Sir Arthur V Quackington

Ingesting all the artwork you ever created by obtaining it illegally and feeding it into my plagarism remix machine is theft of your work, because I did not pay for it.

Separately, keeping a copy of this work so I can do this repeatedly is also stealing your work.

The judge ruled the first was okay but the second was not because the first is “transformative”, which sadly means to me that the judge despite best efforts does not understand how a weighted matrix of tokens works and that while they may have some prevention steps in place now, early models showed the tech for what it was as it regurgitated text with only minor differences in word choice here and there.

Current models have layers on top to try and prevent this user input, but escaping those safeguards is common, and it’s also only masking the fact that the entire model is built off of the theft of other’s work.

There is nothing intelligent about “AI” as we call it. It parrots based on probability. If you remove the randomness value from the model, it parrots the same thing every time based on it’s weights, and if the weights were trained on Harry Potter, it will consistently give you giant chunks of harry potter verbatim when prompted.

Most of the LLM services attempt to avoid this by adding arbitrary randomness values to churn the soup. But this is also inherently part of the cause of hallucinations, as the model cannot preserve a single correct response as always the right way to respond to a certain query.

LLMs are insanely “dumb”, they’re just lightspeed parrots. The fact that Meta and these other giant tech companies claim it’s not theft because they sprinkle in some randomness is just obscuring the reality and the fact that their models are derivative of the work of organizations like the BBC and Wikipedia, while also dependent on the works of tens of thousands of authors to develop their corpus of language.

In short, there was a ethical way to train these models. But that would have been slower. And the court just basically gave them a pass on theft. Facebook would have been entirely in the clear had it not stored the books in a dataset, which in itself is insane.

I wish I knew when I was younger that stealing is wrong, unless you steal at scale. Then it’s just clever business.

Terrible judgement.

Turn the K value down on the model and it reproduces text near verbatim.

Provided there is an “upper limit” on what scale we are talking, Ive often wondered, couldn’t private users also host a sharded copy of a server instance to offset load and bandwidth? Like Folding@Home, but for site support.

I realize this isn’t exactly feasible today for most infra, but if we’re trying to “solve” the problem, imagine if you were able to voluntarily, give up like 100gb HDD space and have your PC host 2-3% of an instance’s server load for a month or something. Or maybe just be a CDN node for the media and bandwidth heavy parts to ease server load, while the server code is on different machines.

This kind of distributed “load balancing” on private hardware may be a complete pipe dream today, but it think if might be the way federated services need to head. I can tell you if we could get it to be as simple as volunteers spinning up a docker, and dropping the generated wireguard key and their IP in a “federate” form to give the mini-node over to an instance, it would be a lot easier to support sites in this way.

Speaking for myself, I have enough bandwidth and space I could lend some compute and offset a small amount of traffic. But the full load of a popular instance would be more than my simple home setup is equipped for. If contributing hosting was as easy as contributing compute, it could have a chance to catch on.