The consultation is mostly addressed to developers. And there's a lot of interesting stuff in there:
- Always-on custom hot words (via DSP)
- Access to App search
- Interoperability with Magic Cue
- Access to contextual/ambient data (e.g. to build Magic Cue alternatives)
- Ability to run App Functions exposed by other apps (these are currently system privileged, unlike standard Android intents)
- Screen automation
- APIs for better integration with Google cloud services
- On-device models
- Background execution
This stuff can be very exciting for a lot of apps, people working on Android apps should reply to the public consultation even just to say that they agree with the measures.
It is all about the openness and complexity of the technical interfaces, INCLUDING THE SDK (Software Development Kit).
Those interfaces must be straight to the point without convolution, I should be able to write assembly code which would implement directly those interfaces, that with some reasonable amount of effort.
This is a good benchmark, even though there are always ways to sneak complexity somewhere, namely stay focus on the dependencies of those interfaces, absolutely including the SDK.
If they cannot trick you, they'll go super aggressive, be careful.
i don't care much about running llms on my phone, but i would like to have at least some APIs for accessing the TPU which was apparently such a big focus in my pixel 9[0] (running graphene)
not even sure what i'd run on it but cmon. maybe some ocr/clip models for my photo gallery
[0] the "3rd Gen Edge TPU (Rio)" on the Tensor G4, apparently only accessible by the proprietary google AI Core thing, to run small gemini models or something
https://digital-markets-act.ec.europa.eu/dma100220-consultat...
The consultation is mostly addressed to developers. And there's a lot of interesting stuff in there:
- Always-on custom hot words (via DSP)
- Access to App search
- Interoperability with Magic Cue
- Access to contextual/ambient data (e.g. to build Magic Cue alternatives)
- Ability to run App Functions exposed by other apps (these are currently system privileged, unlike standard Android intents)
- Screen automation
- APIs for better integration with Google cloud services
- On-device models
- Background execution
This stuff can be very exciting for a lot of apps, people working on Android apps should reply to the public consultation even just to say that they agree with the measures.
Those interfaces must be straight to the point without convolution, I should be able to write assembly code which would implement directly those interfaces, that with some reasonable amount of effort.
This is a good benchmark, even though there are always ways to sneak complexity somewhere, namely stay focus on the dependencies of those interfaces, absolutely including the SDK.
If they cannot trick you, they'll go super aggressive, be careful.
not even sure what i'd run on it but cmon. maybe some ocr/clip models for my photo gallery
[0] the "3rd Gen Edge TPU (Rio)" on the Tensor G4, apparently only accessible by the proprietary google AI Core thing, to run small gemini models or something