Neural Engines may be used in CoreML models. I don't know it can be used with Apple's BNNS library [1]. You can use with TensorFlow Lite with coreML delegate as well [2]. And some tried to reverse engineer it and used it for model training [3].

[1] https://developer.apple.com/documentation/accelerate/bnns

[2] https://www.tensorflow.org/lite/performance/coreml_delegate

[3] https://github.com/geohot/tinygrad#ane-support-broken

Adobe Photoshop, Premiere etc make use of it for scene detection, content aware fill, "neural filters" and so on

I think the most important use of the neural engines so far is for the internal camera postprocessing. Better camera postprocessing is the reason why people buy new iPhones.

Adobe uses them Lightroom / Photoshop for some functions.

https://www.digitalcameraworld.com/news/apple-m1-chip-makes-...

Probably object tracking in videos will be the best use of them.

Or, there will be some new form of video generation (like the ones generating video from Deep Dream etc, but something aimed at studio production) using ML that wasn't practically usable before.

It opens many doors, but it will take at least many months, if not years, to see some new "kind" of software to emerge that efficiently makes use of them.

CoreML models may run on them to macOS's discretion. If you manage to get your neural network in CoreML's format you may use it.

Would something like huggingface transformers ever be able to support this? Or is it best fit to just use the GPU.

It's to build a giant distributed Aleph in which a preserved digitized Steve Jobs can live once again.

I thought it's also used to activate Siri by voice without any CPU usage

Or does any popular ML libraries support it ?