Elon Musk’s xAI Secretly Distilled Anthropic’s Claude to Train Grok Coding Models
Despite Elon Musk's public posture of building independent frontier AI, a report from The Information has revealed that his startup, xAI, spent months using outputs from Anthropic's Claude to train and distill its own Grok coding models.
When Anthropic detected the activity and revoked xAI's official API access in January, xAI engineers reportedly bypassed the ban by continuing the distillation project underground. They utilized personal accounts and the intermediary service Blackbox AI to maintain access to Claude's data.
The revelation highlights a chaotic year for xAI's engineering team, which has struggled to keep pace with rivals. The company's pretraining team has reportedly shrunk to under five people, and four of its Grok code leads have departed in recent months alongside several co-founders. Furthermore, an employee accidentally deleted critical training data, costing the company two to three weeks of work. Ironically, while xAI was covertly scraping Anthropic's models for training, Anthropic was simultaneously negotiating a massive $1.25 billion per month deal to lease GPU capacity from SpaceX's data center in Memphis, Tennessee.
Verbatim Quotes
"Elon Musk's xAI spent months distilling Anthropic's Claude to train its own coding models... After Anthropic revoked official access in January, xAI engineers kept going through personal accounts and the intermediary service Blackbox AI.1" — The Decoder
"Internally, xAI looks troubled. The pretraining team shrank to under five people. Four Grok code leads left within months... One employee accidentally deleted critical training data, costing two to three weeks of work, according to The Information." — The Decoder
-
An instance of Platform bans cannot prevent rivals from distilling superior model outputs to train their own. — xAI quickly bypassed Anthropic's platform bans and API blocks to continue distilling Claude outputs for its own coding systems. ↩︎