One detail that stood out is that latency and multimodal performance seem to have been major factors in Apple selecting Gemini. It suggests that the practical constraints of on device inference are still a significant bottleneck for assistants like Siri.