Two AI21 engineers, Yuval Belfer and Asaf Gardin, detailed on the AI Engineer podcast two silent bugs in vLLM, the open-source inference engine, that corrupted outputs from Jamba, AI21's hybrid transformer-Mamba model. The first, a scheduler error that ran decode before prefill, caused roughly one in 1,000 requests to return gibberish by reading stale memory state. The second, a uint32 integer overflow in a Mamba kernel, silently corrupted log-probabilities during reinforcement learning training. Both produced confident, fluent, completely wrong answers with no crash, warning, or error, and were found through log-probability forensics comparing vLLM's outputs against a vanilla Hugging Face baseline rather than stack traces. The fixes were small: a scheduler classification rule and a single data-type change from uint32 to size_t. The engineers warned that as AI models become stateful and complex, the distance between high confidence and silent failure is shrinking, and the only reliable detection tool is rigorous differential testing against simpler implementations.