Apple bets on Google Gemini, raises hard questions

Apple dropped two announcements back to back: 'Siri AI', a rebranded and supposedly more capable assistant, and a new AI architecture explicitly built around Google Gemini models. The HN thread on the Gemini news is sharp and skeptical. People want to know whether Apple is using flagship Gemini models behind custom prompts, fine-tuning them, or doing something more architecturally significant. The fact that it's not launching in the EU is being read as a red flag about its ability to meet regulatory requirements.

The pattern here: Apple is essentially confirming it can't build competitive frontier AI on its own in the timeframe the market demands. Choosing Google over Anthropic or OpenAI is strange, given that Google is also a direct competitor in mobile. The worry in the thread is that Apple is locking itself into a dependency that disadvantages them in the model quality race, since they'll always be one step behind whatever Google ships for themselves.

The Core AI Framework thread adds a wrinkle. Apple is simultaneously building serious on-device ML infrastructure, a new way to run PyTorch models across CPU, GPU, and the Neural Engine, possibly replacing CoreML. So the picture is: outsource the frontier stuff to Google, double down on on-device efficiency. That's a coherent strategy, but it depends entirely on the on-device models closing the gap over time.