Elon Musk reveals that xAI utilized OpenAI’s models for Grok’s training.

In a federal courtroom in California on Thursday, Elon Musk shared insights about his AI venture, xAI, revealing that it has utilized OpenAI’s models to enhance its own technology.

The discussion centered around model distillation, a prevalent practice in the AI industry where a larger model serves as a “teacher” to impart knowledge to a smaller “student” model. While this technique is often employed legitimately within companies to train their own models, it can also be exploited by smaller labs aiming to replicate the performance of larger competitors.

When questioned about his understanding of model distillation, Musk defined it as using one AI model to train another. However, when asked if xAI had distilled OpenAI’s technology, he sidestepped the question, stating that “generally all the AI companies” engage in such practices. When pressed further, he responded with a cautious, “Partly.”

Musk elaborated that it is standard practice to leverage other AIs to validate one’s own. The rise of model distillation has sparked controversy among AI labs, as the legal boundaries regarding its use often exist in a gray area. Companies like OpenAI and Anthropic have accused Chinese firms of distilling their models, with OpenAI expressing concerns about DeepSeek, while Anthropic specifically called out DeepSeek, Moonshot, and MiniMax. Google has also taken measures to combat what it terms “distillation attacks,” viewing them as a form of intellectual property theft that breaches its terms of service.

In a blog post, Anthropic noted, “Distillation is a widely used and legitimate training method. For instance, leading AI labs routinely distill their own models to create smaller, more affordable versions for their clients. However, distillation can also be misused: competitors might leverage it to gain powerful capabilities from other labs much faster and at a lower cost than developing them independently.”