Grok keeps sending gibberish responses to users

Users of xAI’s Grok Lite web interface began receiving incoherent token streams on Wednesday morning, with one example showing a fragment such as “match it without and your they and two for planets can practical and often cheese…” that continued for several paragraphs. The output resembled random word salad and, in some cases, included URLs pointing to reinforcement‑learning research pages. Reports emerged on Reddit and TechCrunch, though the issue could not be reproduced in the outlet’s own tests, indicating that only a subset of sessions were affected.

The glitch appeared limited to direct queries made through Grok.com; the Grok account on X.com remained unaffected. Refreshing the chat or starting a new conversation typically cleared the nonsense, although a few users reported persistence after several reloads. xAI’s status page (https://status.x.ai) showed all services as operational, and the official Grok account on X characterized the problem as a “rare temporary generation glitch,” advising users to regenerate or start a fresh chat. Key takeaways: the issue affected only Grok Lite, manifested as nonsensical token output with occasional research‑site links, and was resolved by session reset.

Why this matters

Source fact: the glitch was confined to Grok Lite and disappeared after refreshing the session or starting a new chat. Inference: the behavior suggests a transient fault in the model’s generation pipeline—such as a mis‑aligned token‑probability buffer or a temporary corruption in the KV‑cache—that causes the sampler to emit random strings until the internal state is reset. This underscores the need for robust state‑management and health‑checking mechanisms in large‑scale language‑model serving systems.