[Terretta]

Not all big tech are at it.

The reason you hear a global sigh that “my battery life is terrible!” at every major iOS release, is the photo face to people clustering is done entirely on-device, and which photos are in the cluster is not shared across devices through iCloud.

There are several such features where Apple has made life difficult for themselves and the UX a little less seamless for users, in service of not giving themselves all the data.

Depending how the politics of privacy go, we could find an aggregation of personal data is not an asset, but a liability.

In such an environment, avoiding “hoovering up all data” is both a marketing point, and a reduction of risk exposure.

[prepend]

Is Apple the only big tech that’s using AI on device in a privacy protecting way?

I appreciate their approach to user functionality that doesn’t dump all my data into a near infinite number of collectors.

I also think it’s a neat way of distributed computing to use the cpus of all these phones sleeping at night to crunch stuff is actually cheaper than cloud. I think because the phones are sunk cost that is idle a few hours a day.

I miss the BOINC/Folding/SETI at Home clients that used to run as my screensaver and try to do good.

[dawg-]

I think Google is trying to do this with the new update to Assistant - the language model is only 500mb so it can all be processed on the local device. I saw in a blog post somewhere they are trying to get to where individual devices hold data and compile these aggregated profiles on the device, the send the profiles to google for their ad services and other big data stuff

[prepend]

I use Google Translate and that works well with offline datasets to do CV and translation without any network or cloud usage.

I don’t think the photo app or assistant app allows this functionality although they likely could.

[solidasparagus]

Believe it or not, Google is the leader in the space of privacy-centric AI - see Federated Learning.

[Arelius]

I don't buy the fact that "Ferderated Learning" is privacy centric so much as it is liability and/or privacy-marketing centric. I feel that in the long run we will discover that using "ML" as a black box to smuggle user data off device won't hold up to scrutiny.

[prepend]

I find that hard to believe as Google’s definition of privacy is not mine. I believe Google could be awesome if they wanted, but it’s in conflict with their business mode.

[froindt]

I'm not an iOS user. Is there a reason the clustering can't be done while charging? Does it simply take longer than a charge cycle if a person has thousands of photos?