xAI Grok Model Sends Repetitive Gibberish to Users During Platform Outage

Subscribers to xAI Grok platform encountered widespread output degradation today after Grok Lite and Grok Pro models began emitting repetitive nonsensical text fragments across search and chat interfaces. Affected users reported that simple queries triggered multi-paragraph loops of random character strings and syntax errors.

According to internal system status updates, the breakdown stemmed from a corrupted KV-cache sync across inference nodes following a routine model weights deployment. The cache desynchronization caused self-attention layers to enter degenerate loop states during sampling execution.

xAI engineering teams rolled back the faulty GPU cluster nodes and restored standard response generation within two hours, adding pre-sampling token sanity checks to prevent loop propagation in future releases.