Yann LeCun confirms Meta's Llama 4 benchmarks were "fudged a little bit" (tech.slashdot.org)

Yann LeCun, Meta's outgoing chief AI scientist, has revealed that the benchmarks for the Llama 4 language model were “fudged a little bit” prior to its April 2025 launch, indicating that different models were used selectively to improve performance results. This admission comes after Llama 4 faced criticism and accusations of inflated capabilities at its release, prompting CEO Mark Zuckerberg to express significant disappointment and distrust towards the team involved. As a result, Zuckerberg has reportedly sidelined the entire GenAI organization, leading to increased turnover within the team. LeCun's departure from Meta marks a significant shift, as he has been at the forefront of AI research for over a decade. He is set to launch a new venture called Advanced Machine Intelligence Labs, emphasizing a critical stance on the future of large language models (LLMs), which he views as inadequate for achieving true superintelligence. This incident not only raises questions about ethical practices in AI research and development but also highlights a potential shakeup within Meta's AI strategy at a time when transparency and reliability are paramount in the rapidly evolving field of artificial intelligence.