[rm999]

>Google is about to introduce a next-generation Tensor chip, updating the ones found in the company’s Pixel 6 phones. It is basically an artificial-intelligence accelerator in your pocket, allowing your phone to adapt to you—your own personal neural network to train.

Is this accurate? Everything I've read online only discusses the tensor chip for inference (running the neural network to make predictions), not training a new model for that specific device. For example this blog only discusses how they are making inference faster: https://ai.googleblog.com/2021/11/improved-on-device-ml-on-p....

[nostrademons]

Google announced Federated Learning about 4 years ago:

https://ai.googleblog.com/2017/04/federated-learning-collabo...

I work adjacent to some teams using it. It's not idle research speculation; there are consumer product teams actively commercializing it.

(IMHO - speaking only for myself and not my employer - Federated Learning will be the most important invention to come out of Google, and its primary beneficiary will be someone other than Google, in the same way that the Alto was the most important invention to come out of Xerox but made a lot of people other than Xerox rich. It lets you build machine-learned models off of billions of devices without having centralized infrastructure or data storage, which finally makes the convenience and functionality of modern consumer software compatible with decentralized architectures and user privacy.)

[skybrian]

Federated learning is neat in some ways but I’m wondering how this solves anyone’s privacy concerns? As we saw with the FLoC controversy, people aren’t going to be any happier about collecting data via an opaque algorithm that they can’t audit.

I think it does avoid collecting data. The problem is that it doesn’t help for gaining trust, so only an already-trustworthy organization could use it.

[nostrademons]

The algorithm itself is public - the blog post has arXiv papers on it.

People have shown themselves very willing to trust algorithms that are opaque (as in difficult to understand) but auditable (by other people). Witness Bitcoin or Ethereum, where they will put their life savings into an algorithm. They're just not willing to trust organizations or people, where the organization can change the rules of the game at whim. (Even this has exceptions - ETH flourished over ETC despite doing a hard fork to rewrite history after the DAO hack.)

[sifar]

>> The algorithm itself is public - the blog post has arXiv papers on it.

This is misleading. The implementation (running on a phone) would not be and there is no way to verify.

[nostrademons]

Why not? The implication of this comment thread is that the implementation would not be Google's.

[skybrian]

The software and hardware supply chains need to be trusted. This would require more transparency than we get from Apple or Google. It’s doable, but who is going to do it?

I don’t think people trusting cryptocurrency is much of an argument because cryptocurrency investors in general are a minority, and also, people do a lot of stupid things with cryptocurrency.

[ffhhj]

> made a lot of people other than Xerox rich

But this time Google should be able to peek into what other people do with its invention and get rich too, right?

[nostrademons]

Potentially, but that's not what killed Xerox (who had a decade head start on Microsoft and Apple and had full visibility on what they were doing in the marketplace for another decade before GUIs became commonplace).

They fell victim to the Innovator's Dilemma: when a new technology comes out that fundamentally rewrites the rules about who your customers are and how markets should be organized, existing market leaders cannot maintain their lead. That's because their market itself gets restructured - where previously one market with a dominant firm and set of existing customers used to exist, now oftentimes new markets with new dominant firms and a new set of customers exist. Basically it rewrites all the assumptions that the business is structured around.

[guerrilla]

I'm pretty sure TechTechPotato (whom I consider a reliable souece) said it would be used for training as well as inference but didn't mention use for specific devices.

[timpattinson]

who's probably known better as Dr Ian Cuttress