So let's see if I get this right: the report says Apple is paying Google $1B per year to build some 1.2T parameters model to power Siri and other resource-intensive tasks, besides the native local ~3B model that already comes preinstalled on devices (Mac & iPhone).
They announced in WWDC24 that Apple is all in on building foundational models to do what is now offloaded to Google. Apple’s going to retain user data by deploying models on private cloud compute clusters.
Big win for G and finally Apple might have saved its AI infra from this partnership. After building on the foundational model framework for quite some time and testing the on-device features – I’ve had a mixed experience to be honest. There have been no major updates on the MPS backend and by looking at PRs – the development seems to have stalled.
Without doubt, Apple Silicon is a great architecture. There has been promising work by the MLX team and I hope they continue to improve training and inference on the silicon. But the flip side is a sign of incompetence from Apple. Seeing the trajectory of how Microslop is sabotaging itself with Windows copilot features,
Apple could've taken some more time, hired engineers to work on training models in-house. Users don't need more AI unless someone cracks a powerful AI-native solution on edge. Sure ..i’ll be fine as long as there is an option to disable Siri.
