reuters) = https://www.reuters.com/business/ex-meta-ai-chief-yann-lecuns-ami-raises-103-billion-alternative-ai-approach-2026-03-10/
(Wired) = https://www.wired.com/story/yann-lecun-raises-dollar1-billion-to-build-ai-that-understands-the-physical-world/
(OpenReview) = https://openreview.net/pdf?id=BZ5a1r-kVsf
(Meta Engineering) = https://engineering.fb.com/2016/06/20/ml-applications/a-path-to-unsupervised-learning-through-adversarial-networks/
(Stanford HAI) = https://hai.stanford.edu/ai-index/2025-ai-index-report
(Scientific Reports) = https://www.nature.com/articles/s41598-025-22940-0
