Elon Musk testified on Thursday in a California federal court that his AI startup, xAI, employed a technique known as "distillation" to train its Grok chatbot using OpenAI's models.
Distillation involves extracting knowledge from a larger, more capable AI model—often by prompting it extensively—to train a smaller, cheaper model. While OpenAI and Anthropic have criticized Chinese firms for using distillation to create competitive open-weight models, Musk's admission reveals that the practice is also common among US labs.
During the trial, where Musk is suing OpenAI, CEO Sam Altman, and Greg Brockman for allegedly abandoning the nonprofit mission, Musk was asked whether xAI had used distillation on OpenAI models. He replied that it is a general practice among AI companies, and when pressed for a "yes or no" answer, he said "Partly."
The strategy of distillation threatens major AI labs by undermining the competitive advantage they gain from expensive compute infrastructure. Ironically, this comes as OpenAI itself faces scrutiny for alleged copyright violations in its training data.