"They" refers to Qubic (by Sergey Ivancheglo), a blockchain network that uses a "Useful Proof-of-Work" system, so it is not built for traditional cryptocurrency mining that solves arbitrary puzzles. Instead, it uses the collective processing power of its miners to train an AI. Qubic's AI-training work is performed by CPUs, same as used by RandomX (Monero's mining algo).
Qubic was able to orchestrate its network of miners to temporarily halt their AI-related tasks and redirect their collective CPU power to mine on the Monero network instead.
Also, Qubic has implemented an economic strategy that involves selling the Monero it mines for a stablecoin like USDT and then using those funds to benefit its own ecosystem and attract more miners, and renting hardware to gain more hash power. The proceeds from the sale of XMR are used to buy Qubic's native token (QUBIC) from exchanges. These purchased tokens are then "burned" or permanently removed from circulation.
I've looked into the "source code", and it doesn't. There is no such thing as useful PoW. Qubic isn't actually a decentralized cryptocurrency. It's closed source, runs as a EFI executable, and is only accessible from their discord channel.
The attack is no different than paying miners to join a malicious pool. It works as long as money flows in.
There is such a thing as useful proof of work. Qubic may not be doing it but it does exist. The linked papers [1][2] are examples of way to do it. They aren't 100% "useful" but rather achieve partial efficiency by essentially forcing miners down random paths in a manner that limits the ability to complete work ahead of time or otherwise "cheat".
Not really-- or, rather, the security provided by proof of work is only proportional to the part of the cost above the fair value of the useful work.
One of the main idea behind POW security is that you spend energy and the thing you get for it is income in the blockchain. And so if you mine unfaithfully your work will end up on a chain of debased value or won't end up in the eventual consensus chain at all.. so your effort is burnt out.
Now imagine a POW that costs $5 in energy and does $5 in "useful work" --- well in that system you can now attack for 'free'. Or say it costs $6 in energy to mine plus due $5 in "useful work". There your security is related to the $1, the $5 is mostly coming along for a ride.
There are other problems with "useful" proof of work: e.g. A POW function should ideally be approximation free and optimization free... if an attacker invents a better version they gain an advantage. So e.g. if the miner detects that this particular work instance is 'hard' they can just discard it and try another. This makes it really hard to do much of anything 'useful' except the most contrived kinds of 'useful' without creating vulnerabilities.
But difficulties aside, the fact that outside benefits don't contribute to security (or at least don't contribute much) makes the whole idea space kind of unexciting.
I will have to read these papers then. My intuition is that it's impossible to usefully use PoW to train neural networks because you have to rely on user-submitted training data in order to work which allows you to cheat by pre-determining the solution to your own work.
It's not a terrible idea, but I've yet to see it be inplemented. Gridcoin is one typical example where it's just PoS with "useful PoW" tacked on for token distribution, and doesn't actually use PoW for security.
Qubic was able to orchestrate its network of miners to temporarily halt their AI-related tasks and redirect their collective CPU power to mine on the Monero network instead.
Also, Qubic has implemented an economic strategy that involves selling the Monero it mines for a stablecoin like USDT and then using those funds to benefit its own ecosystem and attract more miners, and renting hardware to gain more hash power. The proceeds from the sale of XMR are used to buy Qubic's native token (QUBIC) from exchanges. These purchased tokens are then "burned" or permanently removed from circulation.