Yoshua LeCun, a Turing Award winner, finally warned Meta: I've been working on AI for 40 years, and large models are a dead end.

The wind at Meta has changed, and Yann LeCun admits he's leaving soon!

According to reports from multiple authoritative media outlets, Yann LeCun, Meta's Chief AI Scientist in charge of "Fundamental AI Research" (FAIR), is expected to leave the company soon.

This 65-year-old veteran in the AI field holds a core position at Meta, one of the world's largest technology companies, and has access to seemingly limitless resources.

Meta has been extremely generous with its spending. It has been poaching top AI experts from its competitors with sky-high salaries.

In July, Mark Zuckerberg even claimed that "superintelligence is within reach."

So, why is LeCun leaving Meta? Is it just because of the personnel upheaval at Meta? What's the hidden story behind it?

Has Zuckerberg Changed Direction and LeCun Lost His Influence?

This summer, 28-year-old Alexandr Wang became Meta's Chief AI Officer, making this young and enthusiastic advocate of large language models LeCun's superior.

In addition, Meta also appointed another relatively young Chief Scientist, Shengjia Zhao, whose position is also above LeCun's.

In an official announcement, Meta highly praised Shengjia Zhao's "breakthroughs" in scaling. However, LeCun has lost confidence in scaling.

He also advised doctoral students: "Don't work on LLM."

If you're curious why both LeCun and Zhao are Chief Scientists, it's because Meta's AI department has a rather unique organizational structure, divided into multiple independent teams.

The media has been reporting that Meta is going to restructure its AI organization.

Last month, Meta's super AI lab laid off hundreds of employees, including 10-year veteran Yuandong Tian. It's said that this was to sort out the chaotic situation.

This is already the fourth adjustment of Meta's AI business within half a year.

The once-glorious FAIR, which was led by LeCun, has lost its former luster. According to current and former employees, this department has experienced layoffs, budget cuts, and a significant decline in internal influence.

There was a time when FAIR was the most intellectually active "ivory tower" within Meta. Researchers could explore various future paths of AI and even conduct experiments that "might not succeed," without having to worry about productization.

Now, Meta's newly established AI research department has recruited a large number of high - paid new employees, led by Wang, with clear goals: to be fast, to achieve real - world applications, and to productize.

He's Been Ahead in AI for 40 Years

LeCun has always been at the forefront of the times -

He started researching "machine learning" when it was not yet recognized by the mainstream.

He once worked in Geoffrey Hinton's lab in Toronto when Hinton was not yet an AI legend.

After that, most of his career was spent at Bell Labs in New Jersey, an institution famous for numerous innovative inventions.

In 1947, Bell Labs invented the transistor.

"The most exciting thing for me is to work with people smarter than me because it magnifies your abilities," LeCun said in a 2023 magazine interview.

At Bell Labs, LeCun was involved in the development of handwriting recognition technology, which was later widely used in banks to automatically read checks. He also participated in a project to digitize paper documents and distribute them over the Internet.

LeCun has said that he was interested in physics since childhood and mainly collaborated with physicists at Bell Labs, reading many physics textbooks.

I learned a lot of things that seem unrelated to AI or computer science on the surface (I majored in electrical engineering in college and had little formal training in computer science).

In 2003, LeCun began teaching computer science at New York University and later became the founding director of the university's Center for Data Science.

In 2013, Zuckerberg personally invited him to join Facebook (before it was renamed Meta) to establish a new AI lab.

He led the team for four years and stepped down in 2018, becoming the company's Chief AI Scientist and continuing to explore the technological frontier as an "individual researcher."

In 2018, he, along with Geoffrey Hinton and Yoshua Bengio, won the Turing Award - the highest honor in the computer field - in recognition of their foundational work in neural networks.

Since then, LeCun has gradually become a "symbolic figure." He was not involved in the development of Meta's first open - source large language model, Llama, and has long since stopped participating in the daily work of such projects.

According to people who have worked with him, LeCun is now mainly working on his own research projects and often attends various technology conferences to share his views on AI technology.

Facing media reports, Yann LeCun only pointed out "minor errors" in the reports and did not deny the news of his upcoming departure.

He knows that he has been marginalized in the entire Silicon Valley technology circle, including at Meta. At a seminar at MIT last month, 65 - year - old LeCun said bluntly:

Over the years, I haven't been very popular in many parts of Silicon Valley, including Meta, because I've been saying that within 3 to 5 years, world models will become the mainstream AI architecture, and no one will want to use the current LLM anymore.

But he firmly believes in his judgment of the future of AI. His old friend Léon Bottou once told the media that LeCun is "cutely stubborn" - he will listen to others' opinions but has his own firm beliefs.

Now, it seems that LeCun can "no longer bear it" at Meta and is finally leaving.

In fact, he has hinted at the answer many times.

On the path to general artificial intelligence, LeCun has recently become well - known for his sharp criticism of large language models.

He believes that no matter how much tech giants scale them up, the large language models as we currently understand them are "at the end of their rope," a "detour, a distraction, and a dead end."

He has been involved in AI research for 40 years, and his judgments about AI have often come true. Now, he believes that most people are wrong.

He has laid many foundations for modern AI. Now he firmly believes that most people in the field have been led astray by the "siren song" of large language models.

This provides more possible explanations for his departure.

LeCun May Leave Meta for World Models

Previous reports indicated that he is discussing with industry peers to start a company, seek investment, and form a team focused on world models.

The so - called "world model" is similar to how small animals or infants actively learn the laws of the world through visual and other sensory data; while LLM is just a model that relies on a large amount of text for prediction.

LeCun himself has never shied away from explaining why he believes that "world models" are the answer to AI.

Models like Meta's Llama, OpenAI's GPT, and Google's Bard are all trained with a large amount of data. LeCun estimates that it would take about 100,000 years for a person to read all the text required for their training.

But the main way humans learn is not by reading text.

We obtain much more information from interacting with the world. LeCun estimates that the amount of data a normal four - year - old child has been exposed to is 50 times that of the largest current LLM.

Most human knowledge is not in the form of language.

So these systems will never reach human - level intelligence unless you completely change their architecture.

And he himself has long prepared an alternative. He calls it "objective - driven AI."

The objective - driven AI system is built to achieve specific goals set by humans.

Unlike being driven only by pure text data, they are trained through sensor and video data to understand the physical world.

The "world model" thus constructed can present the impact of actions, and all potential changes are updated in real - time in the system's memory.

Why Is He So Obsessed with World Models?

At the "Paris AI Summit" at the beginning of the year, Yann LeCun clearly stated that he is a firm believer in wearable devices.

He believes that in the future, we need to interact with wearable devices just as we communicate with people, and large language models don't understand the world like humans do.

With large language models, we can't even replicate the intelligence of a cat or a mouse, let alone a dog.

These animals can accomplish amazing feats because they understand the physical world. Any domestic cat can plan extremely complex actions because they have a causal model of the world.

To illustrate this point, LeCun designed a thought experiment: "Imagine a cube floating in the air in front of you. Okay, now rotate this cube 90 degrees around the vertical axis. What will it look like?"

He believes that any human can easily do this, while large language models are powerless:

"It's very easy for a person to build a mental model of a rotating cube in their mind."