Earlier this year, Jeff Dean, the leader of Google AI, delved into the concept of “distillation” during a podcast episode. He explained that Google stumbled upon AI distillation technology while refining its AI models. The core idea behind this technique is to enable a “student model” to learn from a “teacher model”, facilitating knowledge transfer and model compression. This approach not only boosts system performance but also minimizes dependence on a single, large-scale model.
However, the technology soon found itself at the center of controversy. Some companies began misusing it for what are now known as “distillation attacks.” These attacks involve bombarding AI models with countless repetitive queries to extract their inner workings. The goal? To clone these models or fortify their own AI systems—a practice widely viewed as intellectual property theft.
