Issue 639  December 8th 2023

You’ll almost certainly have seen links to the new mlx repository during this week. It’s a new project from a machine learning research group inside Apple, so it makes sense that the repository description is “An array framework for Apple silicon”. That’s certainly a description written by a research group rather than a marketing department, especially when you look into what it does.

The examples repository shows more of what it’s capable of, with samples covering transformer language models, LLaMA, LoRA, Stable Diffusion, and OpenAI’s Whisper. There’s also a quick start guide and more documentation if you want to explore further.

Why is this significant? We have all those things already, right? We do, but think about how and where Apple likes to do most of its AI processing: On-device.

This library is optimised for Apple Silicon chips, and speeding up AI models on the devices we have in our pockets, on our wrists, in our ears, on our desks, and coming soon to our faces is essential if Apple wants to continue processing most data locally. It isn’t how other companies approach AI, but it makes perfect sense that it would be how Apple looks at it, given their history and marketing around privacy.

I know nothing about Apple’s plans, but it makes a lot of sense that whatever they do will be primarily (or entirely) on-device, and this repository is one clue that this remains important to them. 🕵️

Dave Verwer