Turing Award winner Yann LeCun: Large models are a "dead end". Which path to bet on next?

On November 19, 2025, Turing Award winner Yann LeCun announced that he would leave Meta and turn to founding a new company focused on Advanced Machine Intelligence (AMI).

This is not an ordinary executive job - hopping.

(CNBC: Yann LeCun is about to leave and start his own startup)

This Turing Award winner didn't choose to join the arms race of large language models (LLMs), but instead plunged into a direction that has been neglected for years: the world model.

LeCun used an extreme term: LLMs are a "dead end" on the path to human - level intelligence.

In a public conversation titled "Do LLMs Understand?" on November 23, he directly pointed out that LLMs are good at language expression but lack an understanding of the real world.

Almost simultaneously, Ilya Sutskever, the former chief scientist of OpenAI, also stated in a podcast on November 25 that "the era of 'Just Add GPUs' is over."

Within a week, two deep - learning pioneers questioned the mainstream approach simultaneously.

This is not a coincidence but a signal of a collective shift in the technological route: the post - LLM era is taking shape.

Section 1 | Why does he say large models are a dead end?

Before discussing the world model, we must first understand why Yann LeCun calls LLMs a "dead end."

His answer is more systematic than the outside world thought.

① The models are getting bigger, but understanding isn't keeping up

LeCun's exact words were: LLMs perform well at the language level, but they don't understand the world. They lack common sense and causal relationships and are just a stack of a large number of statistical correlations.

In other words, scale can make the model more like a person who can talk, but it can't make it more like a person who understands the world.

In fact, Meta's Llama 4 is the best example. After its release in April 2025, its performance in real - world scenarios was far worse than in benchmark tests, and developers even questioned that it had over - optimized the evaluation metrics.

This confirms LeCun's judgment: language fluency has improved, but the understanding of the world hasn't kept up.

② The ceiling of LLM capabilities has been revealed in the laboratory

He emphasized in the public conversation that we are seeing performance saturation. A larger model doesn't necessarily bring higher real - world intelligence.

The training data is approaching its limit, the computing power cost is rising exponentially, and the understanding isn't improving synchronously.

This is what he calls the dead end: continuing to pile on computing power leads to diminishing marginal returns.

Ilya Sutskever, the former chief scientist of OpenAI, also expressed a similar view in an interview: simply increasing the computing power scale by 100 times won't bring about a qualitative change.

③ Language is a by - product, and the physical world is the core of intelligence

His core view is:

Language is a by - product of human intelligence, not the core mechanism.

The logic behind this statement is that language only describes a part of the world, and true intelligence comes from modeling, predicting, and acting in the physical world.

But LLMs can't do this. They don't even understand why a cup won't pass through a table. They know the rules in language but not the rules of the world.

The design inspiration for airplanes comes from birds, but it's not a simple imitation of the way birds fly. Similarly, intelligence isn't produced by imitating the surface rules of language.

④ LLMs can't plan, let alone act

LeCun's main criticism is that LLMs only seem smart in conversations, but their capabilities drop sharply when it comes to multi - step reasoning, long - term planning, and embodied interaction.

He gave a striking comparison:

A teenager can learn to drive in 20 hours. But we still don't have level 5 autonomous driving.

A child can clean the table and load the dishwasher on the first try. But we don't even have robots that can do housework.

These comparisons show that intelligence isn't the ability to talk but the ability to act, which is precisely the weakness of LLMs.

LeCun's logic isn't to oppose large models but to believe that predicting language won't lead to the end - goal.

To enable AI to truly have the ability to understand, reason, and act, a new architecture is needed.

Section 2 | The world model: How should the next - generation AI view the world?

If language models can't understand the world, how should we build true intelligence?

LeCun's answer is to let AI learn to see the world.

He pointed out that future AI must be able to build an internal understanding of the world from multi - modal inputs, just like humans and animals, and then predict and act based on this understanding.

