3 ms·

I do kind of like the idea of joining AI and cryptocurrencies though. AI workloads could be an actually useful piece of "work" that there can be "proof" of. E

I do kind of like the idea of joining AI and cryptocurrencies though. AI workloads could be an actually useful piece of "work" that there can be "proof" of. EDIT: maybe I had assumed too much about the technical feasibility...

Cheer2171 2y ago
[flagged]
ben_w 2y ago
To the extent that you could have an P!=NP type thing going on where an AI does a lot of work to reach an easy-to-verify solution, you're likely to have some combination of: 1. The answers have a value uncorrelated with the price: either the problems are stupid (just like BTC's are) or they're so much more valuable than the mere mining reward that you'd do it anyway, with very little "correctly priced". 2. If the problems are completely arbitrary you get all the stupid spin-off coins just like we saw with cryptocurrency; and if they're not completely arbitrary then you vary between having lots of new problems and hardly any in exactly the same way that gold mines were suddenly found and then got mined out back when the gold standard was a thing, and IIRC that's one of the reasons against the gold standard.
- AnthonyMouse 2y ago
  > The answers have a value uncorrelated with the price: either the problems are stupid (just like BTC's are) or they're so much more valuable than the mere mining reward that you'd do it anyway, with very little "correctly priced". Why can't you solve the high value problems by having the person who wants their problem solved bid to have the miners solve their problem instead of someone else's? Then the mining reward includes the bid and stays high as long as people are willing to pay to have their problems solved. Transaction fees are near-zero because the miners are paid by the bidders, and the amount of hardware you need to do a 51% attack goes way up because mining is more profitable this way so there is more competition.
WJW 2y ago
The main problem is that an AI training workload eventually ends, or (if inference is also included) at most generates more work proportional to the rate of user queries. If you have 10 times as many workers, you are done in 1/10th of the time. Crypto on the other hand generates as more and more work as more miners join the network, so that the overall time taken remains constant. This is an essential part of the system, to prevent improvements in technology devaluing all previously mined crypto. The two have a fundamental incompatibility.

Cheer2171 2y ago

[flagged]

ben_w 2y ago

To the extent that you could have an P!=NP type thing going on where an AI does a lot of work to reach an easy-to-verify solution, you're likely to have some combination of: 1. The answers have a value uncorrelated with the price: either the problems are stupid (just like BTC's are) or they're so much more valuable than the mere mining reward that you'd do it anyway, with very little "correctly priced". 2. If the problems are completely arbitrary you get all the stupid spin-off coins just like we saw with cryptocurrency; and if they're not completely arbitrary then you vary between having lots of new problems and hardly any in exactly the same way that gold mines were suddenly found and then got mined out back when the gold standard was a thing, and IIRC that's one of the reasons against the gold standard.

AnthonyMouse 2y ago
> The answers have a value uncorrelated with the price: either the problems are stupid (just like BTC's are) or they're so much more valuable than the mere mining reward that you'd do it anyway, with very little "correctly priced". Why can't you solve the high value problems by having the person who wants their problem solved bid to have the miners solve their problem instead of someone else's? Then the mining reward includes the bid and stays high as long as people are willing to pay to have their problems solved. Transaction fees are near-zero because the miners are paid by the bidders, and the amount of hardware you need to do a 51% attack goes way up because mining is more profitable this way so there is more competition.

WJW 2y ago

The main problem is that an AI training workload eventually ends, or (if inference is also included) at most generates more work proportional to the rate of user queries. If you have 10 times as many workers, you are done in 1/10th of the time. Crypto on the other hand generates as more and more work as more miners join the network, so that the overall time taken remains constant. This is an essential part of the system, to prevent improvements in technology devaluing all previously mined crypto. The two have a fundamental incompatibility.