[icebraining]

I'm not sure. Say you have a few TBs of random sensor data to crunch, and you allocate chunks of a few GBs to each node. What can anyone in practice do with that? Especially if you have checks to avoid tampering with results (e.g. processing every chunk twice).

[gwern]

"Especially if you have checks to avoid tampering with results (e.g. processing every chunk twice)."

See, that's the problem every time this comes up (and it's come up about as often as micropayments). By the time you get done accounting for the unreliability, hostile end-users, lack of access, customer support, low-grade obsolete inefficient un-fixable infrastructure, increased avenues for attack, duplicate processing, super-high latency, super-low bandwidth etc, you've created something that's several times slower, more expensive, more unreliable, more complex, and harder to upgrade than if you'd just stuck it all in a datacenter where you don't have to propose insane mechanisms like double your computations.