[aleem]

From a comment I had left on a related thread,

> Though very uncanny, this is rather easy to extrapolate even using the basic facets FB has available. But the most overlooked is photos. They are offline beacons and are to offline tracking what websites are to online tracking. For example you and a group of 10 other people took photos at the same location which FB sees as a small gathering of intimate friends. It will need to qualify the location is not a public restaurant to be sure its an intimate gathering. The chances of you connecting with that person are then really high. And if you are and the other person are in each others photos, then it's almost a certainty you will end up connecting on FB should FB recommend the connection. The more connected FB's network is, the more it can extrapolate based on commonalities from the first degree graph of your network. It's also the reason why Google is betting so heavy on photos and offering free unlimited storage -- remember, there's no such thing as a free lunch. Google wants to build a social network graph based on location and facial recognition to draw on proximity. With AI this will get even more uncanny. For example a wedding will have a certain photography profile (number of pics taken, the time of day, the location and venue based on Google Maps or past photographing histroy, the lighting in the photos, etc). Once you throw AI into the mix you realize Google doesn't need to draw out any conclusions. It can throw all these parameters into the AI engine and draw up proximity. In this case, Google or FB will not be able to tell you how they drew the connection, because even they won't know. All you can assume is the AI engine will take dozens of parameters today and hundreds tomorrow. Google's deep investment in AI infrastructure is a bigger testament to this.

[semi-extrinsic]

On the exact day when my daughter started first grade this year, Google Photos (which I use pretty often) notified me that it had made a new video montage. It started with a title slide saying "Look how fast they grow", and then showed photos and videos of her (not a single mistake with one of my other kids or any of her friends) growing up from ages 0 to 6, ending with the photo I had just taken of her in front of school. It was really nicely made, but it also freaked me out quite a bit how they not only detect "today is first day of school" and act on that, but also "this photo is of Kid Y, not Kid X" at all ages growing up.

[bilbo0s]

Imagine how much Google, Facebook, et al will know about your daughter when she finally turns 13 and can open an account.

Cradle to grave data collection!

[rbosinger]

I've had it make me around 3 if those already for my 2.5 year old. It's the same video but they just keep adding a few more months into it. I always save them. It's amazing, creepy and well done. I had a Premiere project where I was painstakingly trying to compile a similar thing and now I just gave up on that.

[ktsmith]

I have dozens of these videos now for my two kids and even one or two that are called "Dog days" or something like that with my dog featured instead of the kids. They are usually pretty good but every once in a while there's a blurry pic or they include a video instead of photos and there's some background motion that's distracting. They've gotten much better since they were first introduced.

[otachack]

I agree. It's just so convenient to use Google Photos as a backup and method of sharing albums that the threshold for leaving the service is super high, at least in my book.

[asteli]

Google's facial recognition engine spotted me in a photo my partner took months before we knew each other. I was visible in a crowd in the corner of a photo at the Women's March.

Clearly Google knew who I was and only chose to expose this tag info once they knew I had a connection to the person whose photo it was. It'd be amazing, and so so creepy, to see a photo of a crowd with the identification engine run unrestricted. They have the info already.

[aorav]

Can you substantiate any of what you present as facts with actual sources?

[cjsawyer]

I totally buy this

[deanCommie]

> In this case, Google or FB will not be able to tell you how they drew the connection, because even they won't know. All you can assume is the AI engine will take dozens of parameters today and hundreds tomorrow.

Sure there are ways to implement systems like this where you can't understand what they are doing. Just like you can launch websites and webservices without any metrics, monitors, or logging.

But why would you?

It requires more work, and it makes the underlying architecture more complicated but any sensible organization with a good engineering excellence fundamentals would ensure they can introspect and debug the outputs of any production algorithm.