on april 30, local time, elon musk testified in a california federal court, admitting that his startup xai had used openai’s models to help train its own chatbot, grok. this testimony came in the context of musk’s lawsuit against openai and its ceo sam altman, in which musk accuses the company of having abandoned its original nonprofit mission.
when asked whether xai had employed “distillation” techniques on competitors’ technologies, musk said this is a common industry practice and acknowledged that xai had “partially” used it. distillation is akin to a “teacher–student” training approach: a more powerful large model serves as the teacher, while a smaller, more efficient new model acts as the student. developers use the large model’s responses to train the new system, enabling it to achieve performance close to that of the high-level model. while this practice is not necessarily illegal—many companies use it to develop lower-cost versions of their own models—the controversy arises when the distillation target is a competitor’s model, as it can be seen as taking a shortcut—a strategy that, for a latecomer like xai, can significantly cut down on r&d time and costs.
musk’s admission comes at a time of intense industry tension. recently, companies such as openai, google, and anthropic have been trying to prevent third parties from “distilling” their model outputs, even labeling it as intellectual-property theft. yet musk’s testimony suggests that u.s. ai labs may also be quietly employing the same tactics to maintain their competitive edge. speaking about the competitive landscape, musk believes anthropic currently ranks first, with openai and google close behind. despite grok’s rapid progress, he still describes xai as a much smaller company—with only a few hundred employees, compared with several thousand at its main rivals.
