Generative models have been used in healthcare for a while for things like drug design and data generation. Not to mention all the algorithms (and probably ML) used in generating results for MRI and CT scans. I don't think this is that crazy provided they can prove it's effective.
I'd wager that for 99.9% of "Apple Intelligence" tasks, Google's models perform just as well as other frontier labs. Google also has done more work on getting LLMs running on edge devices compared to anthropic and openAI.
The source also says
> The new architecture centers on Apple Foundation Models co-developed with Google, which Apple says are adapted to run both on-device and on servers through its existing Private Cloud Compute infrastructure
Which could mean Google and Apple have trained some custom models, probably the on-device ones, specifically tailored towards Apple's hardware.
> A control vector is a vector (technically a list of vectors, one per layer) that you can apply to model activations during inference to control the model's behavior without additional prompting
Neat! It seems like Qwen 9b took the same amount of time as gemma4-e4b too, which is interesting. I haven't been able to get Qwen to stop thinking so much
Not to be nitpicky, but many of the 4-12b models are somewhere between GPT-3.5 and GPT-4o-mini. It's hard to find a good comparison though, because the benchmarks people score models against change so often. For reference, Sonnet 3.6 came out about a year after GPT 3.5
I'm a software engineer, but I use it to write down hypotheses: the cause of a bug, how I'm guessing a system works, potential fixes for said bugs, what I think this piece of documentation means, etc