xAI built its coding models on Claude outputs — until Anthropic cut the cord
the information reports xai spent months training on claude outputs before access was revoked. the reliance on rival data highlights the chaos at musk's startup as it races to compete.
xaì has been distilling anthropic's claude model to train its own coding models, a practice that continued for months via personal accounts even after anthropic revoked official access in january. the report details a chaotic internal environment, citing a pretraining team of fewer than five people, the departure of four code leads, and lost training data due to employee error. while musk leases compute to google and anthropic, xai struggles to stand independently, raising questions about its ability to catch up to industry leaders without using their outputs as crutches.