Recent developments in the artificial intelligence landscape reveal a growing conflict surrounding "distillation," a process where new AI models are trained by leveraging publicly accessible chatbots and APIs of existing models. This technique has drawn attention primarily due to Chinese firms reportedly using it to develop open-weight models that rival U.S. offerings in capability but at significantly reduced costs. However, it has been widely speculated that American AI labs might also be employing these techniques on each other to maintain competitive parity.

Confirmation of such practices emerged during a California federal court trial where Elon Musk, founder of xAI, testified. When questioned about xAI's use of distillation techniques on OpenAI models to train its Grok AI, Musk affirmed it as a "general practice among AI companies," responding "Partly" when asked for a direct "yes." This admission is particularly notable given Musk's ongoing lawsuit against OpenAI, CEO Sam Altman, and Greg Brockman, alleging a breach of OpenAI's foundational nonprofit mission by transitioning to a for-profit structure, a trial which commenced this week.

Musk's testimony underscores the significant threat distillation poses to established AI giants. By enabling other software makers to create nearly equally capable models cheaply, distillation undermines the substantial advantage built by frontier labs through their vast investments in compute infrastructure. This situation is fraught with irony, considering the controversies surrounding alleged copyright infringements by these same frontier labs in their quest for sufficient training data. It is unsurprising that xAI, founded in 2023, years after OpenAI, would seek to learn from and adapt techniques pioneered by leaders in the field.

While the explicit legality of distillation remains unclear, it is more likely to fall under violations of terms of service set by companies for the use of their products. In response to these challenges, OpenAI, Anthropic, and Google have reportedly initiated a collaborative effort through the Frontier Model Forum. This initiative aims to share information and strategies to combat distillation attempts, particularly those originating from China. Their approach typically involves identifying and preventing suspicious mass queries that are characteristic of systematic model probing to understand internal workings.

Further into his testimony, Musk addressed a prior claim about xAI's future standing among AI providers. He revised his assessment, ranking Anthropic as the current leader, followed by OpenAI, Google, and then Chinese open-source models. He characterized xAI as a much smaller entity, employing only a few hundred individuals, indicating a more modest current position compared to his earlier predictions.