LeCun exposed that Meta cheated on the rankings. Tian Yuandong: I didn't expect this ending.

After leaving Meta, Yann LeCun struck back hard: the widely panned Llama 4 really cheated on the rankings!

It's true that the results were tampered with a bit. To get better results, the team used different models for different benchmark tests.

They say don't provoke those who have left the company. Not only Yann LeCun, but Tian Yuandong also revealed some inside information in his personal year - end summary.

I was dragged in to rescue Llama 4. I anticipated four possible scenarios in advance, but Mark Zuckerberg gave me a fifth...

(Smiling wryly with a hand on the forehead.jpg)

Anyway, as for their plans after leaving, both great minds said in unison:

Start a business!

The wave of departures triggered by Llama 4's ranking fraud

DeepSeek put too much pressure on Mark Zuckerberg.

Previously, there were reports that the emergence of DeepSeek left Llama 4 far behind even before its release, forcing Mark Zuckerberg to frantically increase AI investment.

This was also confirmed by Yann LeCun.

Mark Zuckerberg was really panicked. He put pressure internally, asking the GenAI department to accelerate AI development and deployment, and dragged in Tian Yuandong from the former FAIR team to rescue the situation.

As a result, the communication between teams broke down completely. Yann LeCun and his colleagues wanted to do something new, but Mark Zuckerberg wanted proven and directly applicable technologies.

The real fuse was the fiasco of Llama 4.

Not only did it lose, but its reputation also plummeted due to ranking fraud. So Mark Zuckerberg completely lost confidence in the entire team and marginalized them all.

This directly led to a major reshuffle at Meta: on one hand, poaching people at sky - high prices from various Silicon Valley companies, and on the other hand, massively laying off old employees.

Tian Yuandong and his team were the first to be affected.

They were forced to join the Llama 4 project in January, and were discarded right after Llama 4.5 was trained in October.

Tian Yuandong really has a grievance to air...

Due to the pressure from above, the whole group was forced to put down their ongoing work and take over the mess of Llama 4. There were only two months left until the release deadline, and they had to handle all the dirty work.

For this reason, Tian Yuandong even drew a 2x2 return matrix at that time to calculate the four possible outcomes of this task:

But since the boss had spoken, they had to do it. Tian Yuandong thought he would just do his best and have a clear conscience.

After working hard for several months, Mark Zuckerberg didn't choose any of the four scenarios and came up with Plan E: without holding the person - in - charge accountable, he just kicked Tian Yuandong and his team members out.

No wonder when Tian Yuandong reviewed this experience, he also said:

This has also given me a deeper understanding of the complexity of society.

Fortunately, Tian Yuandong is open - minded. He said he had been at Meta for more than a decade, and in recent years, he had the mentality of 'I hope the company will fire me soon'. So it can be considered a blessing in disguise, and it also provided a lot of new material for his next novel (doge).

(Ahem) Getting back to the point, these months were not fruitless for Tian Yuandong. He also made some new explorations in the core issues of reinforcement learning training.

First is large - model reasoning. After the team's public release of continuous latent space reasoning (coconut) at the end of 2024 attracted wide attention, the team further clarified the advantages of continuous latent space reasoning through the theoretical work Reasoning by Superposition.

At the same time, they also tried to improve model reasoning efficiency from different angles: Token Assorted reduces computational overhead through discrete tokens in the latent space, DeepConf terminates reasoning early based on confidence, ThreadWeaver creates parallel thought chains to accelerate reasoning, and RL is used to learn reasoning ability in models of different scales.

In terms of interpretability, the research focuses on the Grokking phenomenon, starting from the sudden change process from memory to generalization, trying to explain what the model has learned, its relationship with the input data, and the level of generalization it can achieve, that is, to open the black box of the model.

In short, regarding being dragged in to "take the blame", Tian Yuandong was lenient towards his former employer. Although the company was unfair, he has let it go～

However, Yann LeCun, who left soon after, was not so soft - hearted. He directly criticized Meta for being overly addicted to large - language models (LLMs), especially the members of the newly poached Super Intelligence Lab.

He specifically criticized Alexander Wang: Young and inexperienced.

Although he learns quickly, he doesn't understand research at all, doesn't know how to do it, and doesn't know how to get along with researchers.

In front of Yann LeCun, this 27 - year - old young man is like a little baby.

But large - language models are not as wonderful as they think. Yann LeCun said bluntly that LLMs are useful but essentially limited by language. To put it more exaggeratedly:

LLMs are a dead end.

To achieve human - level intelligence, one must understand the operating laws of the physical world, which is the world model that Yann LeCun has long focused on. But Meta has lost interest in this.

The disagreement in the research direction also forced Yann LeCun to leave the company and start the next chapter of his life - starting a business.

What to do after leaving Meta? Start a business

Yann LeCun's new company is called Advanced Machine Intelligence (AMI), which focuses on the world model he has always cared about, and is fully committed to open - source.

However, according to his own disclosure, he will only serve as the executive chairman of the new company instead of the CEO.

I'm a scientist. I can inspire people to work and guess which technology will succeed, but I'm not good at organization and management, and I'm too old.

Yann LeCun will have the same research freedom as he did at Meta, and the person in charge of leading AMI will be Alex LeBrun, the co - founder and CEO of the French medical AI startup Nabla.

They are focusing on the V - JEPA architecture, trying to understand the physical world by learning video and spatial data, enabling AI to complete planning, reasoning, and have long - term memory, which is what he often calls advanced machine intelligence.

Yann LeCun expects to see the initial version of this technology within 12 months and achieve large - scale progress in the next few years. Although it is far from super - intelligence, at least there is hope on the road to AGI.

As Yann LeCun's former subordinate, Tian Yuandong also rejected the olive branches thrown at him by big companies and has just officially announced starting a business!

The specific details have not been revealed yet, but he said:

While I'm still young, I'll be the co - founder of a new startup.

Anyway, he'll just work quietly for a while.

Reference links:

[1]https://www.ft.com/content/e3c4c2f6-4ea7-4adf-b945-e58495f836c2

[2]https://zhuanlan.zhihu.com/p/1990809161458540818

This article is from the WeChat official account "QbitAI". Author: Focusing on cutting - edge technology. Republished by 36Kr with authorization.