I've no problem with the intuition. But I would hope for a lot more focus in the marketing materials on proving the (statistical) correctness of the implementation. 15% better inference speed is not worth it to use a completely unknown inference engine not tested across a wide range of generation scenarios.