Meta Spent $72 Billion on AI, then the Co-Inventor of Deep Learning Quit and Revealed They Rigged Their Benchmarks.
Inside Yann LeCun’s explosive departure: the fudged Llama 4 scores that made Zuckerberg furious, the 28-year-old now running legendary researchers, and why $72B can’t buy direction
$72 billion. That’s what Meta spent on AI in 2025 alone. An astronomical sum that could fund several moon missions.
And yet, despite these Pharaonic investments, something isn’t working at the Menlo Park giant.
A few weeks ago, one of AI’s most respected scientists slammed the door on Meta. Yann LeCun, the co-inventor of deep learning and a key figure in modern AI, has resigned after 12 years.
What he revealed on his way out sent shockwaves through the entire sector. And what we learned goes far beyond simple strategic disagreements.
The Rigged Benchmark That Made Zuckerberg Furious
LeCun didn’t just leave politely. In an interview with the Financial Times, he revealed that the benchmarks for Llama 4, Meta’s flagship model launched in April 2025, had been rigged.
In his own words: “The results were somewhat fudged.”
Here’s what happened. The team used different versions of the model for divergent tests to obtain better scores. Meta presented the world with a version of its AI that wasn’t the one made available to the public.
It’s like a car manufacturer announcing 5 liters per 100 km after testing the vehicle on a downward slope with the engine off.
Mark Zuckerberg’s reaction was swift. According to LeCun himself, the CEO was furious and lost confidence in the entire team involved. Result? The GenAI organization was sidelined, and that’s not a figure of speech. In October 2025, Meta laid off 600 employees from its AI division, a good portion from the FAIR team, the very one LeCun had founded and led.
“Many people have left, and many who haven’t left yet will leave,” he declared.
What makes this situation particularly ironic is that LeCun had been defending a radically different approach from the one the industry adopted. While everyone bet on large language models, those famous LLMs powering ChatGPT, Claude, Grok, Gemini, and LeCun kept repeating that this path was a dead end for achieving true artificial intelligence.
His conviction? We need what experts in the field call world models. Systems capable of understanding physics, maintaining persistent memory, and planning complex actions, and not simply predicting the next word in a sentence.
The $3.5 Billion Bet Against His Former Employer
Here’s where it gets pretty exciting.
