My fears about Android 17

lpsry

Google showed a new firmware, where everything is very strongly tied to their AI. Yes, I'm not too well aware of this system, so I can make it so, but the trend looks very frightening. If Gemini already allow so much to climb into the phone system and personal data of a person, what remains in a new update, except it? Yes, in the system could have made better security, some new chips (by the type of meaningless with the switch on wifi and mobile internet), but given that all the main functions are tied to Gemini, which will so and so come cut, what will remain from the system? Given the current Google policy, there will be too much in the future, there will be too much on the neural network - you will not have time to look back as for your device's safety. Yes, of course, it was all expected, but I would like to know someone else's opinion about it, as well as ideas how to deal with the same developers of grapheneos?

And, yeah, sorry, my English is not so good. I hope, my idea was clear

other8026

lpsry It seems you're more worried about the stock OS than GrapheneOS. Gemini isn't installed on GrapheneOS by default. Even if a GrapheneOS user installed Gemini, it wouldn't have any special access just because it's a Google app.

de0u

lpsry Google showed a new firmware, where everything is very strongly tied to their AI.

Though it is possible to regard an Android OS release as "firmware", I think most members of the GrapheneOS community use that term for low-level software that directly drives specific hardware, e.g., cellular modem firmware, Wi-Fi firmware, Bluetooth firmware, GPU firmware -- as opposed to the kernel, system services, or apps that users interact with.

So far it does not seem to be the case that Google is putting Gemini (or any other "AI") into device firmware. Also, despite lots of marketing about Gemini being "integrated" or "fundamental", so far it is deployed in specific apps, and so far those apps are not part of AOSP and thus are not part of GrapheneOS.

Overall, it is not expected that the Android 17 version of GrapheneOS will include any "AI" features from Google. This is exactly what happened with Android 16: lots of marketing messages about "AI" being integrated in Android 16, but zero addition of AI features to GrapheneOS.

Developer-Dude

lpsry GrapheneOS never has had AI integrated into it and very likely Android 17 won't either (which is one of the reasons I love GrapheneOS!). The AI features are all on Stock OS, as @other8026 said.

Carrousel7956

I haven't looked into this, so this could be totally wrong, but I think you are confusing the Android that GrapheneOS uses and the operating system that Google uses on Pixels. All those Gemini features are on top of the AOSP. Some you may be able to install as apps like the way you can install Google Camera on GrapheneOS and get the AI features.

Johnnyloans

Developer-Dude

The new speech app is AI.

All it does is figure out how some written words would sound like

Developer-Dude

Johnnyloans The new speech app is AI.

Really? Where was that mentioned? I mean, I saw this;

Our Speech Services app uses a fully open source model for text-to-speech which we created ourselves using existing open source code and data. We'll be able to train a better model to replace this one now that we have an RTX 5090 for this purpose. We also plan to make voices for other languages too.

https://discuss.grapheneos.org/d/35722-initial-release-of-grapheneos-speech-services-for-text-to-speech

But I assumed it was a voice recording or whatever

de0u

Johnnyloans The new speech app is AI.

"AI" sounds like it means something, but unfortunately it means too many things to mean anything in particular. "AI" is like "vehicle" or "plant". Some vehicles fly, others "swim"; some have weapons attached, but most do not. Some plants are too small to see, and some are taller than buildings.

Johnnyloans All it does is figure out how some written words would sound like.

Yes.

soupslurpr It's a text-to-speech model, it isn't a general purpose AI model [...]

Exactly. The GrapheneOS text-to-speech model:

...can transform words that are sent to it into sounds
...doesn't "understand" pictures or audio at all
...doesn't scan any part of the system for anything
...doesn't report anything to anybody
...doesn't learn, plan, or form opinions

Johnnyloans

Yeah, I suppose for lots of people it wouldn't be enough to only say this.

Johnnyloans The new speech app is AI.

All it does is figure out how some written words would sound like

Because, people wouldn't understand that something is simply a classifier, vision model. They could be imagining it's a murderous image processing demon working undercover as a cashier recognizing items while it plots its revenge.

What you said here was good.

The GrapheneOS text-to-speech model:
...can transform words that are sent to it into sounds
...doesn't "understand" pictures or audio at all
...doesn't scan any part of the system for anything
...doesn't report anything to anybody
...doesn't learn, plan, or form opinions

edit: sorry to ping you twice.

soupslurpr

Developer-Dude it's a text-to-speech model, it isn't a general purpose AI model and isn't built the same way as those. It's built in a way where it's very unlikely to have hallucinations. There hasn't been any in testing or any reported by users unlike some other text-to-speech models.

Johnnyloans

The words model and train tells us it's AI

... fully open source model for text-to-speech ...
We'll be able to train a better model ...

Or the tool listed in the app store description for speech is for AI models.

Developer-Dude But I assumed it was a voice recording or whatever

This would be like recording an interview of someone and then cut each word then switch it around. It'd sound very strange. About 800GB of audio files are used to train an AI which takes up 200MB (random guess). There are small and big models.

soupslurpr

Johnnyloans the main model was trained on around 23 hours of speech, and is around 26 MB.

Johnnyloans

soupslurpr

Wow that's a lot smaller than Mozilla's common voice English dataset. I don't remember how big exactly, it's been years.

That's amazing for 23hours

gristle-cause

The devs are pretty openly passionate about attack surface reduction - by assuming every piece of software could have a vulnerability, you become motivated to minimize your software footprint.

AI models like what is being discussed in Android17 are massive. It's a tremendous increase in attack surface

On those grounds alone, I spend no time worrying about future AI integration on GOS

Johnnyloans

de0u

I think if the audience is largely laypersons then yes it's confusing; but -- then again -- there will be ample explanation present since we are taking the audience into account.

To me, "AI" is an informal, less-wordy alternative for "neural network." Which is an accurate description for anyone in the know. It wouldn't be the first time a computer topic had to be "dumbed down" for my accountant brother or similar.

de0u Some vehicles fly, others "swim"

You are right that a word like "travel" could also mean many things but it's used plenty.

Based on context -- most of the time -- we won't have to spell out: "a neural network that transforms written text into audio output." I prefer writing that flows better.

Finally, in this specific case, one must use the word AI for the conversation.
"there isn't any AI in GOS"
"Yes there is AI. It turns written word into sound"

A real useless umbrella term is: "information & technology". Hardly helps narrow it down haha.

de0u

Johnnyloans Finally, in this specific case, one must use the word AI for the conversation.

The problem with answering either that there is "AI" in GrapheneOS or that there isn't "AI" in GrapheneOS is that the word means opposite things to different people. Clearly Google believes they can sell more Pixels by promising that Pixels contain "AI", but just as clearly many users of this forum do not want to have any "AI" in GrapheneOS.

The difference between the GrapheneOS TTS model on the one hand, and either Gemini Nano or Gemini Intelligence on Google's stock OS on the other hand, is like the difference between a bicycle and a helicopter. Does it make sense to have a "vehicle-free policy" for one's neighborhood because one doesn't want a helicopter landing on one's front lawn? Does it make sense to have an "AI-free policy" for one's phone because one doesn't want an AI agent with Internet access scanning one's photos? I think using broad category names often confuses the issue to the point of fear or paralysis, instead of simplifying it.

These are confusing times. I don't think it's going to work out in general for people to decide either "AI is good" or "AI is bad", and I think getting past that will often require zooming in past the general "AI" label to things like "text-to-speech model" versus "AI agent".

Johnnyloans

There are certainly many situations where one should read the room to determine whether: a disclaimer, warning, or [class]101 should be tacked onto an otherwise simple reply.