Musk Admits xAI Partially Used OpenAI Models to Train Grok in Court
During the OpenAI lawsuit hearing, Elon Musk testified that xAI "partially" used OpenAI models to train Grok through distillation technology, which is a common practice in the AI industry.
When asked if xAI distilled OpenAI models, Musk initially stated, "All AI companies do this," and then confirmed the "partial" usage.
This testimony occurred in a federal court in California regarding the lawsuit over OpenAI's shift in mission.
Source: Public Information
ABAB AI Insight
Elon Musk, as a co-founder of OpenAI, testified in the lawsuit acknowledging that xAI used OpenAI models to continue the rapid launch of Grok after his departure in 2023. Earlier, xAI trained using a large GPU cluster and real-world data, and this distillation acknowledgment explains the rapid iteration of Grok's early versions, which was used by OpenAI's lawyers in court to counter their accusations against OpenAI.
In terms of capital strategy, xAI uses OpenAI's output as a teacher model for distillation, generating smaller, more efficient student models (the Grok series), motivated by reducing training costs and accelerating catch-up. Meanwhile, supported by the Colossus supercluster, it gradually shifts towards its own data and architecture, transitioning from "leveraging" to independent leadership, moving computational resources from reliance on external models to xAI's self-controlled training loop.
Similar to the Meta Llama series or many startups using GPT distillation to accelerate development, xAI is currently in a mid-stage transformation from early catch-up to independent cutting-edge models, focusing on enhancing real-world reasoning capabilities in versions like Grok 4.
Essentially, this is a technological substitution: compressing the knowledge of a large teacher model (OpenAI) into its own model through model distillation, which lowers the computational power threshold and time for training, allowing xAI to quickly gain high-performance capabilities in the resource-limited early stages. In the context of the lawsuit, it uses the common industry practice as a defense, shifting the pricing power of AI model iteration from a single giant to a multi-company knowledge-sharing structure.