This ability is lacking in GPT - 4, Claude, and Gemini. But cats have it, infants have it, and humans have it.

① What is the world model?

LeCun explained that we train language models by predicting the next word because the vocabulary in language is limited and can be enumerated. But the real world is infinitely rich, and predicting the future at the pixel level is simply impossible.

The real world is a high - dimensional, continuous, and chaotic sensory stream. Humans don't understand the world by predicting the next word but by observing, memorizing, and summarizing to form an internal projection of an abstract world in their minds.

For example:

An infant doesn't need someone to tell them what gravity is; they understand it after dropping things a few times.

A cat doesn't need language instruction; it knows how high to jump to reach the table after observing a few times.

Humans can master driving in 20 hours, not by memorizing rules but by building an intuitive model of speed, distance, and inertia.

What LLMs lack is this projection space; they don't have an internal representation of the world.

This is the new path LeCun is building: Joint Embedding Predictive Architecture (JEPA), a joint embedding prediction architecture.

② JEPA: A brand - new learning paradigm

The core differences between JEPA and LLMs are reflected in multiple aspects.

• In terms of input form, LLMs only process language tokens, while JEPA can process multi - modal data such as videos, images, and sensor data.
• In terms of learning goals, LLMs predict the next word, while JEPA predicts changes in abstract states.
• In terms of learning methods, LLMs rely on discrete sequence modeling, while JEPA combines representation learning and causal modeling.
• Most importantly, LLMs don't have the ability to act, while JEPA naturally has planning and execution interfaces.

LeCun used a vivid analogy: Using an LLM to understand the real world is like teaching someone to drive by just talking. You can memorize all the traffic rules, but you'll never learn to drive for real. Because language can't describe the feelings of friction, inertia, and blind spots, which are the core of action intelligence.

③ Start from the simulated world to train the next - generation AI

What LeCun is promoting at AMI is an AI training model similar to animal learning: first, let the AI interact autonomously in a simulated environment, then extract causal relationships from the interactions, form continuous memories, and finally gain the ability to plan actions.

This model no longer relies on more tokens but on a better world model.

He said: We don't need an AI that can recite encyclopedias; we need an AI that can understand the world with its eyes and hands.

If LLMs are masters of language, the world model is an apprentice in the physical world.

Yann LeCun chose to bet on the latter. This is not only a fork in the technological route but also a redefinition of the essence of AGI.

Section 3 | Not just LeCun: Exploration in another direction

LeCun isn't the only one questioning the path of LLMs. Sutskever also believes that the scaling era is over, and the next - generation intelligence needs a new architectural foundation.

The two deep - learning pioneers have reached a consensus, but the answers they provided are completely different.

① LeCun bets on the world model, and Sutskever bets on safe super - intelligence

LeCun's direction is clear: enable AI to have the ability to understand and act in the physical world. Through self - supervised learning, representation modeling, and causal prediction, build a system that can truly see and understand the world. He predicts that a prototype of embodied AGI will appear within 10 years.

Sutskever's focus is on the other side: the generalization ability of current AI systems is far inferior to that of humans. They perform well in benchmarks but are prone to getting into error loops in real - world scenarios. If this vulnerability isn't resolved, the larger the scale, the higher the risk. He founded the SSI company to ensure that AI remains safe and controllable while its capabilities continue to improve.

In a nutshell, LeCun wants to teach AI to understand the world and act, while Sutskever wants to make AI controllable as it becomes more powerful.

② Different concerns behind the two routes

This divergence stems from their different focuses.

LeCun is concerned about how AI can effectively generalize and act in the real world. He emphasizes that what we lack isn't computing power or data but architecture.

Sutskever is concerned about the safety and controllability of AI. He believes that simply pursuing the improvement of capabilities is dangerous before resolving the vulnerability of generalization.

They represent two directions in the post - LLM era: the architectural innovation school and the safety - first school.

