4 ms·

But then these giant corporations do not have any incentive to provide service like photos for free. Currently Google is feeding all these user photos to create

But then these giant corporations do not have any incentive to provide service like photos for free. Currently Google is feeding all these user photos to create ML models. There AI is getting stronger with every photo upload. With the approach you suggested its impossible for them to use this data.

matt-attack 7y ago
How does merely possessing a photo help ML? Wouldn’t you have to have known-to-be-correct metadata on the image as well for it to be useful?
- cowpig 7y ago
  Through something called Unsupervised Learning[1]. Have you ever noticed something in the world, and then later someone gives it a name, and you immediately understand how to identify it? For example, did you know the little plastic table in the middle of a pizza is called a "box tent"? Now you know its name, but you've noticed it before and understood what it was just by seeing it, and combining those experiences with a bunch of other real-world knowledge that you've accumulated. Same thing for unsupervised learning. edit: and to underline how that can be scary, remember that people are uploading tons of personal photos, and so google's neural networks can learn to identify anything in those photos, add them to knowledge about who those people are, who they talk to and say over email or hangouts/video chat, where they go every day with their phone, the things they read about on the internet, etc, and someday (when the algorithms further improve) end up with a more complete understanding of who that person is than even their family members can possibly know. And they are a publicly-traded company, so they have a "fiduciary duty" to use all that knowledge for profit, and we can rest assured our "the internet is a series of tubes" legislators aren't going to come up with reasonable rules about all this. [1] https://en.wikipedia.org/wiki/Unsupervised_learning https://en.wikipedia.org/wiki/Unsupervised_learning
  - matt-attack 7y ago
    Thanks for this. Truly unexpected and fascinating!

matt-attack 7y ago

How does merely possessing a photo help ML? Wouldn’t you have to have known-to-be-correct metadata on the image as well for it to be useful?

cowpig 7y ago
Through something called Unsupervised Learning[1]. Have you ever noticed something in the world, and then later someone gives it a name, and you immediately understand how to identify it? For example, did you know the little plastic table in the middle of a pizza is called a "box tent"? Now you know its name, but you've noticed it before and understood what it was just by seeing it, and combining those experiences with a bunch of other real-world knowledge that you've accumulated. Same thing for unsupervised learning. edit: and to underline how that can be scary, remember that people are uploading tons of personal photos, and so google's neural networks can learn to identify anything in those photos, add them to knowledge about who those people are, who they talk to and say over email or hangouts/video chat, where they go every day with their phone, the things they read about on the internet, etc, and someday (when the algorithms further improve) end up with a more complete understanding of who that person is than even their family members can possibly know. And they are a publicly-traded company, so they have a "fiduciary duty" to use all that knowledge for profit, and we can rest assured our "the internet is a series of tubes" legislators aren't going to come up with reasonable rules about all this. [1] https://en.wikipedia.org/wiki/Unsupervised_learning https://en.wikipedia.org/wiki/Unsupervised_learning
- matt-attack 7y ago
  Thanks for this. Truly unexpected and fascinating!