GMaps is horrible not only as a service that forces you to enable all their spyware technologies or lose most of the functionality, not only it’s hostile to any other mapping applications and makes it literally impossible to even copy the coordinates (I’m not even talking about converting links or using geo-uri’s), but even within itself it’s absolutely horrible for any non-car activity: hiking, walking, cycling, running, commuting — the UX of all that is absolutely abysmal.

So if you find a solution that works for you without GMaps (and preferably the one that contributes back to something like OSM), good for you, stick to that solution a forget that absolute garbage produced by Google.

No, the actual AI runs locally, on the phone. What MLKit does is two things:

1. Downloads the actual AI models from Google’s servers — not sure, but maybe they can be bundled or downloaded from other sources.

2. Send the usage analytics about those models — again, don’t remember exactly what’s being sent but the actual prompts/source images/model responses shouldn’t be sent in normal operation.

Why I highlighted the normal operation thing is because Google is kinda famous for collecting data it shouldn’t be collecting, e.g. read this README for example: https://github.com/PlqnK/magisk-supl-replacer

Not to bash them or something, but just FYI: I got interested in how they’ve implemented AI client-side, and they use Android MLKit in their Android app for that.

The problem with MLKit is that it phones back to… ta-dam!.. Google, even if it’s not actually used by the app, and that telemetry can’t be legally (and neither in any convenient and reliable way te technically) disabled, even by the app developer.

It doesn’t seem to be sending any sensitive information in that telemetry, but I don’t know Rick: changing Google for… Google?

Yes, with something like OpenRouter (or Mistral’s own API) you should be able to integrate it everywhere. Also, OpenRouter, while being a US company AFAIK, seems to be pretty transparent and lets you evaluate a lot of models from different developers and running on different platforms.

I’ve used their devstral (latest one) + goose for a side project. It worked pretty decently, on par with Claude 3.7-ish Sonnet, maybe even better. And it’s not the largest: 123B. If you can have access to their larger models, that should be even better.

It’s also useful to have a look at this great resource: https://eylenburg.github.io/android_comparison.htm