Hacker News Reader

Comment by tentacleuno | Hacker News Reader

This is unfair. When it works, local AI features are absolutely brilliant -- take Google's Now Playing[0] (and plenty more of their local AI features) for instance: they're able to enhance the user experience without draconian data collection.

Personally, I'm absolutely fine with local AIs, even if they're proprietary. That is, of course, assuming the AI layer is able to perform the task without storing its inputs and then re-uploading PII (or any data, depending on the feature).

Of course, if any data is being sent for training, then that should always be made clear to the user when they are enabling the feature (and they should be able to disable it at any time).

[0]: https://support.google.com/pixelphone/answer/7535326?hl=en#z...

View on HN

> This is unfair. When it works, local AI features are absolutely brilliant -

How about make it opt-in instead and also permanently disableable for user who do not want these features running in the background.

Of course, I completely agree that AI should not be forced upon users in any context, whether on-device or cloud-based (urgh).

> This is unfair. When it works, local AI features are absolutely brilliant

To best represent your position: Are you asserting that CoPilot is where Microsoft draws the line at sharing user data with third parties?

To clarify, when I say "local AI" I mean things which primarily function offline (although in the case of federated learning, auxiliary data / testing / better models can happen while online, but this is not required for the current model to provide usefulness).

I use "local AI" as a phrase to describe a model which doesn't need to wait for HTTP 200 to actually provide any usefulness (so where the model is local). Perhaps "on-device" is a more descriptive identifier?

As far as I understand, "CoPilot" is just a reskin of ChatGPT, so it's not "local AI".

Edit: now you mention it, I'm not really aware of any real local AI on Windows. Perhaps someone more familiar with Windows can offer some examples (I use Linux).

> To clarify, when I say "local AI" I mean things which can function offline

Understood. We'll reconcile your area of discussion with the topic at hand.

(rising to your edit!)

Let's say you found a local MS AI product in front of you. Do you have confidence it would not be a meaningful part of collecting user data?

> Do you feel that local Microsoft AI offerings will not be a meaningful part of collecting user data?

This ultimately depends on how Microsoft implements the AI -- if they use well-understood techniques such as federated learning[0] and differential privacy[1], there's a strong chance for Microsoft to offer useful features without the user sacrificing privacy.

Of course, the elephant in the room is that, since 10, privacy on Windows has taken a nosedive. I'm really not sure if they'd even want to spend the R&D work on doing things The Right Way.

> Or if you allow that they could, do you believe that MS will be steadfast in not sharing it?

There are good techniques for ensuring local AI (or "on-device AI"?) is private as of now, so I'm not sure.

EDIT: (Rising to yours, too!)

> Let's say you found a local MS AI product in front of you. Do you have confidence it would not be a meaningful part of collecting user data?

With the history of privacy on Windows, I would be very hesitant that it would preserve privacy without doing some serious in-depth analysis on it, to be honest.

[0]: https://federated.withgoogle.com/#top

[0.1]: https://en.wikipedia.org/wiki/Federated_learning

[1]: https://en.wikipedia.org/wiki/Differential_privacy

Oh hey. Look what I found on page 2.

Intel, Microsoft discuss plans to run Copilot locally on PCs instead of in the cloud

Discussion: https://news.ycombinator.com/item?id=39852166

Oh yeah! I totally forgot about Intel's on-chip AI acceleration[1]. I'm not sure how well it works, but I know that it exists and nothing major is currently taking advantage of it AFAIK.

Local Copilot sounds interesting, but I'd imagine it would be akin to the old Siri on iPhones where its use would be limited to controlling on-device settings. I've heard and read quite a lot about the dangers of that, though: with prompt injections being an issue with Copilot[0] (and Bing Chat frequently going crazy[2]), I can understand why Microsoft are very hesitant to give it any real control over the system.

[0]: https://embracethered.com/blog/posts/2024/whoami-conditional...

[0.1]: https://kai-greshake.de/posts/llm-malware/

[1]: https://www.tomshardware.com/news/intel-details-meteor-lakes...

[1.1]: https://www.extremetech.com/computing/intel-announces-an-int... (hard to find a good explanation of these accelerators)

[2]: https://www.forbes.com/sites/siladityaray/2023/02/16/bing-ch...

Unfair? Eh.

That doesn't sound very useful to me. You can just ask the person playing, or enjoy the ephemeral moment in ignorance. If it's really important you can record it and then find out later what it was.

Also that feature seems to be a good fit for phones. I don't think it would be useful on a laptop or desktop PC.

This could be said for thousands of features.

For instance, you can also control your central heating / AC with a rocker switch, but most people opt for thermostats. If something can be made easier and more enjoyable (which ultimately enhances the experience), why not? In the case of Now Playing, spend 5 minutes online and you'll see that people absolutely love it (I too, before I opted for a custom ROM).

Then again, I believe we're both missing the point here: this is just one feature (out of thousands!) that can be pulled off with on-device AI. It shouldn't be used as a deciding factor for AI in general.

Reply on news.ycombinator.com