I'm definitely sympathetic to the ideas described in this article. I strongly feel that software should be designed to serve the interests of users, and running locally is structurally necessary to accomplish this.
For some applications, it seems plausible that mesh networking and cooperative routing could potentially replace centralized communications infrastructure. It's probably OK if an email takes half an hour or more to wind its way through intermittently connected machines to get to a recipient, and for things like maps and restaurant menus you can take advantage of physical locality.
Most of the time, though, the decentralized approach means information propagates slower, less reliably, and with greater need for redundancy (and thus overhead), so anything that's already challenging current infrastructure, like livestreaming video, is right off the table. And then there's the part where if you aren't living in an urban center there may well not be any other machines for miles to mesh with.
I'm also doubtful about the assertion that mobile devices are up to the task of "machine learning". Cell phones have severely limited power and thermal dissipation capacity. The tensor units that have started to show up on mobile devices are exclusively designed to _evaluate_ pre-trained networks; training is far more expensive and practical today only because of the economies of scale in a datacenter: train once, run many times. But to train fancy NNs, you need big heaps of data, too, and not always the sort of data you can source in an... entirely above-board manner.
Perhaps the answer is simply that some of these things aren't done anymore, or at least not for free?
For some applications, it seems plausible that mesh networking and cooperative routing could potentially replace centralized communications infrastructure. It's probably OK if an email takes half an hour or more to wind its way through intermittently connected machines to get to a recipient, and for things like maps and restaurant menus you can take advantage of physical locality.
Most of the time, though, the decentralized approach means information propagates slower, less reliably, and with greater need for redundancy (and thus overhead), so anything that's already challenging current infrastructure, like livestreaming video, is right off the table. And then there's the part where if you aren't living in an urban center there may well not be any other machines for miles to mesh with.
I'm also doubtful about the assertion that mobile devices are up to the task of "machine learning". Cell phones have severely limited power and thermal dissipation capacity. The tensor units that have started to show up on mobile devices are exclusively designed to _evaluate_ pre-trained networks; training is far more expensive and practical today only because of the economies of scale in a datacenter: train once, run many times. But to train fancy NNs, you need big heaps of data, too, and not always the sort of data you can source in an... entirely above-board manner.
Perhaps the answer is simply that some of these things aren't done anymore, or at least not for free?