In the past decade, the competition in AI was about model scale and training data. But when the two pioneers left the big companies one after another, they told us that the rules have changed.

In the next stage, the competition will be about who can invent a new architecture first and whose system is both powerful and controllable.

This is the end of one era and the start of another.

Section 4 | A shift is happening

When a Turing Award winner publicly questions the mainstream approach, when OpenAI launches a hardware project, when Google hires the CTO of Boston Dynamics, and when billions of dollars in investments start flowing into embodied intelligence, a question emerges: what will the post - LLM era look like?

① The quiet shift in the industry

Although LLMs are still developing rapidly, some key changes are already taking place.

OpenAI's hardware ambitions are surfacing. On November 24, the company confirmed that its first AI hardware prototype was completed, which is the result of cooperation with Jony Ive, the former chief designer of Apple. According to the plan, this screenless AI device will be released within two years, completely changing the way humans interact with AI.

Google's multi - route strategy is also worth noting. It released Gemini 3 Pro on November 18 and hired Aaron Saunders, the former CTO of Boston Dynamics, on November 21 to turn Gemini into a general - purpose robot control platform. The goal is to make the same model adaptable to robots of any form and ready to use out of the box.

After raising $230 million in financing, Fei - Fei Li's World Labs released its first commercial product, Marble, a generative world model platform, on November 12.

The field of embodied intelligence is even more lively: Figure AI is valued at $39 billion, and Tesla Optimus plans to start mass - production in 2026.

These actions point to a consensus: the next - generation AI won't only exist in dialog boxes.

② Both routes need time

Neither LeCun's world model nor Sutskever's safe super - intelligence is a direction where results can be seen in the short term.

LeCun said it would take several years to a decade, and Sutskever said it would take 5 to 20 years. This means that the current LLMs are still the basis for mainstream applications. GPT, Claude, and Gemini will continue to iterate and serve hundreds of millions of users.

But the long - term technological high - ground may not be on this path. Whoever makes a breakthrough in the new architecture first will gain the right to speak for the next decade.

This is a marathon that requires patience, not a sprint.

③ What does it mean for entrepreneurs and developers?

LeCun's shift conveys several important signals:

First, don't blindly believe in scale. A larger model doesn't equal better intelligence, and there is still a huge space for architectural innovation.

Second, there are opportunities in vertical scenarios. The world model may first be implemented in fields that require physical interaction, such as robotics, autonomous driving, and industrial control, rather than in general - purpose AGI.

Third, open - source is still important. LeCun has always been a firm supporter of open - source, and his new company, AMI, will continue this route, which means that small teams also have the opportunity to participate in the exploration of the new paradigm.

Finally, be prepared for the long term. This isn't a direction where returns can be seen in one or two years, but it may be the most important direction in the next decade.

LeCun once said that true intelligence isn't on the surface of language but in the deep understanding of the world.

This isn't a negation of LLMs but a greater imagination for the future of AI. Large models have proven the power of scale, but the next breakthrough may come from a completely different architecture.

True AGI won't be confined to dialog boxes but will appear in systems that can understand the world and execute tasks.

The exploration on this path has just begun.

📮 Original article links

https://jannalevin.substack.com/p/do-llms-understand-ai-pioneer-yann

https://www.businessinsider.com/openai-cofounder-ilya-sutskever-scaling-ai-age-of-research-dwarkesh-2025-11

https://techcrunch.com/2025/01/23/metas-yann-lecun-predicts-a-new-ai-architectures-paradigm-within-5-years-and-decade-of-robotics/

https://www.abzglobal.net/web-development-blog/ilya-sutskever-yann-lecun-and-the-end-of-just-add-gpus

https://www.bloomberg.com/opinion/articles/2025-11-12/yann-lecun-meta-and-mark-zuckerberg-pick-groupthink-over-ai-godfather

This article is from the WeChat official account "AI Deep Researcher", author: AI Deep Researcher, published by 36Kr with permission.