Emanuele Fabbiani heads up AI at xtream. He focuses on decoding how large language models actually function under the hood, moving beyond surface applications to understand their internal behaviors.
His work explores the hidden mechanics of models like Llama 2 and Claude 3. He breaks down complex phenomena—like monosemantic features and cross-lingual bias—to help developers grasp why models memorize personal data and how to mitigate those privacy restrictions.
