The prompts shown literally invite the LLM to complete the copyrighted text by providing unedited selections and asking the machine to finish those. Even if this is problematic in a small number of cases it is not a use case that undermines the business model of the newspaper since it requires the reader to have access to the original text. Nor will it be easy to demonstrate economic harm since this is not how readers consume news and is very far from how users interact with LLMs. Nor are the archival materials used for training remotely reflective of the "time-sensitive" articles that newspapers sell. And archival materials are easily available elsewhere so where is the case for economic harm?
The courts are going to rule that LLM training is a transformative use case that is protected as fair use under copyright law. They may rule that if an LLM-powered service is explicitly designed to enable copyright violation that is illegal, but there is no way any court is going to look at these examples and see it as anything other than the NYT fishing to try and generate a violation by using the LLM in a way that is very different than the service is intended to be used and which -- even if abused -- doesn't hurt the business model under which the text has been produced.
The most likely outcome is that LLM providers will add some sort of filter on output to prevent machines from regurgitating source documents. But this isn't a court case the NYT can win without gutting fair use protections, and that would be a terrible thing.
Perhaps you are being a bit cynical? There are certainly no problems that can't be solved by giving more money to government initiatives like the MaRS Centre in Ontario. These institutions also create jobs and are businesses of a certain sort, so funding them is supporting business directly.
And who can argue with the results? Several promising entrepreneurs each year take advantage of these critical programs to consult with pro-business advisers who assist in job creation by recommending they apply for 3k tax breaks.
Do you want your transaction fee to pay for mining, or do you want it to pay for the servers that run the network? If it pays for mining, who pays for the servers that run the network?
You seem to be concerned about censorship attacks -- attacks that necessarily involve orphaning work in some capacity -- either orphaning blocks produced by honest nodes, or orphaning (refusing to include) tx-embedded routing work being sent from honest users.
The payout lottery does make censorship costly for all attackers who orphan work, but you'd need to specify the exact attack method if you want a discussion of specific work-orphaning attack vectors. Even nodes with a majority of "routing work" do not have the ability to costlessly orphan work produced by other nodes, so it isn't clear what exact attack you have in mind or why you think controlling a bunch of first-hop routing nodes under different identifies somehow makes these costs go away.
You're not describing a sybil attack. You're talking about censorship.
If you want to address that, your best strategy is to incentivize people to run access points that do not censor. This requires payouts to routing nodes in proportion to the value of the transactions they process, which requires the ability to make routing payouts, which requires a sybil-proof routing mechanism.
> There are no excess hops in the attack I am describing.
As above, you're not describing a sybil attack. Adding a bunch of 1st hop nodes to a network doesn't mean that those nodes are conducting a sybil attack.
If you want to leverage them in an attack, you would have to have them send their TXS to a central block producer that gathers the TXS from all of these first-hop nodes and then tries to use all of their collected-work to orphan the blocks/work from honest nodes in the network. If you work through the paper, you'll see that this attack is provably-costly.
You could theoretically avoid the penalty if you gave all of your 1st hop nodes the same keypair, but in that case you don't have multiple identities on the network and you don't have a sybil attack...
