The failure mode is that only long context tasks benefit, short ones work fast enough with full attention, and better. It's amazing that OpenAI never used them in any serious LLM even though training costs are huge.