Yann LeCun confirms Meta's Llama 4 benchmarks were "fudged a little bit"

4 months ago

Yann LeCun, Meta's outgoing chief AI scientist, admitted that the Llama 4 language model's benchmark results were manipulated before its release.
LeCun stated that different models were used for different benchmarks to improve results, calling it 'fudged a little bit.'
Llama 4 was widely criticized at launch, with accusations of benchmark manipulation to make the model appear more capable.
CEO Mark Zuckerberg was reportedly 'really upset' and lost confidence in the team involved, leading to the sidelining of the GenAI organization.
Many employees have left or plan to leave Meta, including LeCun, who is starting a new AI research venture called Advanced Machine Intelligence Labs.
LeCun criticized Meta's new hires as 'completely LLM-pilled,' calling LLMs a 'dead end' for superintelligence.
Comments reflect skepticism about Meta's integrity, leadership failures, and the broader implications of AI development.

Hasty Briefsbeta